Home

DIET Users Manual

Contents

1. Set an environment variable called JXTA_LIB containing the path to the JXTA JAR files They are by default provided in the lt diet_root gt src lib direc tory At last the command to be launched to run a MA is java cp lt JXTA_JARS gt JXTAMultiMA lt DIET_MA_config file gt Ensure that this command is launched inside the right directory indeed only one peer can be launched by directory information concerning this peer is available in a jxta directory under the directory where you launched the peer Delete this directory before launching a peer if you have already used it on another machine in order to clean the platform configuration e Each time anew JXTA peer is launched you have to configure it On the first setup screen the name of the peer is required and must be unique for instance MA1 for the first IMA you load The second screen named advanced displays the TCP and HTTP settings When using DIET on a single machine the configuration is as shown on Figure 12 2 else just replace localhost by the IP address of the machine Please note that for each peer on a single machine the TCP and HTTP ports have to be different For instance 9701 and 9700 for the first peer 9703 and 9702 for the second etc The third setup screen deals with the web access If you want to access peers outside the local network references of rendezvous and relay peers placed at the disposal of JXTA users by the JXTA community
2. Before compiling DIET itself first install the above mentioned cf Section 2 1 5 dependen cies Then untar the DIET archive and change current directory to its root directory 2 2 1 Obtaining and installing cmake per se DIET requires using cmake at least version 2 4 3 For many popular distributions cmake is incorporated by default or at least apt get or whatever your distro package installer might be is cmake aware Still in case you need to install an up to date version cmake s official site distributes many binary versions alas packaged as tarballs which are made available at http www cmake org HTML Download html Optionally you can download the sources and recompile them this simple process bootstrap make make install is described at http www cmake org HTML Install html 2 2 2 Configuring DIET s compilation cmake quick introduc tion If you are already experienced with cmake then using it to compile DIET should provide no surprise DIET respects cmake s best practices e g by clearly separating the source tree from the INRIA ENS Lyon UCBL Page 16 DIET User s Manual BOSE TE binary tree or compile tree by exposing the main configuration optional flag variables prefixed with DIET_ and by hiding away the technical variables and by not postponing configuration difficulties in particular the handling of external dependencies like libraries to compile stage Cmake classically provides two ways for
3. e a native method that launches the MA pies e a native method submitJXTA that unpacks the description of the problem to be solved in order to build a DIET problem description calls the DIET submit function and thus gets a response extracts and returns the SeD reference s to the MA INRIA ENS Lyon UCBL Page 108 DIET User s Manual BOSE E The SeD pie To solve the client s computation requests the SeD needs to call the SeD pie solve function In the same manner as above to allow this the SeD z launches the SeD pie via a native method and calls the solve function via another The SeD pie contains e a native method that launches the SeD Diet e a native method solveJXTA that unpacks the problem to be solved and builds a DIET profile calls the solve function extracts and returns the response to the SeD 7 123 The future of DIET 12 3 1 Remaining problems e An unsolved problem dealing with omniORB and JNI results in a failure when a JNI SeD piet registers to a DIET Agent not launched via JNI Because of that to deploy some LAs between a DIET MA and a DIET SeD they have to be launched via JNI Moreover a DIET MA won t be able to know LAs or SeDs not launched via JNI The current DIET tree is unable to contain classic LAS piet or SeDs Diet e The current version of the DIET platform works only for problems having two input matrices and one output matrix The seriali
4. on convertors see section 5 5 If mode is DIET_PERSISTENCE_MODE_COUNT or if base_type is DIET_BASE_TYPE_COUNT or if order is DIET_MATRIX_ORDER_COUNT or if size nb_rows nb_cols or length is 0 or if path is NULL then the corresponding field is not modified int diet_scalar_set diet_arg_tx arg void value diet_persistence_mode_t mode diet_base_type_t base_type int diet_vector_set diet_arg_t arg void value diet_persistence_mode_t mode diet_base_type_t base_type size_t size Matrices can be stored by rows or by columns typedef enum DIET_COL_MAJOR O DIET_ROW_MAJOR DIET_MATRIX_ORDER_COUNT diet_matrix_order_t int diet_matrix_set diet_arg_t arg void value diet_persistence_mode_t mode diet_base_type_t base_type size_t nb_rows size_t nb_cols diet_matrix_order_t order int diet_string_set diet_arg_t arg char value diet_persistence_mode_t mode The file size is computed and stocked in a field of arg Warning The path is not duplicated int diet_file_set diet_arg_t arg diet_persistence_mode_t mode char path 3 4 2 Access functions A NULL pointer is not an error except for arg it is simply IGNORED For instance INRIA ENS Lyon UCBL Page 28 DIET User s Manual BOSE TE diet_scalar_get arg amp value NULL will only set the value to the value field of the arg structure NB these are macros that let the us
5. start printing CoRI values cpu average load 0 56 CPU O cache 1024 Kb number of processors 1 CPU O Bogomips 5554 17 diskspeed in reading 9 66665 Mbyte s diskspeed in writing 3 38776 Mbyte s total disk size 7875 51 Mb available disk size 373 727 Mb total memory 1011 86 Mb available memory 22 5195 Mb end printing CoRI values 8 3 2 FAST FAST as collector of CoRI gives the user the same information as without CoRI see table 3 1 to know which information FAST can provide INRIA ENS Lyon UCBL Page 80 DIET User s Manual BOSE E 8 3 3 CoRI Easy The CoRI Easy collector makes some basic system calls to gather the information CoRI Easy is only available if DIET is compiled with the option DDIET_USE_CORI set to ON The last collumn of the table 8 1 corresponds to the CoRI Easy s functionality There is an example on how to use CoRI Easy in the lt diet_src gt src examples cori directory 8 3 4 CoRI batch With the help of the CoRI batch collector a SeD programmer can use some information obtained from the batch system It is only available if DIET is compiled with the option DDIET_USE_BATCH set to ON For the moment only simple information can be accessed but functionalities will be improved as well as the number of recognizable batch systems There is an example on how to use CoRI batch in the lt diet_src gt src examples Batch Cori_cycle_stealing directory 8 4 Future Work There
6. USER S MANUAL VERSION DATE PROJECT MANAGER EDITORIAL STAFF AUTHORS STAFF Copyright 2 4 April 2010 Fr d ric DESPREZ Yves CANIOU Eddy CARON and David LOUREIRO Abdelkader AMAR Rapha l BOLZE ric Borx Yves CANIOU Eddy CARON Pushpinder Kaur CHOUHAN Philippe COMBES Sylvain Danan Holly DAIL Bruno DELFABRO Peter FRAUENKRON Georg HOESCH Benjamin ISNARD Mathieu JAN Jean Yves L EXCELLENT Gal LE MAHEC Christophe PERA Cyrille PONTVIEUX Alan Su C dric TEDESCHI and Antoine VERNOIS INRIA ENS Lyon UCBL DIET User s Manual BOSE TE Contents Introduction 8 1 A Diet platform 11 1 1 DIET components ccoo 28455 eh bed oe Pete ee stade di 12 1 2 Communications layer cs 4 4 4 bee 4 da bu ge ae da du eue 12 13 DIET initialigation 14 su sua oe die han tt ae us Ho Rite ee 13 LA Soie problem s p s assis qua a ER las RUES oO ea date nd 13 US DIET Extensions oe Lo A BOR EU Don wh as 14 Tol MMM Le du um uen sh Nam Rte RAM UN ape 14 Moe PAST a oy hae we des Due ee une eh ee Be Un eS 14 too CORI yk e oY PY Ba he RA ee ee ae F 14 2 DIET installation 15 21 Dependenties oo soset da kee a ia eee EE oe A ee Se e dd ane 15 2 1 1 General remarks on DIET platform dependencies 15 21 2 Hardware dependencies lt co ccosa kew a eee eee ee 15 213 Supported comipilers rocs s ei aep o a ae A ee oe le de 15 2 1 4 Operating system dependencie
7. data ue du Du Bee ER due a RE ee ae Bee 48 6 5 Batch system 1444442042 8482 Re bambou ane ee ee es 48 6 6 Client extended API 48 6 7 Batch server extended API and configuration file 49 6 8 Server APL 2 64 2 Seon a moia Bae SR a a a eS 49 6 8 1 Registering the service 50 6 8 2 Server configuration file 50 6 8 3 Server API for writing services 50 6 8 4 Example of the client server concatenation problem 51 7 Scheduling in DIET 53 Tel Introduction uc 8 58464 a a a EG he a Ree ee 53 7 2 Default Scheduling Strategy gt se sane ae Wo ha eR ee mans 53 73 Plugin Scheduler Interface cocinera qu Bah Se ee 54 1 3 1 Estimation Metric Vector s o c s a 4 sorib be peewee tae ee bea sg 54 has otandard Estration WAGs me ee a Ru oe De ee ee 54 INRIA ENS Lyon UCBL Page 4 DIET User s Manual BOSE TE 10 0 Estimabion FUNCHON s3 a ne ke eoa A ae a ele ok ae 56 7 3 4 Aggregation Methods 57 TA NORTE ee lucia Pe be et eA ee ee Ne RE Da ann eee he 58 Ta Scheduler at agents level sia e koe dus pe Bek Rie Rh woe Roe ek ee a 59 7 5 1 Scheduling from the agents side 60 7 5 2 Aggregation methods overloading 61 7 5 3 The UserScheduler class ca sacs 86 du unes vu sagas Ew 61
8. separated list of integers for example 2 3 for an array containing 2 arrays of 3 elements Iteration strategies Iteration strategies must be defined when the processor has two or more input ports By default the workflow parser will use a dot iteration strategy for all inputs These operators use the index of data items received or produced by workflow processors to combine them The index of a data item corresponds for data items produced by a source to the order number in the source data file and for data items produced by a standard processor to the index of input data items eventually combined by the operators There are 4 data manipulation operators e dot groups 2 or more ports data from the different ports are processed together when their index match exactly data with index 0 of one port is matched with data with index 0 of the other ports The output index is the same as the index of the input data e cross groups 2 ports processes each data instance of the first port with each data instance of the second port This processor will increase by one the index depth of the output for example if data inputs have indexes 0 and 1 then the outputs have the index 0_1 e flatcross groups 2 ports same as cross but with a different output indexation scheme This operator does not increase the depth of the output index but creates INRIA ENS Lyon UCBL Page 120 DIET User s Manual BOSE E new indexes for example
9. tion times The following option is ignored if nwsUse is set to 0 nwsNameserver default none Specify the host port address of the NWS name server Multi MA options To federate resources each MA tries periodically to contact other MAs These options define how the MA connects to the others neighbours default empty list List of known MAs separated by commas The MA will try to connect itself to the MAs named in this list Each MA is described by the name of its host followed by its bind service port number see bindServicePort For example host1 domain com 500 host4 domain com 500 host domainB net 2001 is a valid three MAs list By default an empty list is set into neighbours maximumNeighbours default 10 This is the maximum number of other MAs that can be connected to the current MA If another MA wants to connect and the current number of connected MAs is equal to maximumNeighbours the request is rejected minimumNeighbours default 2 This is the minimum number of MAs that should be connected to the MA If the current number of connected MA is lower than minimumNeighbours the MA tries to connect to other MAs updateLinkPeriod default 300 The MA checks if the connected MAs are alive every updateLinkPeriod seconds bindServicePort default none The MAs need to use a specific port to be able to federate themselves This port is only used for initializing connections between MAs If this parameter is not se
10. 3 MAIN int main int argc char argv int res std string service_name dyn_add_rem_0 checkUsage argc argv Add service diet_service_table_init 1 add_service service_name c_str Print service table and launch daemon diet_print_service_table res diet_SeD argv 1 argc argv return res 16 2 3 Going further Finally another example is provided in src examples dynamicServiceMgr showing how to dynamically load and unload libraries containing services Hence a client can send a library to as server and for as long as the library is compiled for the right architecture the server will be able to load it and instanciate the service present in the library The service can further be called by other clients and whenever it is not required anymore it can be easily removed INRIA ENS Lyon UCBL Page 145 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 146 DIET User s Manual BOSE TE Appendix A Appendix A 1 Configuration files traceLevel e Component All e Mode All e Type Integer e Description traceLevel for the DIET agent 0 DIET prints only warnings and errors on the standard error output 1 default DIET prints information on the main steps of a call 5 DIET prints information on all internal steps too 10 DIET prints all the communication structures too gt 10 traceLevel 10 is given to the ORB to print CORBA mes
11. Endpoint Options dietPort default none This option specifies the listening port of an agent or SeD If not specified the ORB gets a port from the system This option is very useful when a machine is behind a firewall By default this option is disabled dietHostname default none The IP address or hostname at which the entitity can be contacted from other machines If not specified let the ORB get the hostname from the system by default omniORB takes the first registered network interface which is not always accessible from the exterior INRIA ENS Lyon UCBL Page 86 DIET User s Manual BOSE E This option is very useful in a variety of complicated networking environments such as when multiple interfaces exist or when there is no DNS LogService options useLogService default 0 This activates the connection to LogService If this option is set to 1 then the LogCentral must be started before any DIET entities Agents and SeDs will connect to LogCentral to deliver their monitoring information and they will refuse to start if they cannot establish this connection See Section 10 1 to learn more about LogService lsOutbuffersize default 0 lsFlushinterval default 10000 DIET s LogService connection can buffer outgoing messages and send them asynchronously This can decrease the network load when several messages are sent at one time It can also be used to decouple the generation and the transfer of messages The b
12. Generally a parallel resource is managed by a batch system and jobs are submitted to a site queue The batch system is responsible for managing parallel jobs it schedules each job and it determines and allocates the resources needed for its execution There are many batch system among which Torque a fork of PSB Loadleveler devel opped by IBM SunGrid Engine SGE developped by Sun OAR developped at the IMAG lab Each one implements its own language syntax with its own mnemonics as well as its own scheduler Jobs can generally access the identity of the reserved nodes through a file during their execution and are assured to exclusively possess them 6 6 Client extended API Even if older client codes must be recompiled because internal structures have evolved they do not necessarily need modifications DIET provides means to request exclusively sequential services parallel services or let DIET choose the best implementation of a problem for efficiency purposes according to the scheduling metric and the performance function To explicitly call a sequential service diet_error_t http old clusterresources com products torque 3http www clusterresources com pages products torque resource manager php http www 03 ibm com servers eserver clusters software loadleveler html Shttp www sun com software gridware Shttp oar imag fr INRIA ENS Lyon UCBL Page 48 DIET User s Manual BO
13. base_type elements pointed by value with the persistence mode mode and stores the data ID in JD dagda_put_matrix void value diet_base_type_t base_type diet_persistence_mode_t mode size_t nb_rows size_t nb_cols diet_matrix_order_t order char ID This macro adds to the platform the base_type matrix of dimension nb_rows x nb_cols stored in order order The data ID is stored on ID dagda_put_string char value diet_persistence_mode_t mode char ID This macro adds to the platform the string pointed by value with the persistence mode mode and stores the data ID into JD dagda_put_file char path diet_persistence_mode_t mode char ID This macro adds the file of path path with the persistence mode mode to the platform and stores the data ID into JD DAGDA get data macros functions The following API functions are defined to obtain a data from DAGDA using its ID dagda_get_scalar char ID void value diet_base_type_t base_type The scalar value using the ID D is obtained from DAGDA and the value ar gument is initialized with a pointer to the data The base_type pointer content is set to the data base type This last parameter is optional and can be set to NULL if the user does not want to get the base_type value dagda_get_vector char ID void value diet_base_type_t base_type size_t size The vector
14. e Description The directory on which DAGDA will store the data files By default tmp is used USE_SPECIFIC_SCHEDULING e Component Client e Mode Custom Client Scheduling CCS e Type String e Description This option specifies the scheduler the client will use whenever it submits a request BURST_REQUEST round robin on the available SeD BURST_LIMIT only allow a certain number of request per SeD in parallel the limit can be set with void setAllowedReqPerSeD unsigned ix INRIA ENS Lyon UCBL Page 155 DIET User s Manual BOSE E clientMaxNbSeD e Component Client Mode All Type Integer e Description The maximum number of SeD the client should receive INRIA ENS Lyon UCBL Page 156 DIET User s Manual BOSE E Bibliography 1 D Arnold S Agrawal S Blackford J Dongarra M Miller K Sagi Z Shi and S Vadhiyar Users Guide to NetSolve V1 4 Computer Science Dept Technical Report CS 01 467 University of Tennessee Knoxville TN July 2001 http www cs utk edu netsolve 2 F Cappello F Desprez M Dayde E Jeannot Y Jegou S Lanteri N Melab R Namyst P Primet O Richard E Caron J Leduc and G Mornet Grid 5000 A large scale reconfigurable controlable and monitorable grid platform In Proceed ings of the 6th IEEE ACM International Workshop on Grid Computing Grid 2005 Seattle Washington USA November 2005 3 E Caron and F Suter Parallel E
15. format The file can be used as input data file for another workflow execution int diet_wf_save_transcript_file diet_wf_desc_t profile const char transcript file name saves the transcript of the current work flow list of tasks with their status and data This file can be used as input tran script file for another workflow execution tasks already done with output data still available on the platform will not be exe cuted again int diet_wf_sink_get diet_wf_desc_t wf_profile const char id char datalD gets a container DAGDA data containing all the data received by a sink node Table 14 3 Diet workflow Functional specific API INRIA ENS Lyon UCBL Page 126 DIET User s Manual BOSE E can be fully execution and result retrieving executed with this client include lt string h gt include lt unistd h gt include lt stdlib h gt include lt stdio h gt include lt sys stat h gt include DIET client h int main int argc char argv i diet_wf_desc_t profile char x fileName long l if argc 3 4 fprintf stderr Usage s lt file cfg gt lt wf_file gt in argv 0 return 1 if diet _initialize argv l argc argv fprintf stderr DIET initialization failed n return 1 fileName argv 2 profile diet_wf_profile_alloc fileName test DIET WF DAG if diet_wf_call profile pri
16. private UserScheduler const char moduleName static UserScheduler instance void module These two methods are obtained from the loaded module constructor constructs destructor destroys The aggregate method takes 4 arguments e corba_response_t aggrResp the result of the aggregation has to be set in this argument aggrResp is an array of corba_server_estimation_t objects e size t max srv this argument gives the maximum number of responses to return in aggrResp This value can be ignored without any risk and it is sometimes useful to ignore it because this parameter is hard coded in the DIET sources e const size_t nb_responses this argument gives the number of responses in responses INRIA ENS Lyon UCBL Page 62 DIET User s Manual BOSE E e const corba_response_t responses the responses are stored in this argument It is an array of corba_response_t which is a CORBA structure containing a CORBA sequence of corba_server_estimation_t The corba_response_t structure is defined as follows struct corba_response_t typedef _CORBA_ConstrType_Variable_Var lt corba_response_t gt _var_type CORBA ULong reqID CORBA Long myID SeqserverEstimation_t servers void operator gt gt cdrStream amp const void operator lt lt cdrStream 4 F The _var_type field is an internal CORBA object The scheduler developer does not have to use it The two operators operator gt gt and ope
17. profile print on standard output all the results of the current executed workflow or dag Table 14 1 Diet workflow common API Workflow function Description int diet wf scalar get const char id void value retrieves a workflow scalar result id the output port identifier int diet_wf_string get const char id char value retrieves a workflow string result id the output port identifier int diet_wf_file_get const char id size_t size char path retrieves a workflow file result id the output port identifier int diet_wf_matrix_get id void value nb_rows nb_cols order retrieves a workflow matrix result id the output port identifier Table 14 2 Diet workflow DAG specific API INRIA ENS Lyon UCBL Page 125 DIET User s Manual eae D Workflow function Description void diet_wf_set_data_file diet_wf_desc_t profile const char data_file name specifies the file containing the data de scription used to generate the workflow void diet_wf_set_transcript_file diet wfdesc_t profile const char transcript file name specifies the file containing the tasks sta tus and data used to restart a dag or workflow int diet_wf_save data file diet_wf_desc_t profile const char data_file name saves the input and output data descrip tion source and sink nodes in XML
18. s Manual BOSE TE e create an advertisement including its input pipe reference allowing clients to connect to it back and publish it with a hardcoded lifetime e process each client or agent message by a thread if the source of the message received is a client call the MA pa with the packed problem and get SeD reference s if any send it to the client else search other MA y forward the query to all other MA5 7 discovered and send a response containing all SeD references thus received to the client if the source is an agent x call the MA pie on the problem received and get SeD references found in its own DIET tree propagate the request to the other MA s in order to find the fastest path to reach all the MAs on the network send a response including SeD reference s to the MA from which it received the request and forward the responses from other MAs it has reached first back to the MA that reached first this MA 7 12 2 2 Interfacing JXTA and DIET with JNI JNI is a technology allowing programmers to call native methods written in C C from a program written in java As seen before the DIET components having a DIET part and a JXTA part are the MA and the SeD The MA pie To submit the client s requests to DIET the MA needs to call the MApDie submit function To allow this the MA launches a MA pie via a native method and calls the submit function via another The MA piet contains
19. we will present the current version of DAGDA which is an alternative data manager for DIET with several advanced data management features 15 1 Overview DAGDA allows data explicit or implicit replications and advanced data management on the erid It was designed to be backward compatible with previously developed applications for DIET which benefit transparently of the data replications Moreover DAGDA limits the data size loaded in memory to a user fixed value and avoids CORBA errors when transmitting too large data regarding to the ORB configuration DAGDA offers a new way to manage the data on DIET The API allows the application developer to replicate move add or delete a data to be reused later or by another appli cation Each component of DIET can interact with DAGDA and the data manipulation can be done from a client application a server or an agent through a plug in scheduler A DAGDA component is associated to each element in a DIET platform client Master Agent Local Agent SeD These components are connected following the DIET deploy ment topology Figure 15 1 shows how the DAGDA and DIET classical components are connected In contrary of a DIET architecture each DAGDA component has the same role It can store transfer or move a data The client s DAGDA component is isolated of the architecture and communicates only with the chosen SeDs DAGDA components when necessary When searching for a data DAGDA uses its hierarchical top
20. 7 5 4 Easy definition of a new scheduler class 65 7 5 5 Creation and usage of a scheduler module 70 7 5 6 SeD plugin schedulers and agent schedulers interactions 71 7 5 7 A complete example of scheduler 71 Tio Futra Work s Lu ee sa ue dada s ea tam La mm a ae ga 73 8 Performance prediction 75 Bol ARTO RO do seg ce 2 LS pal cee ge SN Re de DUR we Ree eS 75 8 2 FAST Fast Agent s System Timer 75 22 1 Bulding PAST issos 4 Lun RR eue 4 76 8 2 2 Using FAST in the plug in scheduler 77 8 2 3 Building a server application with FAST 77 8 2 4 Example with convertors 78 8 3 CoRI Collectors of Ressource Information 79 8 3 1 Functions and tags e s oos 4 s 4 miiir eee eee eee 80 Oe PAST s gees pee a dus end d an nou ee a 80 S39 CORRESP 4 a wie bun eo app Re monde Re Me Anais 81 Bol COR Baten pa ie pis Yee Pa a doe he Lt e See Bee E 81 SA Future Work s s co 648 ee eA BOE Se eA ange a eH ewe QE 81 9 Deploying a DIET platform 83 O41 Deployment Dass L sack aos a Gok pa ok BO Pate A 83 Od Using CORBA vce VOR a ee eek Ge ee ee Be 83 O12 DIET configuration Me sa oia 0 ous ee Re we due eR ee ee 85 g1 Example wey ikke a a RAR NS Rue HE 89 UL GODET o s re a A e OS ie A ee eee ee da 91 10 DIET da
21. Darwin 8 and on Windows Cygwin 1 5 25 and Cygwin 1 7 1 2 1 5 Software dependencies As explained in Section 1 2 CORBA is used for all communications inside the platform The implementations of CORBA currently supported in DIET is omniORB 4 which itself depends on Python NB We have noticed that some problems occur with Python 2 3 the C code generated by idl could not be compiled It has been patched in DIET but some warnings may still appear omniORB 4 itself also depends on OpenSSL in case you wish to secure your DIET plat form If you want to deploy a secure DIET platform SSL support is not yet implemented in DIET but an easy way to do so is to deploy DIET over a VPN In order to deploy CORBA services with omniORB a configuration file and a log direc tory are required see Section 9 1 1 for a complete description of the services Their paths can be given to omniORB either at runtime through the well known environment variables OMNIORB_CONFIG and OMNINAMES LOGDIR and or at omniORB compile time with the with omniORB config and with omniNames logdir options Some examples provided in the DIET sources depend on the BLAS and ScaLAPACK libraries However the compilation of those BLAS and ScaLAPACK dependent examples are optional 2 2 Compiling the platform DIET compilation process moved away from the traditional autotools way of things to a tool named cmake mainly to benefit from cmake s built in regression tests mechanism
22. Data replication After a data has been added to the DAGDA hierarchy the users can choose to replicate it explicitely on one or several DIET nodes With the current DAGDA version we allow to choose the nodes where the data will be replicated by hostname or DAGDA component ID In future developments it will be possible to select the nodes differently To maintain backward compatibility the replication function uses a C string to define the replication rule dagda_replicate_data const char id const char rule The replication rule is defined as follows lt Pattern target gt lt identification pattern gt lt Capacity overflow behavior gt e The pattern target can be ID or host e The identification pattern can contain some wildcards characters for example lyon grid5000 fr is a valid pattern e The capacity overflow behavior can be replace or noreplace replace means the cache replacement algorithm will be used if available on the target node a data could be deleted from the node to leave space to store the new one noreplace means that the data will be replicated on the node if and only if there is enough storage capacity on it For example host capricorne lyon replace is a valid replication rule 15 5 On correct usage of DAGDA Some things to keep in mind when using DAGDA as data manager for DIET e All the data managed by DAGDA are entirely managed by DAGDA
23. JuxMem allows DIET to store data in the JuxMem service Correspondingly the capability to log and visualize these transfers has been added to VizDIET VizDIET is available for download See the web page http graal ens lyon fr DIET vizdiet html for more information INRIA ENS Lyon UCBL Page 100 DIET User s Manual BOSE E Chapter 11 Multi MA extension The hierarchical organization of DIET is efficient when the set of resources is shared by few individuals However the aim of grid computing is to share resources between several individuals In that case the DIET hierarchy become inefficient The Multi MA extension has been implemented to resolve this issue This chapter explains the different scalability issues of grid computing and how to use the multi MA extension to deal with them 11 1 Function of the Multi MA extension The use of a monolithic architecture become more and more difficult when the number of users and the number of resources grow simultaneously When a user tries to resolve a problem without the multi MA extension DIET looks for the better SeD that can solve it This search involves the fact that each SeD has to be queried to run a performance prediction as described in Section 1 4 The need to query every SeD that can resolve a problem is a serious scalability issue To avoid it the multi MA extension proposes to interconnect several MA together So instead of having the whole set of SeD available under
24. Li LL di da aa hee ds Pa 28 J42 Access TNCHONS oe s 24 45 40 kOe du Ree Se EE we Se ERS bn 28 3 5 Data Management functions 29 Souk Pres functions e s sors sde du ha RE ee ee ee due ek ee RE 30 0 Problem description oa so sarau wb hat aie te e mue tu dou r ie 31 Gat o LL eg ae LOL hee SEINE SR el ee a A ee 31 3 7 1 Example 1 without persistency 31 3 7 2 Example 2 using persistency 32 4 Building a client program 35 4 1 Structure of a client program 35 42 Chat APL seee g i pe e a h Bee ES 36 AS EXamples ouch ke ace Bon aa Ae ae ee ge Ee hope eee Ge 36 18 1 Synchronous all cues eri De 4 RES Ee wa ee AE 36 43 2 Asynchronous call Len gi kG ba BP Ge 48 ow le Ale GeO 37 AA COMPIM IONM A ante Li ala tk ES SE A ad ae a OR PA eS 39 44 1 Compilation USE emake 0204 5 ee 4 kei Re Roe matins ae 39 5 Building a server application 41 5 1 Structure of the program o 4 2 du pocs ee ee ee Al 62 Semen AP l 4 4 Gh beeen ts due a a de dei Ee de ee 42 Oo PAMPA oil LEE ke Dam RNA EEE EK Rita tea 43 SA Compilation e i sosse Lao be ee ee heed a DORE Neue 45 6 Batch and parallel submissions 47 BL Introduction ARE eS ewe Pe ee EY 47 0A TEMOR op oe a Oe a ee ep a ee 47 63 Configuration for compilation 1 0 44 44 dea ds daa a ue RO 48 OA Paralleleystems 4
25. Page 119 DIET User s Manual BOSE E workflow engine only processors containing a lt diet gt tag can be used The lt diet gt tag contains the path attribute that matches exactly the DIET service name and option nally contains the estimation attribute with value keyword constant whenever the computation time estimation for this service does not depend on input data using this option may reduce considerably the load on the DIET platform because the request for performance estimation is done only once by the MaDag instead of being done for each task Processor ports Processor input and output ports are named and declared A port may be an input lt in gt tag or an output out tag For each input output the following attributes can be defined e type mandatory contains the base type of data i e a basic type identifier that describes the type of the data received generated by the port When data is scalar this is the actual data type when data is an array this is the type of the data leaves of the array e depth optional default is 0 contains the depth of the array if applicable e cardinality optional only for out ports with depth gt 0 contains the number of elements of the generated array This value can be provided only if it is a constant i e the number of elements does not vary for each instance of data When the data depth is greater than 1 the format for the cardinality attribute is a column
26. SeD e Mode Batch e Type String e Description Path to an NFS directory where you have read write rights pathToTmp e Component SeD e Mode Batch e Type String e Description Path to a temporary directory where you have read write rights internOARbatchQueueName e Component SeD e Mode Batch e Type String e Description only useful when using CORI batch features with OAR 1 6 initRequestID e Component MA e Mode All e Type Integer e Description When a request is sent to the Master Agent a request ID is associated and by default it begins at 1 If this parameter is provided it will begins at initRequestID ackFile e Component Agent and SeD e Mode Acknowledge file e Type String INRIA ENS Lyon UCBL Page 153 DIET User s Manual BOSE TE Description Path to a file that will be created when the element is ready to execute maxMsgSize Component All Mode DAGDA Type Integer Description The maximum size of a CORBA message sent by DAGDA By default this value is equal to the omniORB giopMaxMsgSize size maxDiskSpace Component All Mode DAGDA Type Integer Description The maximum disk space used by DAGDA to store the data If set to 0 DAGDA will not take care of the disk usage By default this value is equal to the available disk space on the disk partition chosen by the storageDirectory option maxMemSpace e Component All e Mode DAGDA e Type Integer e Descri
27. XML and described in the section 14 4 1 Because of large amounts of computations and data involved in some workflow appli cations the number of tasks in a DAG can grow very fast The need for a more abstract way of representing a workflow that separates the data instances from the data flow has led to the definition of a functional workflow language called the Gwendia language A complex application can be defined using this language that provides data operators and control structures if then else loops To execute the application we need to provide both the workflow description see 14 4 2 and a file describing the input data set The DIET workflow engine will instanciate the workflow as one or several tasks DAGs sent to the MApac agent to be executed in the DIET platform INRIA ENS Lyon UCBL Page 115 DIET User s Manual BOSE E 14 2 Quick start Requirements and compilation The workflow supports in DIET needs the following e The Xerces library the XML handling code is written with Xerces C using the provided DOM API e The XQilla library the conditions in conditional or looping workflow structures are written in XQuery language and parsed using the XQilla library e Enable the workflow support when compiling DIET In order to build DIET with workflow support using cmake three configuration parameters need to be set DIET_USE_WORKFLOW as follow DDIET_USE_WORKFLOW BOOL 0N XERCES_DIR d
28. a standard DIET request that will use the scheduler configured by the SeD e A Multi HEFT scheduler option heft this scheduler applies the HEFT heuristic to all workflows submitted by different clients to the MApac This means that the priorities assigned by the HEFT heuristic are used to order the tasks of all dags processed by the MApac and following this order the tasks are mapped to the first available ressource e A Multi AgingHEFT scheduler option aging heft this scheduler is similar to Multi HEFT but it applies a correction factor to the priorities calculated by the HEFT algorithm This factor is based on the age of the dag ie the time since it was submitted to the scheduler Compared to Multi HEFT this scheduler will increase the priority of the tasks of a workflow that has been submitted earlier than other dags e A FOFT Fairness on Finish Time scheduler option fairness this scheduler uses another heuristic to apply a correction factor to the priorities calculated by the HEFT algorithm This factor is based on the slowdown of the dag that is calculated by comparing the earliest finish time of the tasks in the same environment without any other concurrent workflow and the actual estimated finish time 14 6 2 SeD requirements for workflow scheduling The workflow schedulers Basic Multi HEFT Multi AgingHEFT and FOFT use in formation provided by the SeDs to be able to run the HEFT heuristic So the SeD programmer must pro
29. add_service std string dyn_add_rem_ toString NB c_str std cout lt lt Services added lt lt std endl diet_print_service_table Removing std cout lt lt Removing service lt lt pb gt pb_name lt lt std endl ifdef HAVE_ALT_BATCH pb gt parallel_flag 1 endif diet_service_table_remove pb std cout lt lt Service removed lt lt std endl Print service table diet_print_service_table return 0 usage function int usage char cmd std cerr lt lt Usage lt lt cmd lt lt lt SeD cfg gt lt lt std endl return 1 add_service function declares SeD s service int add_service const char service_name diet_profile_desc_t profile NULL unsigned int pos 0 Set profile parameters profile diet_profile_desc_alloc strdup service_name 0 0 0 diet_generic_desc_set diet_param_desc profile pos DIET_SCALAR DIET_INT Add service to the service table if diet_service_table_add profile NULL service return 1 Free the profile since it was deep copied INRIA ENS Lyon UCBL Page 144 DIET User s Manual BOSE E diet_profile_desc_free profile std cout lt lt Service lt lt service_name lt lt added lt lt std endl return 0 int checkUsage int argc char argv if argc 2 usage argv 0 exit 1 return 0
30. are two primary efforts for the CoRI manager e Improving CoRI Easy Some evaluation functions are very basic and should be revised to increase their response time speed and the accuracy of the information There is a need for other information 1 e information about the network Every operating systems provide other basic functions to get the information CoRI Easy doesn t know all functions Use the diet_estimate_cori_print function to test what CoRI Easy can find on your SeD Send us a mail if not all functions are working properly e Improving CoRI batch add new functionalities to access dynamic information as well as some kind of performance predictions for more batch systems e New collectors Integrating other external tools like Ganglia 22 or Nagios 14 to the CoRI Manager can provide more useful and exact information INRIA ENS Lyon UCBL Page 81 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 82 DIET User s Manual BOSE E Chapter 9 Deploying a DIET platform Deployment is the process of launching a DIET platform including agents and servers For DIET this process includes writing configuration files for each element and launching the elements in the correct hierarchical order There are three primary ways to deploy DIET Launching by hand is a reasonable way to deploy DIET for small scale testing and verification This chapter explains the necessary services how to write DIET configuration
31. can be downloaded Otherwise don t do anything with this screen The last screen deals with username and password but these parameters are filled with default values e Second step registering a SeD to the MA Be sure that the parentName inside the configuration file matches the name of the MA pe previously launched The command to run is java cp lt JXTA_JARS gt JXTASeD lt DIET_SeD_config file gt lt computation_abilities gt If you want to put LA s between the MA and the SeD launch the following com mand before loading the SeD java LA lt DIET LA config file gt INRIA ENS Lyon UCBL Page 110 DIET User s Manual BOSE TE Check the DIET tree coherence and the parentName variables inside the configura tion files e Third step Launch a client with the command java cp lt JXTA_JARS gt JXTAClient lt pb gt At this point you still haven t tested the Multi MA To achieve this launch at least one MA and launch again the client Scripts have been left at your disposal You just need to check the environment variables and paths required As said before only one JXTA peer can be run in one directory so each script is inside a different one These directories have to be edited for configuration are named MMA1 MMA2 MMA3 LA1 SeD1 SeD2 and client and are located in lt DIET_root gt src examples JXTA scripts INRIA ENS Lyon UCBL Page 111 DIET User s Manual BOSE E TXTA Config
32. fast html a grid aware dynamic forecasting library Although the detection of FAST should be correctly handled by cmake since detection is based on the FAST provided fast config utility the installation of FAST can be a lengthy process and depending on your platform potentially quite difficult This is due to the dependency of FAST towards numerous sub libraries on which it relies GSL BDB NWS LDAP Thus the activation of this option can only be recommended for advanced users As already http www netlib org blas INRIA ENS Lyon UCBL Page 19 DIET User s Manual BOSE E mentioned on activation of the DIET_USE_FAST option cmake will search among the well known system path for the fast config command and set the FAST_CONFIG_EXECUTABLE with the result Upon failure it is up to the user to manually set the full path name to this command e g with cJcmake command line argument DFAST_CONFIG_EXECUTABLE PATH HOME 1local bin fast config e DIET_USE_FD for activating Fault Detector e DIET_USE_JUXMEM activates DIET support of JuxMem which allows the user to manage persistent data When this option is activated disabled by default a SeD can store data blocks within JuxMem Chapter 13 describes in more details JuxMem and its use inside DIET e DIET_USE_JXTA activates the so called MULTI Master Agent support This option is which is based on the JXTA layer refer to http www jxta org allows the user to de pl
33. files and in what order DIET elements should be launched See Section 9 1 for details GODIET is a Java based tool for automatic DIET deployment that manages con figuration file creation staging of files launch of elements monitoring and reporting on launch success and process cleanup when the DIET deployment is no longer needed See Section 9 2 for details Writing your own scripts is a surprisingly popular approach This approach often looks easy initially but can sometimes take much much longer than you predict as there are many complexities to manage Learn GODIET it will save you time 9 1 Deployment basics 9 1 1 Using CORBA CORBA is used for all communications in DIET and for communications between DIET and accessory services such as LogService VizDIET and GODIET This section gives basic information on how to use DIET with CORBA Please refer to the documentation of your ORB if you need more details The naming service DIET uses a standard CORBA naming service for translating an user friendly string based name for an object into an Interoperable Object Reference IOR that is a globally unique identifier incorporating the host and port where the object can be contacted The naming service in omniORB is called omniNames and it must be launched before any INRIA ENS Lyon UCBL Page 83 DIET User s Manual BOSE E other DIET entities DIET entities can then locate each other using only a string based name and the
34. function has been redefined an alternative method for aggregation is needed Currently a basic priority scheduler has been implemented enabling an application developer to specify a series of performance values that are to be optimized in succession A developer may implement a priority scheduler using the following interface diet_aggregator_desc_t diet_profile_desc_aggregator diet profile desc t profile int diet_aggregator_set_type diet_aggregator_desc_t agg diet_aggregator_type_t atype int diet aggregator priority max diet aggregator desc t agg diet _est_tag_t tag int diet _aggregator priority min diet aggregator desc t agg diet est tag t tag int diet aggregator priority maxuser diet aggregator desc t agg int val int diet _aggregator priority minuser diet aggregator desc _t agg int val The diet profile desc aggregator and diet aggregator set type functions fetch and con figure the aggregator corresponding to a DIET service profile respectively In particular a priority scheduler is declared by invoking the latter function with DIET AGG PRIORITY as the agg parameter Recall that from the point of view of an agent the aggregation phase is essen tially a sorting of the server responses from its children A priority scheduler logically uses a series of user specified tags to perform the pairwise server comparisons needed to construct the sorted list of server responses To define the tags and the order
35. given as an argument to the start option of omniNames You also need to update your LD_LIBRARY_PATH to point to lt install_dir gt lib So your LD_LIBRARY_PATH environment variable should now be LD_LIBRARY_PATH lt omniORB_home gt lib lt install_dir gt lib NB1 In order to avoid name collision every agent must be assigned a different name in the name server since they don t have any children SeDs do not need names assigned to them and they don t register with the name server NB2 Each DIET hierarchy can use a different name server or multiple hierarchies can share one name server assuming all agents are assigned unique names In a multi MA environment in order for multiple hierarchies to be able to cooperate it is necessary that they all share the same name server 9 1 2 DIET configuration file A configuration file is needed to launch a DIET entity Some fully commented examples of such configuration files are given in the directory src examples cfgs of the DIET source files and installed in lt install_dir gt etc Please note that e comments start with and finish at the end of the current line e meaningful lines have the format keyword value following the format of con figuration files for omniORB 4 e for options that accept 0 or 1 0 means no and 1 means yes and e keywords are case sensitive lif there isn t lt install_dir gt etc directory please configure DIET with enable examples and o
36. if data inputs have indexes 1 and 2 with a maximum index of 3 for the right input then the output has the index 6 4 1 2 Note that this operator creates a synchronization constraint among all instances as the maximum index of the right input must be known by the workflow engine before being able to create new indexes e match groups 2 ports processes each data instance of the first port with all the data instances of the second port that have an index prefix that matches the first port s index for example if left data has index 1_1 it will be processed with all right data which have an index beginning with 1_1 The output index is the second port s index Here is an example of a Gwendia workflow to be continued with the links part below lt workflow gt lt interface gt lt constant name parameter type integer value 50 gt lt source name key type double gt lt sink name results type file gt lt interface gt lt processors gt lt processor name genParam gt lt in name paramKey type double gt lt out name paramFiles type file depth 1 gt lt diet path gen estimation constant gt lt processor gt lt processor name docking gt lt in name param type integer gt lt in name input type file gt lt out name result type double gt lt iterationstrategy gt lt cross gt lt port name param gt lt port name input gt lt cross gt lt it
37. in which they should be compared four functions are introduced These functions of the form diet_aggregator_priority_ serve to identify the estimation values to be optimized during the aggregation phase The min and max forms indicate that a standard performance metric e g time elapsed since last execution from the INRIA ENS Lyon UCBL Page 57 DIET User s Manual BOSE E diet_estimate_lastexec function is to be either minimized or maximized respectively Sim ilarly the minuser and _maxuser forms indicate the analogous operations on user supplied estimation values Calls to these functions indicate the order of precedence of the tags Each time two server responses need to be compared the values associated with the tags specified in the priority aggregator are retrieved In the specified order pairs of corresponding values are successively compared passing to the next tag only if the values for the current tag are identical If one server response contains a value for the metric currently being compared and another does not the response with a valid value will be selected If at any point during the treatment of tags both responses lack the necessary tag the comparison is declared indeter minate This process continues until one response is declared superior to the other or all tags in the priority aggregator are exhausted and the responses are judged equivalent 7 4 Example A new example has been added to the DIET di
38. of a particular server in the context of a particular client request for that task These estimates are passed to the server s parent agent agents then sort these responses in a manner that optimizes certain performance criteria Effectively candidate SeDs are identified through a distributed scheduling algorithm based on pairwise comparisons between these performance estimations upon receiving server responses from its children each agent performs a local scheduling operation called server response aggre gation The end result of the agent s aggregation phase is a list of server responses from servers in the subtree rooted at said agent sorted according to the aggregation method in effect By default the aggregation phase implements the following ordered sequence of tests 1 FAST NWS data SeDs compiled and properly configured with FAST 21 and NWS 26 are capable of making dynamic performance estimates If such data were generated by the SeDs these are the metrics on which agents select servers 2 Round robin In the absence of application and platform specific performance data the DIET scheduler attempts to probabilistically achieve load balance by assigning client requests on a round robin basis Essentially each server records a timestamp indicating the last time at which it was assigned a job for execution Each time a request is received the SeD computes the time elapsed since its last execution and among the responses it
39. offers all the same sub services Two clients are designed to use these services one pdgemm_client c is designed to use the pdgemm_ function only and the other one client c to use all ScaLAPACK functions and sub services such as MatPROD e workflow The programs in this directory are examples that demonstrate how to use the workflow feature of diet The files representing the workflows that can be tested are stored in xml sub directory For each workflow you can find the required services in the corresponding xml file check the path attribute of each node element For the scalar manipulation example you can use scalar_server that gathers four different elementary services 2 3 1 Compiling the examples Cmake will set the examples to be compiled when setting the DIET_BUILD_EXAMPLES to ON which can be achieved by toggling the corresponding entry of ccmake GUI s or by adding DDIET_BUILD_EXAMPLES BOOL ON to the command line arguments of c cmake invocation Note that this option is disabled by default The compilation of the examples respectively the installation is executed on the above described invocation of make resp make install stages The binary of the examples are placed in the lt install_dir gt bin examples sub directory of the installation directory Like wise the samples of configuration files located in src examples cfgs are processed by make install to create ready to use configuration files in src examples cfgs a
40. port To use it to define a source or a sink value it must be prefixed with the node id For example if the source of the input port in3 is the port out2 of the node n1 than the element must be described as follow lt in name in3 type DIET_INT source ni out2 gt The link between input and output ports must be described either by a source value in the jing element or by a sink value in the out element Specifying both does not cause an error but duplicates the information The example shown in Figure 14 1 can be represented by this XML description INRIA ENS Lyon UCBL Page 118 DIET User s Manual BOSE E lt dag gt lt node id n1 path succ gt lt arg name ini type DIET_INT value 56 gt lt out name out1i type DIET_INT gt lt out name out2 type DIET_INT gt lt node gt lt node id n2 path double gt lt in name in2 type DIET_INT source ni out1 gt lt out name out3 type DIET_INT gt lt node gt lt node id n3 path double gt lt in name in3 type DIET_INT source ni out2 gt lt out name out4 type DIET_INT gt lt node gt lt node id n4 path sum gt lt in name in4 type DIET_INT source n2 out3 gt lt in name in5 type DIET_INT source n3 out4 gt lt out name out4 type DIET_INT gt lt node gt lt dag gt 14 4 2 Gwendia language The Gwendia language is written in XML and validated by the workflow parser if the pat
41. receives DIET agents select SeDs with a longer elapsed time INRIA ENS Lyon UCBL Page 53 DIET User s Manual BOSE E 3 Random If the SeD is unable to store timestamps the DIET scheduler will chose randomly when comparing two otherwise equivalent SeD performance estimations Warning If DIET is compiled with option DIET_USE_CORI FAST NWS Scheduling is de activated See Chapter 8 for more information about CoRl In principle this scheduling policy prioritizes servers that are able to provide useful perfor mance prediction information as provided by the FAST and NWS facilities In general this approach works well when all servers in a given DIET hierarchy are capable of making such es timations However in platforms composed of SeDs with varying capabilities load imbalances may occur since DIET systematically prioritizes server responses containing FAST and or NWS data servers that do not respond with such performance data will never be chosen We have designed a plugin scheduler facility to enable the application developer to tailor the DIET scheduling to the targeted application This functionality provides the application developer the means to extend the notion of a performance estimation to include metrics that are application specific and to instruct DIET how to treat those data in the aggregation phase We describe these interfaces in the following sections 7 3 Plugin Scheduler Interface Distributed application
42. that will take into account the heterogeneity of the target architecture Grid extensions to classical batch processing provide an alternative approach with projects like Condor G 5 or Sun GridEngine 9 Finally peer to peer 18 or Global computing 7 can be used for fine grain and loosely coupled applications A second approach provides a semi transparent access to computing servers by submitting jobs to servers offering specific computational services This model is known as the Applica tion Service Provider ASP model where providers offer not necessarily for free computing resources hardware and software to clients in the same way as Internet providers offer network resources to clients The programming granularity of this model is rather coarse One of the advantages of this approach is that end users do not need to be experts in parallel program ming to benefit from high performance parallel programs and computers This model is closely related to the classical Remote Procedure Call RPC paradigm On a Grid platform RPC or GridRPC 13 15 offers easy access to available resources from a Web browser a Problem Solving Environment PSE or a simple client program written in C Fortran or Java It also provides more transparency by hiding the selection and allocation of computing resources We favor this second approach In a Grid context this approach requires the implementation of middleware to facilitate client access to remote r
43. the corresponding dagda_wait_ function The arguments of these functions are the same than for the synchronous dagda_get_ functions See Section 15 4 2 for more details dagda_wait_scalar unsigned int transferRef void value diet_base_type_t base_type dagda_wait_vector unsigned int transferRef void value diet_base_type_t base_type size_t size dagda_wait_matrix unsigned int transferRef void value diet_base_type_t base_type size_t nb_r size_t nb_c diet_matrix_order_t order INRIA ENS Lyon UCBL Page 137 DIET User s Manual BOSE E dagda_wait_string unsigned int transferRef char value dagda_wait_file unsigned int transferRef char path A plugin scheduler developer often wants to make an asynchronous data transfer to the local DIET node Problems can arise if you want to wait the completion of the tranfer before returning But with the previously defined functions DAGDA maintains a reference to the transfer thread which will be released after a call to the waiting function To avoid DAGDA to keep infinitely these references the user should call the dagda_load_ functions instead of the dagda_get_ _async ones dagda_load_scalar char ID dagda_load_vector char ID dagda_load_matrix char ID dagda_load_string char ID dagda_load_file char ID 15 4 4 Data checkpointing with DAGDA DAGDA allows the SeD administrator to choose a file where DAGDA
44. the name of the MA plus the number of the session A file is created only when there are some persistent data in the session For example ID_FILE MA1 1 means the identifiers of the persistent data stored are in the file corresponding to the first session in the Master Agent MA1 The file is stored in the tmp directory NB for the moment when a data item is erased from the platform the file isn t updated 3 4 Manipulating DIET structures The user will notice that the API to the DIET data structures consists of modifier and accessor functions only no allocation function is required since diet_profile_alloc see Section 3 6 allocates all necessary memory for all argument descriptions This avoids the temptation for the user to allocate the memory for these data structures twice which would lead to DIET INRIA ENS Lyon UCBL Page 27 DIET User s Manual BOSE E errors while reading profile arguments Please see the example in Section 3 7 for a typical use Moreover the user should know that arguments of the _set functions that are passed by pointers are not copied in order to save memory This is true for the value arguments but also for the path in diet_file_set Thus the user keeps ownership of the memory zones pointed at by these pointers and he she must be very careful not to alter it during a call to DIET 3 4 1 Set functions On the server side these functions should not be used on arguments but only
45. to use a data item that is already stored in the platform diet_use_data diet_arg_t arg char arglD INRIA ENS Lyon UCBL Page 29 DIET User s Manual BOSE TE This function replaces the set functions see Section 3 4 1 NB a mechanism for data identifier publication hasn t been implemented yet So ex changes of identifiers between end users that want to share data must be done ex plicitly e The diet_free_persistent_data method allows the client to remove a persistent data item from the platform diet_free_persistent_data char argID adolfo k kkk kk Add handler argID and text message msg in the identifier file EEEE EEEE EEEE EEE RO RR E k kkk k kkk kkk kk void store_id char argID char msg sets only identifier data is present inside the platform void diet_use_data diet_arg_t arg char argID DCR OO IR OK a k kkk k kkk kk kkk Free persistent data identified by argID BOO RO I I a kk 2k 2 2k 2k 2k 2k 2 kkk int diet_free_persistent_data char argID 3 5 1 Free functions The amount of data pointed at by value fields should be freed through a DIET API function DCR OO RR OR AR LL RK RK kK ak kkk kkk kk Free the amount of data pointed at by the value field of an argument This should be used ONLY for VOLATILE data on the server for IN arguments that will no longer be used on the client for OUT arguments afte
46. using the ID D is obtained from DAGDA The value argument is initialized with a pointer to the first vector element The base_type content are initialized with the base type and size of the vector These two parameters can be set to NULL if the user does not take care about it INRIA ENS Lyon UCBL Page 135 DIET User s Manual BOSE TE dagda_get_matrix char ID void value diet_base_type_t base_type size_t nb_r size_t nb_c diet_matrix_order_t order The matrix using the ID D is obtained from DAGDA The value argument is initialized with a pointer to the first matrix element The base_type nb_r nb_c and order arguments contents are repectively set to the base type of the matrix the number of rows the number of columns and the matrix order All of these parameters can be set to NULL if the user does not take care about it dagda_get_string char ID char value The string of ID D is obtained from DAGDA and the value content is set to a pointer on the first string character dagda_get_file char ID char path The file of ID JD is obtained from DAGDA and the path content is set to a pointer on the first path string character 15 4 3 Asynchronous data transfers With DAGDA there is two way to manage the asynchronous data transfers depending of the data usage e With end of transfer control DAGDA maintains a reference to the transfe
47. void diet_print_service_table The parameter diet_solve_t solve func is the type of the solve service function a function pointer used by DIET to launch the computation The parameter diet_convertor t cvt is to be used in combination with FAST if avail able It is there to allow profile conversion for multiple services or when differences occur between the DIET and the FAST profile Profile conversion is complicated and will be treated separately in Chapter 8 5 3 Example Let us consider the same example as in Chapter 4 where a function scal_mat_prod performs the product of a matrix and a scalar and returns the time required for the computation INRIA ENS Lyon UCBL Page 43 DIET User s Manual BOSE E int scal_mat_prod double alpha double M int nb_rows int nb_cols float time Our program will first define the solve function that consists of the link between DIET and this function Then the main function defines one service and adds it in the service table with its associated solve function include DIET_server h include scal_mat_prod h int solve_smprod diet_profile_t pb double alpha double M float time size_t m n int res Get arguments diet_scalar_get diet_parameter pb 0 amp alpha NULL diet_matrix_get diet_parameter pb 1 amp M NULL amp m amp n NULL Launch computation res scal_mat_prod alpha M m n amp time Set OUT arguments die
48. 5 IEA ime Start as aoe 24 ee A A A A mas 116 14 3 Sofiware architecte L ee a de mu ee EN Ee Ree Bes 117 14 4 Workflow description languages 118 14 41 MaDas language osor 44 eee ba sun wee RE RR a RES 118 14 4 2 Gwendia language 119 14 5 Client APT 2 hg 404 si 44 eR ee ee ee ue et 124 14 51 Steucture of client program ic ome Thoe mi bah A L eau hs 124 145 2 The simplest example so 4 4 245 auia aa dau ga nada 4 a 124 HO a SL dE Line Le Fee ee ne ENTER Mans die 127 14 6 1 Available schedulers 128 14 6 2 SeD requirements for workflow scheduling 128 15 DAGDA extension 131 Tal Overview lt a sak dis pa Ming E ey BO ek ee UNE Boe ek re ES 131 15 2 The DAGDA configuration options ca cos k es e ee ee ua a 133 15 3 Cache replacement algorithm 134 15 4 The DACDAAP oie rrara a tr RA AAA 134 15 4 1 Note on the memory management 134 15 4 2 Synchronous data tramsfers s o ios eo sn aa a na a 134 15 4 3 Asynchronous data transfers 136 15 4 4 Data checkpointing with DAGDA 138 15 4 5 Create data ID alases o 4 so 4 4 4 ke eR RE RRR Re 138 1546 Data replication 2 i622 2 48644 he ooo oe eke eG who ee 139 15 5 On correct usage of DAGDA oc aa oa nd aus a aOR a we 139 toG Wire Works ca
49. AGENT dietHostname ma delta com bindServicePort 2001 neighbours ma gamma com 6000 11 3 Search examples The following section explains how a diet_call is managed when used on the previous architecture If a client sends a diet_call for the problem a to the alpha s MA the alpha s MA will return a reference of one of it s SeD However if its scheduler see Section 7 says that no SeD is available it will forward the request to beta and gamma If beta has an available SeD it will be used to resolve the problem If not the request is forwarded to delta Now if a client performs a diet_call for the problem c to the delta s MA the delta MA does not have a SeD that can resolve this problem So it forwards the request to gamma If gamma has no available SeD the request is forwarded to alpha and beta INRIA ENS Lyon UCBL Page 103 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 104 DIET User s Manual BOSE E Chapter 12 P2P DIET extension DIET y To extend the field of the available services for each client in a transparent manner DIET uses the Multi Agent system to increase scalability To achieve this the MAs access each others resources when processing a client s request Thus each request is not only submitted inside the hierarchy of the MA contacted by the client but also inside the hierarchy of each MAs connected to the first MA if the first submission failed 12 1 P2P and JXTA One
50. API The client API follows the GridRPC definition 24 all diet_ functions are duplicated with grpc_functions Both diet_initialize grpc_initialize and diet_finalize grpc_finalize belong to the GridRPC API A problem is managed through a function_handle that associates a server to a problem name For compliance with GridRPC DIET accepts diet_function_handle_init but the server specified in the call will be ignored DIET is designed to automatically select the best server The structure allocation is performed through the function diet function handle default The function_handle returned is associated to the problem description its profile in the call to diet_call 4 3 Examples Let us consider the same example as in Section 3 7 but for synchronous and asynchronous calls Here the client configuration file is given as the first argument on the command line and we decide to hardcode the matrix its factor and the name of the problem 4 3 1 Synchronous call smprod for scalar by matrix product tinclude lt stdio h gt tinclude lt stdlib h gt include lt math h gt include DIET_client h int main int argc char argv int i double factor M_PI Pi why not double matrix The matrix to multiply float time NULL To check that time is set by the server diet_profile_t profile Allocate the matrix 60 lines 100 columns matrix malloc 60 100 sizeof double Fill in th
51. ARGETS simple_server DESTINATION bin and it s associated simple client ADD_EXECUTABLE simple_client simple_client c TARGET_LINK_LIBRARIES simple_client DIET_CLIENT_LIBRARIES INSTALL TARGETS simple_client DESTINATION bin In order to test drive the cmake configuration of this example and assuming the DIET_HOME points to a directory containing an installation of DIET simply try export DIET_HOME lt path_to_a_DIET_instal_directory gt cd doc ExternalExample mkdir Bin cd Bin cmake DDIET_DIR PATH DIET_HOME DCMAKE_INSTALL_PREFIX PATH tmp DIETSimple make make install INRIA ENS Lyon UCBL Page 39 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 40 DIET User s Manual BOSE TE Chapter 5 Building a server application A DIET server program is the link between the DIET Server Deamon SeD and the libraries that implement the service to offer 5 1 Structure of the program As for the client side the DIET SeD is a library So the server developer needs to define the main function Within the main the DIET server will be launched with a call to diet_SeD which will never return except if some errors occur The complete server side interface is described in the files DIET_data h see Chapter 3 and DIET_server h found in lt install_dir gt include Do not forget to include the DIET server h DIET server h includes DIET data h at the beginning of your server source code inclu
52. ATCH x void diet_set_server_status diet_server_status t st Set the nature of the service to be registered to the SeD int diet _profile_desc_set __sequential diet_profile_desc_t profile INRIA ENS Lyon UCBL Page 49 DIET User s Manual BOSE E int diet_profile_desc_set_parallel diet_profile_desc_t profile A service MUST call this command to perform the submission to the batch system int diet_submit_parallel diet_profile_t profile const char command 6 8 1 Registering the service A server is mostly built like described in section 5 In order to let the SeD know that the service defined within the profile is a parallel one the SeD programmer must use the function void diet_profile_desc_set_parallel diet_profile_desc_t profile By default a service is registered as sequential Nevertheless for code readability reasons we also give the pendant function to explicitly register a sequential service void diet_profile_desc_set_sequential diet_profile_desc_t profile 6 8 2 Server configuration file The programmer of a batch service available in a SeD has not to worry to which batch system to submit except for its name because DIET provides all the mechanisms to transparently submit the job to them DIET is able to submit batch scripts to OAR version 1 6 and 2 0 PBS Torque and loadleveler The name of the batch scheduler managing the parallel resource where the SeD is running has to
53. D_LIBRARY_ PATH value gt lt env gt lt compute gt lt cluster label res3 disk disk 3 login bar gt lt env gt lt var name PATH value gt lt var name LD_LIBRARY_PATH value gt lt env gt lt node label res3_host1 gt lt ssh server hostl res3 fr gt lt end_point contact 192 5 80 103 gt lt node gt lt node label res3_host2 gt lt ssh server host2 res3 fr gt lt node gt lt cluster gt lt resources gt lt diet_services gt lt omni names contact res1_IP port 2121 gt lt config server resl remote_binary omniNames gt lt omni_names gt lt diet_services gt lt diet_hierarchy gt lt master_agent label MA gt lt config server resl remote binary dietAgent gt lt cfg options gt lt option name traceLevel value 1 gt lt cfg options gt lt SeD label SeD1 gt lt config server res2 remote_binary server_dyn_add_rem gt lt cfg_options gt lt option name traceLevel value 1 gt lt cfg options gt lt SeD gt lt SeD label SeD2 gt lt config server res3_hostl lt cfg_options gt lt option name traceLevel value 30 gt lt cfg_options gt lt parameters string T gt lt SeD gt lt SeD label SeD3 gt lt config server res3_host2 lt cfg options gt lt option name traceLevel
54. MNIORB4_DIR PATH HOME local omni0RB 4 0 7 e DIET_BUILD_EXAMPLES activates the compilation of a set of general client server exam ples Note that some specific examples e g DIET_BUILD_BLAS_EXAMPLES require some additional flag to be activated too INRIA ENS Lyon UCBL Page 18 DIET User s Manual BOSE E e DIET_BUILD_LIBRARIES which is enabled by default activates the compilation of the DIET libraries Disabling this option is only useful if you wish to restrict the compilation to the construction of the documentation 2 2 5 DIET s extensions configuration flags DIET has many extensions some of them are still experimental These extensions most often rely on external packages that need to be pre installed One should notice that some of those extensions offer concurrent functionalities This explains the usage of configuration flags in order to obtain the compilation of the desired extensions e DIET_BUILD_BLAS_EXAMPLES option activates the compilation of the BLAS based DIET examples as a sub module of examples The BLAS Basic Linear Algebra Subpro grams are high quality building block routines for performing basic vector and matrix operations Level 1 BLAS do vector vector operations Level 2 BLAS do matrix vector operations and Level 3 BLAS do matrix matrix operations Because the BLAS are effi cient portable and widely available they re commonly used in the development of high quality linear algebra soft
55. Mem with the DJUXMEM_DIR option A JuxMem canonical installation is supposed with JXTA C and APR installed Note that APR Apache Portable Runtime is a requirement of both JuxMem C and JXTA C When DIET is configured to use JuxMem SeD are able to store data blocks inside JuxMem Please be carefull as it does not mean that you have a JuxMem platform deployed and usable In a first step you must deploy a JuxMem platform as described in the documentations of JuxMem This JuxMem platform is currently based on JuxMem INRIA ENS Lyon UCBL Page 113 DIET User s Manual BOSE E J2SE JuxMem C is only used to play the role of a JuxMem client within a DIET SeD Please read the README file of JuxMem to build and deploy a JuxMem platform 13 4 Example A simple example of the JuxMem usage inside DIET can be found in the dmat_manips sample The name of the client is clientJuxMem This example stores DIET matrices inside JuxMem and allows next computations to retrieve these matrices directly from JuxMem Clients therefore avoid unnecessary tranfers of matrices as they only need to transfer the ID of the data returned by JuxMem More documentation and examples will be available in the future 13 5 Troubleshooting If you encounter any problem you can try get help from the JuxMem discuss mailing list lt juxmem discuss lists gforge inria fr gt Do not forget to include in your e mails the exact error message your hardware description yo
56. ORBA have proven to be a good base for building applications that manage access to distributed services They not only provide transparent communications in heterogeneous networks but they also offer a framework for the large scale deployment of distributed applications Being open and language independent CORBA was chosen as the communication layer in DIET As recent implementations of CORBA provide communication times close to that of sockets CORBA is well suited to support distributed applications in a large scale Grid environment New specialized services can be easily published and existing services can also be used DIET INRIA ENS Lyon UCBL Page 12 DIET User s Manual BOSE TE is based upon OmniORB 3 17 or later a free CORBA implementation that provides good communication performance 1 3 DIET initialization Figure 1 2 shows each step of the initialization of a simple Grid system The architecture is built in hierarchical order each component connecting to its parent The MA is the first entity to be started 1 It waits for connections from LAs or requests from clients Figure 1 2 Initialization of a DIET system In step 2 an LA is launched and registers itself with the MA At this step of system initialization two kinds of components can connect to the LA a SeD 3 which manages some computational resource or another LA 4 to add a hierarchical level in this branch When the SeD registers
57. SE E diet_parallel_call diet_profile_t profile diet _error_t diet _sequential_call_async diet_profile_t profile diet_reqID_t reqID To explicitly call a parallel service in sync or async way diet_error_t diet_sequential_call diet_profile_t profile diet_error_t diet_parallel_call_async diet_profile_t profile diet_reqID_t reqID To mark a profile as parallel or sequential The default call to diet_call or diet_call_async will perform a call to the correct previous call x int diet _profile_set_parallel diet_profile_t profile int diet _profile_set_sequential diet_profile_t profile To let the user choose a given amount of resources int diet _profile_set_nbprocs diet_profile_t x profile int nbprocs 6 7 Batch server extended API and configuration file There are too many diverse scenarii about the communication and execution of parallel appli cations the code can be a MPI code or composed of different interacting programs possibly launched via ssh on every nodes input and output files can use NFS if this file system is present or they can be splitted and uploaded to each node participating to the calculus Then we will see what supplementary information has to be provided in the server con figuration file how to write a batch submission meta script in a SeD and how to record the parallel batch service 6 8 Server API Set the status of the SeD among SERIAL and B
58. The user don t have to free them DAGDA avoids memory leaks so the user does not have to worry about the memory management for the data managed by DAGDA e When using more than one DAGDA component on a node the user should define a different storage directory for each component For example the Master Agent and one SeD are launched on the same computer the user can define the storage directory of the Master Agent as tmp MA and the one for the SeD as tm p SeD1 Do not forget to create the directories before to use DAGDA This tip avoids many bugs which are really hard to find e The DAGDA API can be used to transfer the parameters of a service but it should not be used as this If an application needs a data which is only on the client the user should transmit it through the profile The DAGDA API should be used to share replicate or retrieve an existing data Using the API allows the user INRIA ENS Lyon UCBL Page 139 DIET User s Manual BOSE E to optimize their applications not to proceed to a diet_call even if it works fine Indeed the DAGDA client component is not linked to the DIET hierarchy so using the API to add a data and then to use it as a profile parameter makes DAGDA to do additional and useless transfers e DAGDA can be used without any configuration but it is always a good idea to define all the DAGDA parameters in the configuration files For any comment or bug report on DAGDA please conta
59. UCBL Page 136 DIET User s Manual BOSE E dagda_put_scalar_async void value diet_base_type_t base_type diet_persistence_mode_t mode dagda_put_vector_async void value diet_base_type_t base_type diet_persistence_mode_t mode size_t size dagda_put_matrix_async void value diet_base_type_t base_type diet_persistence_mode_t mode size_t nb_rows size_t nb_cols diet_matrix_order_t order dagda_put_string_async char value diet_persistence_mode_t mode dagda_put_file_async char path diet_persistence_mode_t mode After a call to one of these functions the user can obtain the data ID by calling the dagda_wait_data_ID function with a transfer reference dagda_wait_data_ID unsigned int transferRef char ID The transferRef argument is the value returned by a dagda_put_ _async func tion The JD content will be initialized to a pointer on the data ID DAGDA asynchronous get macros functions The only argument needed for one of these functions is the data ID All of these functions return a reference to the data transfer which is an unsigned int This value will be passed to the corresponding dagda_wait_ functions described later dagda_get_scalar_async char ID dagda_get_vector_async char ID dagda_get_matrix_async char ID dagda_get_string_async char ID dagda_get_file_async char ID After asking for an asynchronous transfer the user has to wait the end by calling
60. XXXXXXXXXXXXKX XXXXXXXXXXXXXXXXXXXXXX DIET configuration summary XXXXXXXXXXXXXXXXXXXXXX XXXXXXXXXXXXXXXXXXXXXXXXXXX 2010 03 31 07 47 15 XXXXXXXXXXXXXXXXXXXXXXXXXX XXX System name Linux XXX Install prefix home diet local diet XXX C compiler usr bin gcc XXX version 4 3 4 XXX options Dinline static __inline__ Dconst std gnu99 XXX CXX compiler usr bin c XXX version 4 3 4 XXX options lpthread g D__linux__ XXX OmniORB found YES XXX OmniORB version 4 1 2 XXX OmniORB directory XXX OmniORB includes usr include XXX OmniORB libraries libomniDynamic4 so libomniORB4 so libomnithread so XXX General options XXX Dynamics Libraries ON XXX Examples ON XXX BLAS Examples ON XXX Options set XXX Batch ON XXX CORI ON XXX JXTA ON XXX JuxMem ON XXX Statistics ON XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX eee A more complete yet technical way of making sure is to check the content of the file named CMakeCache txt generated by cmake in the directory from which cmake was invocated When exchanging with the developers list it is a recommendable practice to join the content of this file which summarizes your options and also the automatic package library detections made by c
61. _desc_aggregator profile for this service use a priority scheduler diet_aggregator_set_type agg DIET_AGG_PRIORITY a install our custom performance function INRIA ENS Lyon UCBL Page 58 DIET User s Manual BOSE E diet_service_use_perfmetric performanceFn b define the precedence order diet_aggregator_priority_minuser agg 0 c diet_aggregator_priority_max agg EST_TIMESINCELASTSOLVE c The performance function performanceFn is defined as follows static void performanceFn diet_profile_t pb estVector_t perfValues performanceFn the performance function to use in the DIET plugin scheduling facility static void performanceFn diet_profile_t pb estVector_t perfValues const char target int numMismatch string value must be fetched from description value is NULL target diet_paramstring_get_desc diet_parameter pb 0 gt param numMismatch computeMismatches target store the mismatch value in the user estimate space using tag value 0 diet_est_set perfValues 0 numMismatch also store the timestamp since last execution diet_estimate_lastexec perfValues pb The function computeMismatches defined earlier in server c calculates the number of re quested databases that are not present on the SeD making the evaluation Together these two code segments serve to customize the generati
62. _wait_and n else printf Result data for requestID for i 0 i lt 5 i printf d rstlil for i 0 i lt 5 i Get and print time diet_scalar_get diet_parameter profile i 2 amp time NULL if time NULL printf Error time not set n else printf time f n time Check the first non zero element of the matrix if fabs matrix i 1 1 2 1 factor gt 1e 15 printf Error matrix not correctly set n INRIA ENS Lyon UCBL Page 38 DIET User s Manual BOSE TE Free profiles for i 0 i lt 5 i diet_cancel rst i diet_profile_free profile i free matrix i free time diet_finalize return 0 4 4 Compilation After compiling the client program the user must link it with the DIET libraries and the CORBA libraries 4 4 1 Compilation using cmake The doc ExternalExample directory also contains a CMakeFile txt file which illustrates the cmake way of compiling this simple client server example PROJECT DIETSIMPLEEXAMPLE SET CMAKE_MODULE_PATH DIETSIMPLEEXAMPLE_SOURCE_DIR Cmake FIND_PACKAGE Diet On success use the information we just recovered INCLUDE_DIRECTORIES DIET_INCLUDE_DIR LINK_DIRECTORIES DIET_LIBRARY_DIR Define a simple server ADD_EXECUTABLE simple_server simple_server c TARGET_LINK_LIBRARIES simple_server DIET_SERVER_LIBRARIES INSTALL T
63. a hierarchy of a unique MA there are several MA and each MA manages a subset of SeDs Those MA are interconnected in a way that they can share the access to their SeDs Each MA works like the usual when they received a query from a user they looks for the best SeD which can resolve their problem inside their hierarchy If there is no SeD available in its hierarchy the queried MA forwards the query to another MA to find a SeD that can be used by its client This way DIET is able to support more clients and more servers because each client request is forwarded to a number of SeDs that is independent of the total number of available SeDs 11 2 Deployment example The instructions about how to compile DIET with the multi MA extension are available in Section 2 2 5 and the configuration instructions are available in Section 9 1 2 INRIA ENS Lyon UCBL Page 101 DIET User s Manual BOSE TE The example described here is about four organizations which want to share there resources The first organization named alpha have ten SeDs which give access to the service a The second organization named beta have eight SeDs with the service a and three with the service b The third one named gamma have two SeDs with the service c The last one named delta have one SeD with the service a but the server crash and the SeD is unavailable Each organization has it s own DIET hierarchy All MAs one for each organization are connected with the m
64. ace DAGDA will not take care of the disk disk partition chosen bythe Y Vv Vv aaa storageDirectory option The maximum memory space used by N E DAGA tort the data If set to O o maximum memory usage is maxMemSpace DAGDA beta ae ss of pa ds set Same effect than to choose V iV iv ory usage 0 The cache replacement algorithm used No cache replacement also when DAGDA needs more space to store P 80 cacheAlgorithm a data Possible values are LRU rithm MAGOS never teplace a Yiv vy LFU FIFO i data by another one The DAGDA component shares its file data with all its children when the shareFiles path is accessible by them for exam No file sharing 0 XIV X ple if the storage directory is on a NFS partition Value can be 0 or 1 The path to the file that will be used when DAGDA save all its stored data ae dataBackupFile data path when asked by the q No checkpointing is possible Xx vv Checkpointing DAGDA will load the dataBackupFile restoreOnStart flea gar and ratore al the data wo fle loading on stant z0 W event Possible values are 0 or 1 Figure 15 2 DAGDA configuration options INRIA ENS Lyon UCBL Page 133 DIET User s Manual BOSE TE 15 3 Cache replacement algorithm When a data is replicated on a site it is possible that not enough disk memory space is available In that case DAGDA allows to choose a strategy to delete a persistent data Only a simple persistent data can be del
65. and wait for messages LogTools are typically used within monitoring tools The main interest in LogService is that information is collected by a central point Log Central that receives logEvents from LogComponents that are attached to DIET elements MA LA and SeD LogCentral offers the possibility to re send this information to several tools LogTools that are responsible for analysing these message and offering compre hensive information to the user LogService defines and implements several functionalities INRIA ENS Lyon UCBL Page 95 DIET User s Manual BOSE TE solve service1 service1 Figure 10 1 DIET and LogService Filtering mechanisms As few messages as possible should be sent to minimize net work traffic With respect to the three tier model the communications between applications e g LogComponent and the collector e g LogCentral as well as between the collector and the monitoring tools e g LogTools should be min imized When a LogTool registers with the LogCentral it also registers a filter defining which messages are required by the tool Message ordering Event ordering is another important feature of a monitoring system LogService handles this problem by the introduction of a global time line At generation each message receives a time stamp The problem that can occur is that the system time can be different on each host LogService measures this difference i
66. arameters 2 desc id matrix D of doubles store_id profile2 gt parameters 1 desc id matrix E of doubles diet_profile_free profile2 diet_free_persistent_data id MA1 1 3 free matrices memory diet_finalize Note that when a single client creates persistent data with a first DIET call and uses that data with a second DIET call we will not know in advance the identifier of the data However the identifier is stored in the structure of the first profile For example consider a matrix A built with diet_matrix_set method as follows diet_profile_t profile INRIA ENS Lyon UCBL Page 33 DIET User s Manual BOSE E diet_matrix_set diet_parameter profile 0 E DIET_PERSISTENT DIET_DOUBLE mA nA oA After the first diet_cal1 the identifier of A is stored in the profile in profile gt parameters 0 desc id So for the second call we will have the following instruction in order to use A diet_profile_t profile2 diet_use_data diet_parameter profile2 0 profile gt parameters 0 desc id NB when using this method the first profile here profile must not be freed before using or making a copy of the data identifier INRIA ENS Lyon UCBL Page 34 DIET User s Manual BOSE TE Chapter 4 Building a client program The most difficult part of building a client program is to understand how to describe the problem interface Once this step is done it is fairly easy to build calls t
67. be incorporated with the keyword batchName in the server configuration file Only this makes the SeD know how to submit a job correctly Furthermore if there is no default queue the DIET deployer must also provide the queue on which jobs have to be submitted with the keyword batchQueue You also have to provide a directory where the SeD can read and write data on the parallel resource Please note that this directory is used by DIET to store the new built script that is submitted to the batch scheduler In consequence because certain batch schedulers like OAR need the script to be available on all resources this directory might be on NFS remember that DIET cannot replicate the script on all resources before submission because of access rights Note that concerning OAR v1 6 in order to use the CoRI batch features for OAR 1 6 see Section 8 3 4 the Batch SeD deployer must also provide the keyword internQueue with the corresponding name For example the server configuration file can contain the following lines batchName oar batchQueue queue_9_13 pathToNFS home ycaniou tmp nfs pathToTmp tmp YC intenOARbatchQueueName 913 6 8 3 Server API for writing services The writing of a service corresponding to a parallel or batch job is very simple The SeD programmer builds a shell script that he would have normally used to execute the job i e a script that must take care of data replication and executable invocation depending on
68. choose which collector will provide the information CoRI Manager FAST CORI Easy CORI Batch Figure 8 2 CoRI overview INRIA ENS Lyon UCBL Page 79 DIET User s Manual BOSE TE 8 3 1 Functions and tags The tags for information are of type integer and defined in the table 7 1 The second type of tag diet_est_collect_tag_t is used to specify which collector will provide the in formation EST_COLL_FAST EST_COLL_EASY or EST_COLL_BATCH Three different functions are provided with CoRI The first function initializes a specific collector int diet_estimate_cori_add_collector diet_est_collect_tag_t collector_type void data The second parameter is reserved for initializing collectors which need additional infor mation on initialization For example the BATCH collector needs for its initialization the profile of the service to be solved After the initialization accessing to the information is done by specifying the collector and the information type int diet_estimate_cori estVector_t ev int info_type diet_est_collect_tag_t collector_type void data Cori Easy doesn t need more information but FAST and BATCH need a profile of type diet_profile_t The last parameter is reserved for it The last function is used to test Cori Easy It prints all information Cori Easy finds to the standard output void diet_estimate_coriEasy_print A result could be the following output
69. ct G Le Mahec at the fol lowing e mail address gael le mahec ens lyon fr 15 6 Future works The next version of DAGDA will allow the users to develop their own cache replacement algorithms and network capacity measurements methods DAGDA will be separated in two parts A data management interface and the DAGDA data manager itself DAGDA will implement the GridRPC data management API extension INRIA ENS Lyon UCBL Page 140 DIET User s Manual BOSE E Chapter 16 Dynamic management 16 1 Dynamically modifying the hierarchy 16 1 1 Motivations So far we saw that DIET s hierarchy was mainly static once the shape of the hierarchy chosen and the hierarchy deployed the only thing you can do is kill part of the hierarchy or add new subtrees to the existing hierarchy But whenever an agent is killed the whole underlying hierarchy is lost This has several drawbacks some SeD will become unavailable and if you want to reuse the machines on which those SeD or agents are you need to kill the existing DIET element and redeploy a new subtree Another problem due to this static asignement of the parent children links is that if you have an agent that is overloaded you cannot move part of its children to an underloaded agent somewhere else in the hierarchy without once again killing part of the hierarchy and deploying once again 16 1 2 And thus it began to evolve Hence DIET also has a mode in which you can dynamicall
70. d DIET_DCOMPLEX are not implemented yet 3 1 2 Composite types Composite types are defined in an enum type diet_type_t INRIA ENS Lyon UCBL Page 25 DIET User s Manual cose E Type Possible base types DIET SCALAR all base types DIET_VECTOR all base types DIET_MATRIX all base types DIET_STRING DIET_CHAR DIET_PARAMSTRING DIET_CHAR DIET_FILE DIET_CHAR DIET_CONTAINER all base types Each of these types requires specific parameters to completely describe the data see Figure gJ 3 1 3 Persistence mode Persistence mode is defined in an enum type diet persistence mode t mode DIET_VOLATILE DIET_PERSISTENT_RETURN DIET_PERSISTENT DIET_STICKY DIET_STICKY_RETURN Description not stored stored on server movable and copy back to client stored on server and movable stored and non movable stored non movable and copy back to client NB DIET_STICKY_RETURN only works with DAGDA 3 2 Data description Each parameter of a client problem is manipulated by DIET using the following structure typedef struct diet_arg_s diet_arg_t struct diet_arg_s diet_data_desc_t desc void value typedef diet_arg_t diet_data_t The second field is a pointer to the memory zone where the parameter data are stored The first one consists of a complete DIET data description which is better described by a figure than with C code since it can be set and accessed through API funct
71. de lt stdio h gt include lt stdlib h gt include DIET_server h The second step is to define a function whose prototype is DIET normalized and which will be able to convert the function into the library function prototype Let us consider a library function with the following prototype int service int argl char arg2 double arg3 This function cannot be called directly by DIET since such a prototype is hard to manip ulate dynamically The user must define a solve function whose prototype only consists of a diet profile t This function will be called by the DIET SeD through a pointer int solve_service diet_profile_t pb int argl char arg2 double arg3 diet_scalar_get diet_parameter pb 0 amp argi NULL diet_string_get diet_parameter pb 1 amp arg2 NULL diet_scalar_get diet_parameter pb 2 amp arg3 NULL return service argi arg2 arg3 INRIA ENS Lyon UCBL Page 41 DIET User s Manual BOSE E Several API functions help the user to write this solve function particularly for getting IN arguments as well as setting OUT arguments Getting IN INOUT and OUT arguments The diet_ _get functions defined in DIET_data h are still usable here Do not forget that the necessary memory space for OUT arguments is allocated by DIET So the user should call the diet_ _get functions to retrieve the pointer to the zone his her program should write to Setting INOUT and OUT argu
72. de the implementation specific profile of a service from the user It allows different servers to declare the same service with the same profile using different implementations of the service As FAST relies on the path of the service the INRIA ENS Lyon UCBL Page 77 DIET User s Manual BOSE TE convertor can also change the path of the declared profile to enable a correct evaluation of the incoming requests by FAST If no convertor is passed when declaring a new service a default convertor is assigned to it that does not change its profile nor its path To translate a profile the convertor defines a new destination profile with a new path It then chooses for each argument of the new profile a predefined function to assign this argument from the source profile This allows the following operations Permutation of arguments This is done implicitly by specifying which argument in the source profile corresponds to which argument in the destination profile Copy of arguments Arguments can be simply used by applying the DIET_CVT_IDENTITY function If the same source argument corresponds to two destination arguments it is automatically copied Creation of new arguments New arguments can either contain static values or the proper ties of existing arguments To create a new static value the index for the source argument must be invalid e g 1 and the arg parameter must be set to the static argument To extract a property of an exis
73. duler Overriden aggregate method to schedule jobs with the SRA policy int aggregate corba_response_t aggrResp size_t max_srv const size_t nb_responses const corba_response_t responses using namespace std const char HostnameRR stName UserGS HostnameRR HostnameRR HostnameRR HostnameRR 4 this gt name this gt stName this gt nameLength strlen this gt name int HostnameRR aggregate corba_response_t aggrResp size_t max_srv const size_t nb_responses const corba_response_t responses ServerList iterator itSeD unsigned int nbUsage 0 corba_server_estimation_t selected cout lt lt NOOO HostnameRR ee lt lt endl ServerList candidates CORBA_to_STL responses nb_responses for itSeD candidates begin itSeD candidates end itSeD We select the SeD by its host usage if hostCounter HOSTNAME itSeD lt nbUsage selected itSeD lt aggrResp gt servers length 1 aggrResp gt servers 0 selected return 0 SCHEDULER_CLASS HostnameRR INRIA ENS Lyon UCBL Page 72 DIET User s Manual BOSE E 7 6 Future Work We have two primary efforts planned for extensions to the plugin scheduler e Additional information services We plan to add functionalities to enable the appli cation developer to access and use data concerning the internal state of the DIET server e g the current length of request queues As other performance
74. e INRIA ENS Lyon UCBL Page 13 DIET User s Manual BOSE TE e Each server that can satisfy the request can send his performance and hardware infor mation or an estimation of the computation time necessary to process the request to its parent an LA via performance prediction tools see Chapter 8 e Each LA that receives one or more positive responses from its children sorts the servers and forwards the best responses to the MA through the hierarchy e Once the MA has collected all the responses from its direct children it chooses a pool of fast servers and sends their references to the client 1 5 DIET Extensions 1 5 1 Multi MA A standard DIET platform gives access to SeDs placed under the control of a MA as explained at the beginning of this chapter Sometime it is useful to connect several MA together This happens when several organizations wish to share their resources to offer a larger set of service types and more available servers The Multi MA extension allows this by creating a federation which shares resources between several MA In multi MA mode the behavior of a DIET hierarchy does not change when a client requests a service that is available under the queried MA However if a request sent to a MA does not found a SeD that can resolve its problem DIET will forward the request to other MAs of the federation To read more about multi MA see Chapter 11 and Chapter 12 1 5 2 FAST Fast Agent s System T
75. e http graal ens lyon fr DIET logservice html for more information 10 2 VizDIET VizDIET is the monitoring tool written for DIET to be able to vizualize and analyze the status and activities of a running DIET deployment As described in Section 10 1 all DIET s components integrate a LogComponent and VizDIET implements the Log Tool interface in order to be able to collect all information sent by DIET s components through their LogComponent VizDIET provides a graphic representation of the DIET architecture being monitored There are two ways to use VizDIET Real time monitoring VizDIET is directly connected to the LogCentral using a Corba connection and receives directly all information about the running DIET platform INRIA ENS Lyon UCBL Page 97 DIET User s Manual BOSE E Post mortem monitoring VizDIET reads a log file containing all log messages re ceived by LogCentral This post mortem analysis can also be replayed in real time if the log file is time sorted The log file is created during the real deployment by a special tool provided with LogService that receives all messages from LogCentral and writes them to a file SO XVizDIET 1 1 1 ee File View Actions Options A aj pa Reading SE DIET Platform Numberof MA 1 6 42 Total Request 0 del dad Twat e fpe E ne List of services Figure 10 2 Snapshot of VizDIET As describ
76. e matrix with dummy values who cares 7 for i 0 i lt 60 100 i matrix i 1 2 i Initialize a DIET session diet_initialize client cfg argc argv Create the profile as explained in Chapter 3 profile diet_profile_alloc smprod 0 1 2 last_in last_inout last_out Set profile arguments INRIA ENS Lyon UCBL Page 36 DIET User s Manual BOSE E diet_scalar_set diet_parameter profile 0 amp factor 0 DIET_DOUBLE diet_matrix_set diet_parameter profile 1 matrix 0 DIET_DOUBLE 60 100 DIET_COL_MAJOR diet_scalar_set diet_parameter profile 2 NULL 0 DIET_FLOAT if diet_call profile If the call has succeeded Get and print time diet_scalar_get diet_parameter profile 2 amp time NULL if time NULL printf Error time not set n else printf time f n time Check the first non zero element of the matrix if fabs matrix 1 1 2 1 factor gt 1e 15 4 printf Error matrix not correctly set n Free profile diet_profile_free profile diet_finalize free matrix free time 4 3 2 Asynchronous call smprod for scalar by matrix product include lt stdio h gt include lt stdlib h gt include lt math h gt include DIET_client h int main int argc char argv int i j double factor M_PI Pi why not size_t m n Matrix size do
77. e source profile As it is an IN argument the last parameter is not important diet_arg_cvt_set amp cvt gt arg_convs 0 DIET_CVT_MAT_NB_ROW 0 NULL 0 apply the function DIET_CVT_MAT_NB_COL to determine the 1st argument of the converted profile The function s argument is the Oth argument of the source profile As it is a IN argument the last parameter is not important diet_arg_cvt_set amp cvt gt arg_convs 1 DIET_CVT_MAT_NB_COL 0 NULL 0 apply the function DIET_CVT_IDENTITY to determine the 2nd argument of the converted profile The function s argument is the Oth argument of the source profile and it will be written back to the Oth argument of the source profile when the call has finished diet_arg cvt_set amp cvt gt arg_convs 2 DIET_CVT_IDENTITY 0 NULL 0 NOTE The last line could also be written as diet_arg_cvt_short_set amp cvt gt arg_convs 2 0 NULL add the service using our convertor diet_service_table_add profile cvt solve_T free our convertor diet_convertor_free cvt More examples on how to create and use convertors are given in the files examples dmat_manips server c and examples BLAS server c 8 3 CoRI Collectors of Ressource Information CoRI manages the access to different tools for collecting information about the SeD At present three tools called collectors are implemented FAST CoRI Easy and CoRI batch The user can
78. eD COMMTIME SeD TCOMP SeD TIMESINCELASTSOLVE SeD COMMPROXIMITY SeD TRANSFEREFFORT SeD FREECPU SeD FREEMEM SeD NBCPU SeD CPUSPEED SeD idx TOTALMEM SeD INRIA ENS Lyon UCBL Page 68 DIET User s Manual BOSE E AVGFREEMEM SeD AVGFREECPU SeD BOGOMIPS SeD idx CACHECPU SeD idx TOTALSIZEDISK SeD FREESIZEDISK SeD DISKACCESSREAD SeD DISKACCESSWRITE SeD USERDEFINED SeD idx The macros taking two arguments need an index to choose which CPU measurement is needed Two extra macros are defined e HOSTNAME server The hostname of the SeD e SED_REF server A CORBA reference to the SeD Here is an example of an aggregate function using these macros SORTFUN compBogo BOGOMIPS int MyScheduler aggregate corba_response_t aggrResp size_t max_srv const size_t nb_responses const corba_response_t responses ServerList candidates CORBA_to_STL responses nb_responses ServerList chosen ServerList iterator it for it candidates begin it candidates end it if NBCPU it gt 2 chosen push_back it SORT chosen compBogo STL_to_CORBA chosen aggrResp return 0 This aggregation method first selects only the SeD which have more than 1 CPU and sorts them according to their number of Bogomips INRIA ENS Lyon UCBL Page 69 DIET User s Manual BOSE E 7 5 5 Creation and usage of a scheduler module How to compile a
79. ebugging deployment problems when using GODIET can be difficult especially if you don t fully understand the role of each element you are launching If you have trouble identifying the problem read the rest of this chapter in full and try launching key elements of your deployment by hand GoDIET is available for download on the web An example input XML file is shown in Figure 9 1 see 4 for a full explanation of all entries in the XML You can also have a look at the fully commented XML example file provided in the GODIET distribution under examples commented xml each option is explained To launch GODIET for the simple example XML file provided in the GODIET distribution under examples examplel1 xml run gt java jar GoDIET x x x jar examplei xml XmlScanner constructor Parsing xml file examplel xml GoDIET gt GODIET reads the XML file and then enters an interactive console mode In this mode you have a number of options GoDIET gt help The following commands are available launch launch entire DIET platform launch_check launch entire DIET platform then check its status relaunch kill the current platform and launch entire DIET platform once again stop kill entire DIET platform using kill pid status print run status of each DIET component 2http graal ens lyon fr DIET godiet html INRIA ENS Lyon UCBL Page 91 DIET User s Manual BOSE E history print history of commands executed help pri
80. ed before every release DIET is known to be effective on a wide range of platforms Nevertheless if you encounter installation difficulties don t hesitate to post on DIET s users mailing list diet usr listes ens lyon fr for the archives refer to http graal ens lyon fr DIET mail lists html If you find a bug in DIET please don t hesitate to submit a bug report on http graal ens lyon fr bugzilla If you have multiple bugs to report please make multiple submissions rather than submitting multiple bugs in a single report 2 1 2 Hardware dependencies DIET is fully tested on Linux i386 and Linux i686 platforms DIET is known to be effective on Linux Sparc Linux i64 Linux amd64 Linux Alpha Linux PowerPC AIX PowerPC Ma cOS PowerPC and Windows XP Cygwin i386 platforms At some point in DIET history DIET used to be tested on the Solaris Sparc platform 2 1 3 Supported compilers DIET is supported on GCC with versions ranging from 3 2 X to 4 3 4 Note that due to omniORB 4 see 2 1 5 requirements towards thread safe management of exception handling compiling DIET with GCC requires at least the version 2 96 DIET is also supported on XL compiler IBM and Intel compiler INRIA ENS Lyon UCBL Page 15 DIET User s Manual BOSE E 2 1 4 Operating system dependencies DIET is fully tested on Linux with varying distributions like Debian Red Hat Enterprise Linux REL ES 3 Fedora Core 5 on AIX 5 3 on MacOSX
81. ed by the Master Agent to share its IOR useConcJobLimit e Component SeD e Mode All e Type Boolean e Description should SeD restrict the number of concurrent solves This should be used in conjunction with maxConcJobs maxConcJobs INRIA ENS Lyon UCBL Page 151 DIET User s Manual BOSE TE e Component SeD e Mode All e Type Integer e Description If useConcJobLimit true how many jobs can run at once This shoudl be used in conjunction with maxConcJobs locationID e Component SeD e Mode DAGDA Type String e Description This parameter is used for alternative transfer cost prediction MADAGNAME e Component Client e Mode Workflow Type String e Description the name of the MAp zg agent to wich the client will connect schedulerModule e Component Agent e Mode User scheduling e Type String e Description The path to the scheduler library file containing the implemen tation of the plugin scheduler class moduleConfigFile e Component Agent e Mode User scheduling Type String e Description Optional configuration file for the module batchName e Component SeD e Mode Batch Type String Description The reservation batch system s name INRIA ENS Lyon UCBL Page 152 DIET User s Manual BOSE E batchQueue e Component SeD e Mode Batch e Type String e Description The name of the queue where the job will be submitted pathToNFS e Component
82. ed in Section 1 4 there are two main steps in the treatment of a request in DIET one step to find and schedule a service and one step to solve this service So two main activities are represented schedule and compute information Schedule information When an agent takes a scheduling decision for a task 1 e finding and deciding which SeD can execute a service it is useful to know how the agent made its decision This information is represented by FindRequest in VizDIET Compute information When a SeD is computing a job we need to be aware of its state and know when the computation begins and ends This information is represented by SolveRequest In VizDIET when a SeD is solving a service the SeD changes color to red FindRequests are only attached to agents and SolveRequests are only attached to SeDs Finally the aggregation of one FindRequest and its SolveRequest is concatenated in one request DIETRequest DIETResquest can be see as a job execution in a DIET INRIA ENS Lyon UCBL Page 98 DIET User s Manual BOSE E DOA XDIETStats _ 998 DIET Platform DIET Platform Property Nb DIETRequest 22 Cowes gramin 3 Legend 540 Figure 10 3 Bar taskflow and gantt graphs in vizDIET platform as seen by an end user A DJETRequest is als
83. eduling process is managed in DIET Most of the developers can go directly to the next section All the schedulers developed by users have to inherit from the UserScheduler class This class furnishes the methods to load its subclasses as a Scheduler class for DIET without error The only method a user has to overload is the aggregate method Several useful functions and macros are defined in the UserScheduler hh file The UserScheduler class is defined as follows INRIA ENS Lyon UCBL Page 61 DIET User s Manual BOSE E class UserScheduler public GlobalScheduler typedef GlobalScheduler constructor typedef void destructor UserScheduler public static const char stName UserScheduler virtual UserScheduler These methods are used to load the user module and to obtain an instance of the scheduler static UserScheduler getInstance const char moduleName static GlobalScheduler instanciate const char moduleName void destroy GlobalScheduler scheduler static GlobalScheduler deserialize const char serializedScheduler const char moduleName static char serialize GlobalScheduler GS The method that has to be overloaded to define a new scheduler virtual int aggregate corba_response_t aggrResp size_t max_srv const size_t nb_responses const corba_response_t responses private E The UserScheduler class is a singleton class Its constructor is
84. efines the path to Xerces installation directory for example DXERCES_DIR PATH usr local xerces XQILLA DIR defines the path to XQilla installation directory for example DXQILLA DIR PATH usr local xqilla N B 1 By activating the workflow module the DAGDA module is also activated This is an example of generating command line cmake DMAINTAINER_MODE BOOL 0N DOMNIORB4_DIR usr local omniORB DDIET_USE_WORKFLOW BOOL ON DXERCES_DIR usr local xerces DXQILLA_DIR usr local xqilla Workflow support was tested in the following configurations gcc version 4 0 2 and higher omniORB version 4 1 0 and higher Xerces 3 0 1 XQilla 2 2 0 N B 2 Workflow support is not available on Windows Cygwin platforms Windows XP and Cygwin lt 1 5 25 for Xerces 3 0 1 and XQilla 2 2 0 INRIA ENS Lyon UCBL Page 116 DIET User s Manual BOSE E succ y y O 2y 2y sum je Figure 14 1 DAG example Executing the examples The directory examples workflow includes some exam ples of workflows You can find a simple DAG workflow see Figure 14 1 in the file xml scalar xml and you can test it with the following command where local_client cfg is the DIET configuration file example provided in the etc client wf cfg file generic_client local_client cfg dag scalar xml You need to have a running DIET platform with the MApac agent and the needed services You can launch a single SeD scalar_s
85. ent Agent and SeD e Mode FAST e Type Boolean INRIA ENS Lyon UCBL Page 148 DIET User s Manual BOSE TE e Description If set to 0 all LDAP and NWS parameters are ignored and all requests to FAST are disabled when DIET is compiled with FAST This is useful for testing a DIET platform without deploying an LDAP base nor an NWS platform IdapUse e Component Agent and SeD e Mode FAST e Type Boolean e Description 0 tells FAST not to look for the services in an LDAP base IdapBase e Component Agent and SeD e Mode FAST e Type Address e Description host port of the LDAP base that stores FAST known services IdapMask e Component Agent and SeD e Mode FAST e Type String e Description the mask which is registered in the LDAP base nwsUse e Component Agent and SeD e Mode FAST e Type Boolean e Description 0 tells FAST not to use NWS for its comm times forecasts nwsNameserver e Component Agent and SeD e Mode FAST e Type Address e Description host port of the NWS nameserver nwsForecaster e Component Agent and SeD INRIA ENS Lyon UCBL Page 149 DIET User s Manual BOSE TE e Mode FAST e Type Address e Description NWS forecast module used by FAST useAsyncAPI e Component Agent and SeD e Mode All e Type Boolean e Description No longer used useLogService e Component Agent and SeD e Mode All e Type Boolean e Description 1 to u
86. er Agents The scheduler can use resource avail ability information collected from three different tools from NWS 26 sensors which are placed on every node of the hierarchy from the application centric performance prediction tool FAST 21 which relies on NWS information or from CoRI Easy which is based on simple system calls and some basic performance tests see Chapter 8 Figure 1 1 shows the hierarchical organization of DIET MA Master Agent LA Local Agent SeD Server Daemon Service Figure 1 1 A hierarchy of DIET agents INRIA ENS Lyon UCBL Page 11 DIET User s Manual BOSE TE 1 1 DIET components The different components of our software architecture are the following Client A client is an application which uses DIET to solve problems Many types of clients are able to connect to DIET from a web page a PSE such as Matlab or Scilab or from a compiled program Master Agent MA An MA receives computation requests from clients These requests refer to some DIET problems listed on a reference web page Then the MA collects computation abilities from the servers and chooses the best one The reference of the chosen server is returned to the client A client can be connected to an MA by a specific name server or a web page which stores the various MA locations Local Agent LA An LA transmits requests and information between MAs and servers The information stored on an LA is the list of
87. er clusters of workstations or parallel machines Computational grids are even considered as hierachical sets of parallel resources as we can see in ongoing project like the french research grid project Grid 5000 2 for the moment 9 sites are involved or like the EGEE project Enabling Grids for E science in Europe composed of more than a hundred centers in 48 countries Then in order to provide transparent access to resources grid middleware must supply efficient mechanisms to provide parallel services Because parallel resources are managed differently on each site it is neither the purpose of DIET to deal with the deployment of parallel tasks inside the site nor manage copies of data which can possibly be on NFS DIET implements mechanisms for a SeD programmer to easily provide a service that can be portable on different sites for clients to request services which can be explicitly sequential parallel or solved in the real transparent and efficient metacomputing way only the name of the service is given and DIET chooses the best resource where to solve the problem 6 2 Terminology Servers provide services e g instanciation of problems that a server can solve for example two services can provide the resolution of the same problem one being sequential and the other parallel A DIET task also called a job is created by the request of a client it refers to the resolution of a service on a given server A service can be sequen
88. er not worry about casting int or double etc into void XA XX XX X Type int diet_scalar_get diet_arg_t void diet_persistence_mode_t define diet_scalar_get arg value mode _scalar_get arg void value mode Type int diet_vector_get diet_arg_t void diet_persistence_mode_t size_t define diet_vector_get arg value mode size _vector_get arg void value mode size Type int diet_matrix_get diet_arg_t void diet_persistence_mode_t size_t size_t diet_matrix_order_t define diet_matrix_get arg value mode nb_rows nb_cols order _matrix_get arg void value mode nb_rows nb_cols order Type int diet_string_get diet_arg_t char diet_persistence_mode_t define diet_string_get arg value mode _string_get arg char value mode Type int diet_file_get diet_arg t diet_persistence_mode_t size_t char define diet_file_get arg mode size path _file_get arg mode size char path 3 5 Data Management functions e The store_id method is used to store the identifier of persistent data It also accepts a description of the data stored This method has to be called after the diet_ca11 so that the identifier exists store_id char argID char msg e The diet_use_data method allows the client
89. erationstrategy gt lt diet path dock estimation constant gt lt processor gt lt processor name statisticaltest gt lt in name values type double depth 1 gt lt out name result type file gt lt iterationstrategy gt INRIA ENS Lyon UCBL Page 121 DIET User s Manual BOSE E lt cross gt lt port name coefficient gt lt match gt lt port name values gt lt port name weights gt lt match gt lt cross gt lt iterationstrategy gt lt diet path weightedaverage gt lt processor gt lt processors gt lt links gt lt LINKS see below gt lt links gt lt workflow gt Data links A data link is a connection between a processor output port and a processor input port as exampled below lt links gt lt link from key to genParam paramKey gt lt link from genParam paramFiles to docking input gt lt link from parameter to docking param gt lt link from docking result to statisticaltest values gt lt link from statisticaltest result to results gt lt links gt When a processor A port A out is connected to a processor B port B in through a data link an instance of A one task may trigger a number of B instances that depends on first the data depth at both ends of the link and second the iteration strategy chosen for the B in port within the B processor The data depths on both ends of the link determine the number of data
90. erver that includes all the needed services read Chapter 5 for more details You can also find some examples of functional workflows written in the Gwendia language see file xm1 func_string xml and you can test it with the following command generic_client local_client cfg wf func_string xml data xml You need to have a running DIET platform with the needed services the commands to launch the services are included as comments within the workflow XML 14 3 Software architecture A new agent called the MApag is used to manage workflows in the DIET architecture This agent receives requests from clients containing the description of a workflow in a specific language the MApac XML workflow language for DAGs The role of the MApac is to determine how to schedule the tasks contained in the workflow in order to follow the precedence constraints between tasks and how to map the tasks to appropriate ressources in the DIET hierarchy The execution of the individual tasks is actually delegated by the MApag to the client that submitted the workflow After submitting the workflow the client is put in a waiting mode and it will receive individual requests from the MApac to execute each task of the workflow Therefore all the data transfers are done only from the client to INRIA ENS Lyon UCBL Page 117 DIET User s Manual BOSE E the SeDs and do not transit through the MApac When all tasks are completed the MApac will send a release si
91. esources In the ASP approach a common way for clients to ask for resources to solve their problem is to submit a request to the middleware The middleware will find the most appropriate server that will solve the problem on behalf of the client using a specific software Several environments usually called Network Enabled Servers NES have developed such a paradigm NetSolve 1 Ninf 16 NEOS 6 OmniRPC 23 and more recently DIET developed in the GRAAL project A common feature of these environments is that they are built on top of five components clients servers databases monitors and schedulers Clients solve computational requests on servers found by the NES The NES schedules the requests on the different servers using performance information obtained by monitors and stored in a database DIET stands for Distributed Interactive Engineering Toolbox It is a toolbox for easily developing Application Service Provider systems on Grid platforms based on the Client Agen t Server scheme Agents are the schedulers of this toolbox In DIET user requests are served INRIA ENS Lyon UCBL Page 9 DIET User s Manual BOSE TE via RPC DIET follows the GridRPC API defined within the Open Grid Forum 10 INRIA ENS Lyon UCBL Page 10 DIET User s Manual BASE E Chapter 1 A DIET platform DIET is built upon Server Daemons The process of scheduling the requests is distributed amongst a hierarchy of Local Agents and Mast
92. et diet_param_desc profile 0 agg diet_profile_desc_aggregator profile diet_aggregator_set_type agg DIET_AGG_USER diet_service_table_add profile Usually the developer should define a performance metric function to communicate with the agent scheduler For example if the scheduler uses the number of waiting jobs in the FIFO queue the performance metric could be INRIA ENS Lyon UCBL Page 70 DIET User s Manual BOSE E void metric diet_profile_t profile estVector_t values diet_estimate_waiting_jobs values This metric just fixes the number of waiting jobs in the FIFO queue of the SeD Now at the agent side the scheduler can use this value to aggregate sort and filter the SeDs responses More details are given in the following section about how to use the SeDs plugin schedulers to communicate with the agent scheduler module 7 5 6 SeD plugin schedulers and agent schedulers interactions Most of the time a scheduler needs some information from the nodes to choose where a job should be executed By using the plugin scheduler capacities of the SeDs DIET allows to communicate some useful information for the scheduling The developer just has to define a performance metric function and select DIET_AGG_USER as aggregator Information obtained from the SeD Your plugin scheduler can access to the information obtained from CoRI by initializing the estimation vector using the diet_estimate_c
93. et that the first argument of the method call diet_SeD must be the path of the configuration file above Launching a client Our client must connect to the MA_example file client cfg configuration file for a client MAName MA example traceLevel 1 default INRIA ENS Lyon UCBL Page 90 DIET User s Manual BOSE E Run the executable that you linked with the DIET client library and do not forget that the first argument of the method call diet_initialize must be the path of the configuration file above 9 2 GODIET GODIET is a Java based tool for automatic DIET deployment that manages configuration file creation staging of files launch of elements monitoring and reporting on launch success and process cleanup when the DIET deployment is no longer needed 1 The user of GODIET describes the desired deployment in an XML file including all needed external services e g omniNames and LogService the desired hierarchical organization of agents and servers is expressed directly using the hierarchical organization of XML The user also defines all machines available for the deployment disk scratch space available at each site for storage of configuration files and which machines share the same disk to avoid unecessary copies GODIET is extremely useful for large deployments e g more than 5 elements and for experiments where one needs to deploy and shut down multiple deployments to test different configurations Note that d
94. eted the sticky ones are never deleted by the chosen algorithm DAGDA offers three algorithms to manage the cache replacement e LRU The least recently used persistent data of sufficient size is deleted e LFU The least frequently used persistent data of sufficient size is deleted e FIFO Among the persistent data of sufficient size the oldest is deleted 15 4 The DAGDA API By compiling DIET with the DAGDA extension activated the DIET_Dagda h file is in stalled on the DIET include directory This file contains some data management functions and macros 15 4 1 Note on the memory management On the SeD side DAGDA and the SeD share the same data pointers that means that if the pointer is a local variable reference when DAGDA will use the data it will read an unallocated variable The users should allways allocate the data with a malloc calloc or new call on the SeD and agent sides Because DAGDA takes the control of the data pointer there is no risk of memory leak even if the service allocate a new pointer at each call The data lifetime is managed by DAGDA and the data will be freed according to its persistence mode On the SeD and agent sides DAGDA takes the control of the data amp pointers To free a data may cause major bugs which could be very hard to find The users could only free a DIET data on the client side after the end of a transfer 15 4 2 Synchronous data transfers All of the following function
95. eye ia a Lans ph emule LEE A4 ee Da Le 140 16 Dynamic management 141 16 1 Dynamically modifying the hierarchy s 141 16 0 Wietivasions o cs LU Ne Louve di she ne oe ee due 141 16 1 2 And thus 16 began te evolve cisco du abs et Hate 141 16 13 Example s 3 442 488 du e be Ee ERE eee eee uit ne 142 16 2 Changing oflered Services cos mos d a mime pass Re a e oe ee ees 142 INRIA ENS Lyon UCBL Page 6 DIET User s Manual BOSE TE 102 L PESCA on ee ok a De A Ee Se au 142 1622 Merges ros eo ee de oe Pe ee et 143 16 2 8 Gola MENEE se ech Bk ee ee ee Pe ee Ee Be ee 145 A Appendix 147 A1 Configuration fils 2 444 s eo s deer a us be eo mme bat dead e ves 147 INRIA ENS Lyon UCBL Page 7 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 8 DIET User s Manual BOSE E Introduction Resource management is one of the key issues for the development of efficient Grid environments Several approaches co exist in today s middleware platforms The granularity of computation or communication and dependencies between computations can have a great influence on the software choices The first approach provides the user with a uniform view of resources This is the case of GLOBUS 8 which provides transparent MPI communications with MPICH G2 between distant nodes but does not manage load balancing issues between these nodes It s the user s task to develop a code
96. for the shortest path between them in the graph of monitored links It estimates the bandwidth as the minimum of those in the path and the latency as the sum of those measured This allows the availability of more predictions when DIET is deployed over a hierarchical network INRIA ENS Lyon UCBL Page 75 DIET User s Manual BOSE E DDIET_USE_CORI BOOL OFF BOOL ON Information tag DDIET_USE_FAST starts with EST_ BOOL OFF BOOL ON BOOL OFF BOOL ON TCOMP x FREECPU x x x FREEMEM x x x NBCPU De x x CPUSPEED x De TOTALMEM x x AVGFREECPU x x BOGOMIPS x De CACHECPU x x TOTALSIZEDISK x x FREESIZEDISK x x DISKACCESREAD x De DISKACCESWRITE x x ALLINFOS x x DDIET USE BATCH 0N PARAL_NB_FREE_RESOURCES_IN_DEFAULT_QUEUE x x Table 8 1 Dependencies of the available information on the compiling options In addition to system availabilities FAST can also forecast the time and space needs of certain computational routines as a function of the problem parameters and the machines where the computations would take place FAST is particularly suited to numerical algebra routines whose performance is not data dependent and where a clear relationship exists between problem size and performance As a basis for predictions FAST benchmarks the routines at installation time on each machine for a representative set of parameters After polynomial data fitting the resul
97. g command xmllint your_xml_file dtdvalid path_to_GoDIET dtd noout you will see the different lines where there is problem and a clear description of why your XML file is not compliant INRIA ENS Lyon UCBL Page 93 DIET User s Manual BOSE E lt xml version 1 0 standalone no gt lt DOCTYPE diet configuration SYSTEM GoDIET dtd gt lt diet_configuration gt lt goDiet debug 2 saveStdOut yes saveStdErr yes useUniqueDirs no log no gt lt resources gt lt scratch dir tmp GoDIET_scratch gt lt storage label disk 1 gt lt scratch dir tmp run scratch gt lt scp server res1 login doe gt lt storage gt lt storage label disk 2 gt lt scratch dir tmp run_scratch gt lt scp server res2 login foo gt lt storage gt lt storage label disk 3 gt lt scratch dir tmp run_scratch gt lt scp server res3 login bar gt lt storage gt lt compute label res1 disk disk 1 gt lt ssh server res1 login doe gt lt env gt lt var name PATH value gt lt var name LD LIBRARY PATH value gt lt env gt lt compute gt lt compute label lt ssh server res2 disk disk 2 gt res2 login foo gt lt env gt lt var name PATH value gt lt var name L
98. gnal to the client which will then retrieve the results if the execution was successful To use the MA pag the client configuration file must include the parameter MADAGNAME with the appropriate name When the client uses a functional workflow in Gwendia language the DIET client library provides the logic for instanciating the workflow generating the DAGs and sending them to the MApac agent Note that when several DAGs are generated they are usually not independent as some data generated by one DAG may be used by another one 14 4 Workflow description languages 14 4 1 MaDag language A DAG is described with an XML representation which is close to DIET profile represen tation In addition to profile description problem path and arguments this description represents also the data dependencies between ports source sink the node identifier unique and the precedences between nodes This last information can be removed since it can be retrieved from the dependencies between ports however it can be useful to define a temporal dependency without port linking The general structure of this description is lt dag gt lt node id path gt lt arg name type value gt lt in name type source gt lt out name type sink gt lt out name type sink gt lt node gt The name argument represents the identifier of the
99. h to the DTD is provided using a IDOCTYPE XML entity in the workflow XML file The Gwendia DTD is included in the DIET distribution in the etc FWorkflow dtd file Types Values flowing through the workflow are typed Basic types are integer short double longint float string and file Homogeneous arrays of values can be also used as inputs outputs and can have any depth an array can contain arrays of values depth 2 Arrays are ordered and can eventually contain NULL elements Processors A processor is a data production unit A regular processor invokes a service through a known interface Defined processor types are webservice diet and beanshell Special processors are workflow source a processor with no inbound con nectivity delivering a list of externally defined data values sink a processor with no outbound connectivity receiving some workflow output and constant a processor de livering a single constant value To improve readability the input sink and constant processors are grouped in an lt interface gt tag within the document Other example of processors are grouped in a lt processors gt tag Web services define a lt wsdl gt tag pointing to their WSDL description and the operation to invoke Beanshells define a lt script gt tag containing the java code to interpret DIET services define a lt diet gt tag describing the path to service to invoke When executing the workflow using the DIET INRIA ENS Lyon UCBL
100. he chosen tag and value indicates the value to be associated with the tag Tagged data are used to effect scheduling policies by defining custom server response aggregation methods described in Section 7 3 4 7 3 3 Estimation Function The default behavior of a SeD when a service request arrives from its parent agent is to store the following information in the request profile 1 FAST based execution time predictions DIETSeDs attempt to call FAST routines to obtain execution time predictions based on the type of service requested if FAST was available at compilation time If available such predictions are stored in the performance estimate 2 NWS based dynamic resource information If NWS library functions are available performance estimates may include dynamic resource performance information about CPU availability free memory and network bandwidth 3 Elapsed time since last execution To implement the default round robin behavior in absence of FAST and NWS facilities each SeD stores a timestamp of its last execution When a service request arrives the difference between that timestamp and the current time is added to the performance estimate This is accomplished by using the diet_estimate_fast and diet_estimate_lastexec functions described in Section 7 3 1 To implement a plugin scheduler we define an interface that admits customizable perfor mance estimation routines typedef void diet perfmetric t diet profile tx e
101. he hierarchy Usage disconnect lt SED LA gt lt element name gt e remove cc sends orders to remove an element Usage remove lt SED AGENT gt lt element name gt recursive 011 16 2 Changing offered services 16 2 1 Presentation A SeD does not necessarily need to declare all its services initially i e as presented in Chapter 5 before launching the SeD via diet_SeD One could want to initially INRIA ENS Lyon UCBL Page 142 DIET User s Manual BOSE TE declare a given set of services and then depending on parameters or external events one could want to modify this set of services An example of such usage is to spawn a service that is in charge of cleaning temporary files when they won t be needed nor by this SeD nor by any other SeD or clients and when this service is called it cleans whatever needs to be cleaned and then this service is removed from the service table Adding a service has already been introduced in Chapter 5 using diet_service_table_add you can easily add a new service be it before running the SeD or within a service Well removing a service is as easy you only need to call one of these methods int diet_service_table_remove const diet_profile_t const profile int diet_service_table_remove_desc const diet_profile_desc_t const profile So basically when you want to remove the service that is called you only need to pass the diet_profile_t you receive in the solve fu
102. imer FAST 21 is a tool for dynamic performance forecasting in a Grid environment When DIET is compiled with the appropriate options and FAST has been config ured on the SeD machine SeDs can access FAST to obtain dynamic performance predictions See Chapter 8 for details on using FAST 1 5 3 CoRI Collector of Resource Information CoRI is a manager for collecting hardware and performance information When DIET is compiled with the appropriate option it is possible to get this information via different sub modules like FAST or CoRI Easy if compiled and configured on the SeD machine See Chapter 8 for details on using CoRI INRIA ENS Lyon UCBL Page 14 DIET User s Manual BOSE E Chapter 2 DIET installation 2 1 Dependencies 2 1 1 General remarks on DIET platform dependencies DIET is itself written in C C and for limited parts in java DIET is based on CORBA and thus depends on the chosen CORBA implementation Additionally some of DIET extensions make a strong use of libraries themselves written in C C and java Thus we could expect DIET to be effective on any platform offering decent version of such compilers DIET undergoes daily regression tests see http cdash inria fr CDash index php project DIET on various hardwares a couple of Un x based operating systems under dif ferent distributions MacOSX and AIX and mainly with GCC But thanks to users reports punctual deployments and special tests conduc
103. int_matrix B mB nB oB DIET_ROW_MAJOR diet_matrix_set diet_parameter profile 2 NULL DIET_PERSISTENT_RETURN DIET_DOUBLE mA nB oC INRIA ENS Lyon UCBL Page 32 DIET User s Manual BOSE E if diet_call profile diet_matrix_get diet_parameter profile 2 amp C NULL amp mA amp nB amp oC store_id profile gt parameters 2 desc id matrix C of doubles store_id profile gt parameters 1 desc id matrix B of doubles store_id profile gt parameters 0 desc id matrix A of doubles print_matrix C mA nB oC DIET_ROW_MAJOR diet_profile_free profile free matrices memory diet_finalize Then a client submits the problem D E C with C already present in the platform We consider that the handle of C is id MA1 1 3 double C D E matrices initialization diet_initialize strcpy path MatSUM profile2 diet_profile_alloc path 1 1 2 printf second pb n n diet_use_data diet_parameter profile2 0 id MA1 1 3 diet_matrix_set diet_parameter profile2 1 E DIET_PERSISTENT DIET_DOUBLE mA nB oE print_matrix E mA nB oE DIET_ROW_MAJOR diet_matrix_set diet_parameter profile2 2 NULL DIET_PERSISTENT_RETURN DIET_DOUBLE mA nB oD if diet_call profile2 diet_matrix_get diet_parameter profile2 2 amp D NULL amp mA amp nB amp oD print _matrix D mA nB oD DIET_ROW_MAJOR store_id profile2 gt p
104. ions Figure 3 1 shows the data classification used in DIET Every class inherits from the root class data and could also be a parent of more detailed classes of data in future versions of DIET 3 3 Data management 3 3 1 Data identifier The data identifier is generated by the MA The data identifier is a string field that contains the MA name the number of the session plus the number of the data in the problem incremental plus the string id This is the id field of the diet_data_desc_t structure INRIA ENS Lyon UCBL Page 26 DIET User s Manual BOSE TE e base type I I e type I I e persistence i 1 scalar vector string paramstring e pointer to e size ra e length e pointer to the ID the value e number of columns without the NO pointer to e pointer to e order the array row or col major the first character e pointer to the array Figure 3 1 Argument Data structure description typedef struct char id diet_persistence_mode_t mode diet_data_desc_t For example id MA1 1 1 will identify the first data in the first session submitted on the Master Agent MA1 NB the field id of the identifier will be next replaced by a client identifier This is not implemented yet 3 3 2 Data file The name of the file is generated by a Master Agent It is created during the diet_initialize call The name of the file is the aggregation of the string ID_FILE plus
105. items received by the B in port Three cases are possible e 1 to 1 when depth A out depth B in a data item produced by A out is sent as is to B in e 1 to N when depth A out lt depth B in a data item produced by A out is an array that will be split into its elements when sent to B This will produce several parallel instances tasks of the B processor This is equivalent to a foreach structure in usual programming languages but is here transparent for the user as this is the workflow engines that manages it e N to 1 when depth A out gt depth B in several data items produced by A out by different tasks will be grouped in an array before being sent to B in This is the opposite behaviour from the previous point Note that this structure creates INRIA ENS Lyon UCBL Page 122 DIET User s Manual BOSE E a synchronization barrier among the A tasks as they must all be completed before the B tasks can be launched Conditionals if then else Specific out ports tags lt outThen gt and lt outElse gt are used in that kind of node An outThen port will receive data assigned according to the assignment list in the lt then gt tag only when the condition is evaluated to true If the condition is false this port will not receive data but the lt outElse gt port will receive data according to the assignment list in the lt else gt tag assignment lists are semi column separated lists of assignments of an outThen or o
106. k DIET launch done at Wed Jul 13 09 57 14 CEST 2005 time 11 0 sec The status command will print out the run time status of all launched components The LaunchState reports whether GODIET observed any errors during the launch itself When the user requests the launch of LogService in the input XML file GODIET can INRIA ENS Lyon UCBL Page 92 DIET User s Manual BOSE E connect to the LogService after launching it to obtain the state of launched components when available this state is reported in the LogState column GoDIET gt status Status Element LaunchState LogState Resource PID OmniNames running none localHost 1232 MA_O running none localHost 1262 LA_O running none localHost 1296 SeD_0 running none localHost 1329 Finally when you are done with your DIET deployment you should always run stop To clean up each element GODIET runs a kill operation on the appropriate host using the stored PID of that element GoDIET gt stop Stopping DIET platform at Wed Jul 13 10 05 42 CEST 2005 Trying to stop element SeD_0 Trying to stop element LA_0 Trying to stop element MA_0 Trying to stop element OmniNames DIET platform stopped at Wed Jul 13 10 05 43 CEST 2005 time 0 0 sec Exiting GoDIET Bye On of the main problems when writing a GODIET XML input file is to be compliant with the dtd A good tool to validate a GODIET file before using GODIET is xmllint This tool exist on most platforms and with the followin
107. lt host port gt of the name server To launch the omniORB name server first check that the path of the omniORB libraries is in your environment variable LD_LIBRARY_PATH then specify the log directory through the environment variable OMNINAMES_LOGDIR or with omniORB 4 at compile time through the with omniNames logdir option of the omniORB configure script If there are no log files in this directory omniNames needs to be intialized It can be launched as follows gt omniNames start Tue Jun 28 15 56 50 2005 Starting omniNames for the first time Wrote initial log file Read log file successfully Root context is IOR 010000002b00000049444c3a6f6d672e61f726721f4361f734e616d696e672f4e61 6d696e67436f6e746578744578743a312e300000010000000000000060000000010102000d0000003134 302e37372e31332e34390000f90a0b0000004e616d655365727669636500020000000000000008000000 0100000000545441010000001c0000000100000001000100010000000100010509010100010000000901 0100 Checkpointing Phase 1 Prepare Checkpointing Phase 2 Commit Checkpointing completed This sets an omniORB name server which listens for client connections on the default port 2809 If omniNames has already been launched once e there are already some log files in the log directory using the start option causes an error The port is actually read from old log files gt omniNames start Tue Jun 28 15 57 39 2005 Error log file tmp omninames toto log exists Can
108. lue amp time nb_c 6 order DIET_ROW_MAJOR Here are the lines of C code to generate such a profile double factor double matrix float time Init matrix at least factor and time too would be better diet_profile_t profile diet_profile_alloc 0 1 2 last_in last_inout last_out diet_scalar_set diet_parameter profile 0 amp factor 0 DIET_DOUBLE diet_matrix_set diet_parameter profile 1 matrix 0 DIET_DOUBLE 5 6 DIET_ROW_MAJOR diet_scalar_set diet_parameter profile 2 NULL 0 DIET_FLOAT NB1 If there is no IN argument last_in must be set to 1 if there is no INOUT argument last_inout must be equal to last in and if there is no OUT argument last out must be equal to last_inout NB2 The value argument for set functions 3 4 1 is ignored for OUT arguments since DIET allocates the necessary memory space when the corresponding data are transferred from the server so set value to NULL 3 7 2 Example 2 using persistency Let us consider the following problem C A x B with A B and C persistent matrices double A B C matrices initialization diet_initialize strcpy path MatPROD profile diet_profile_alloc path 1 1 2 diet_matrix_set diet_parameter profile 0 A DIET _PERSISTENT DIET DOUBLE mA nA oA print _matrix A mA nA oA DIET_ROW_MAJOR diet_matrix_set diet_parameter profile 1 B DIET_PERSISTENT DIET_DOUBLE mB nB oB pr
109. m is as follow diet_initialize create the workflow profile call the method diet_wf_call if success retrieve the results free the workflow profile diet_finalize The following tables show a description of methods provided by the DIET workflow API The table 14 1 contains the main methods that are common to the DAG workflows API and to the functional workflows API The table 14 2 contains the methods that are specific to the DAG API The table 14 3 contains the methods that are specific to the functional workflows API 14 5 2 The simplest example This example represents the basic client code to execute a DAG Line 26 indicates that the workflow output is a double value named n4 out4 The example shown in Figure 14 1 INRIA ENS Lyon UCBL Page 124 DIET User s Manual eae E Workflow function Description diet_wf_desc_t diet_wf_profile_alloc const char wf_file name const char wf_name wf_level_t wf_level allocate a workflow profile to be used for a workflow submission wf_file name the file name containing the workflow XML description wf_name the name of the workflow used for logs wf level specifier for workflow type DAG or FUNCTIONAL void diet_wf_profile_free diet_wf_desc_t profile free the workflow profile diet_error_t diet wf_call diet wf_desc_t wf_ profile execute the workflow associated to profile wf_profile int diet_wf_print_results diet_wf_desc_t
110. make INRIA ENS Lyon UCBL Page 21 DIET User s Manual BOSE E Compiling stage You are now done with the configuration stage equivalent of both the bootstrap sh and configure stage of the autotools You are now back to your platform level development tools i e make when working on Unices Hence you can now proceed with the compiling process by launching make Testing If you configured DIET with the BUILD_TESTING you can easily run the regression tests by invoking the make test This is equivalent to invoking ctest command ctest is part of cmake package ctest help provides a summary of the advanced options of ctest among which we recommend the verbose option Installation stage After compiling linking and testing you can optionally proceed with the installation stage with the make install command 2 3 Diet client server examples A set of various examples of DIET server client are provided within the DIET archive here are some of the provided examples e Batch A simple basic example on how to use the batch API is given here no IN or INOUT args the client receives as a result the number of processors on which the service has been executed The service only writes to a file with batch independent mnemonics some information on the batch system e BLAS the server offers the dgemm BLAS functionality We plan to offer all BLAS Basic Linear Algebraic Subroutines in the future Since this function compute
111. measurement and eval uation tools are developed both within and external to the DIET project see Chapter 8 some tools are already available to enable such information to be incorporated in the context of the plugin scheduler e Enhanced aggregation methods The plugin scheduler implemented in the current release enables the DIET system to account for user defined factors in the server selection process However the priority aggregation method is fairly rudimentary and lacks the power to express many imaginable comparison mechanisms We plan to investigate meth ods to embed code into DIET agents e g a simple expression interpreter in a manner that is secure and that preserves performance INRIA ENS Lyon UCBL Page 73 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 74 DIET User s Manual BASE E Chapter 8 Performance prediction 8 1 Introduction As we have seen in Chapter 7 the agent needs some information from the SeD to make an optimal scheduling This information is a performance prediction of the SeD The agent will ask the SeD to fill the data structure defined in Chapter 7 with the information it needs The SeD returns the information and the agent can make the scheduling Performance prediction can be based on hardware information the charge of the SeD the charge of the CPU of the memory or an advanced performance prediction can combine a set of basic performance predictions It is possible to
112. ments To set INOUT and OUT arguments use the diet_ _desc_set defined in DIET_server h these are helpful for writing solve functions only Using these functions the server developer must keep in mind the fact that he cannot alter the memory space pointed to by value fields on the server Indeed this would make DIET confused about how to manage the data If value is NULL or if order is DIET_MATRIX_ORDER_COUNT or if nb_rows or nb_cols is 0 or if path is NULL then the corresponding field is not modified int diet_scalar_desc_set diet_data_t data void value No use of diet_vector_desc_set size should not be altered by server You can alter nb_r and nb_c but the total size must remain the same int diet_matrix_desc_set diet_data_t data size_t nb_r size_t nb_c diet_matrix_order_t order No use of diet_string_desc_set length should not be altered by server int diet_file_desc_set diet_data_t data char path 5 2 Server API Defining services First declare the service s that will be offered Each service is described by a profile descrip tion called diet_profile_desc_t since the service does not specify the sizes of the data The diet profile desc t type is defined in DIET_server h and is very similar to diet profile t And the server developer should not be confused by the fact that diet_scalar_desc_set uses a value since scalar values are copied into the data descriptor 2T
113. module containing a UserScheduler class overloading the aggregate method GlobalScheduler aggregate corba_response_t aggrResp size_t max_srv const size_t nb_responses const corba_response_t responses DIET_AGG_PRIORITY DIET_AGG_USER or DIET_AGG_DEFAULT Scheduler aggregate UserScheduler aggregate DIET_AGG_DEFAULT DIET AGG_DEFAULT Load an external module to instanciate the scheduler wit an FAST NWS estimations no FAST NWS estimations NWSScheduler aggregate RRScheduler aggregate NewScheduler aggregate DIET_AGG DEFAULT and no FAST NWS and no timestamp DIET_AGG_PRIORITY wit user defined metrics FastScheduler MaxScheduler aggregate aggregate Ce NewScheduler cc med RandScheduler MinScheduler N NewScheduler so aggregate aggregate h NewScheduler hh Figure 7 2 Schedulers classes organization in DIET The user defined aggregation method just needs to sort the responses from the SeDs By locating the aggregation method on the agent we can use different scheduling strategies which could not be implemented at the SeD level These schedulers can also avoid some scheduling problems while submitting asynchronous jobs with Round Robin schedulers for example 7 5 3 The UserScheduler class This section presents how the sch
114. n fact this is quite probable as the lifetime of the distributed application can be much longer than the lifetime of a tool As a consequence the system state must be maintained and stored In order to maintain a system state in a general way LogService does not store the system state itself but all messages which are required to construct it Those messages are identified by their tag and stored in a special list This list is forwarded to each tool that connects For the tool this process is transparent since it simply receives a number of messages that represent the state of the application In order to further refine this concept the list of important messages can also be cleaned up by LogService This is necessary as components may connect and discon nect at runtime After a disconnection of a component the respective information is no longer relevant for the system state Therefore all messages which originated at this component can be removed from the list They have become obsolete due to the disconnection of the component and can be safely deleted in order to reduce the length of the list of important messages to a minimum All DIET components implement the LogComponent interface By using LogCentral the DIET architecture is able to relay information to LogCentral and then it is possible to connect to LogCentral by using a LogTool to collect store and analyse this information LogService is available for download See the web pag
115. n outLoop ports according to the as signments of the lt do gt tag semi column separated list of assignments The outputs of the processors that are iterated can be connected to the inLoop ports when the results of one iteration are used by the next one but this is not mandatory When the while con dition is evaluated to false the outLoop data items are handed over to the corresponding out ports according to the final attribute of these They are then sent to the connected processors Finally for one instance of this while processor N gt 0 iterations are done for proces sors connected to the outLoop ports and one data item is produced by the out port s lt loop name WHILE_Example gt INRIA ENS Lyon UCBL Page 123 DIET User s Manual BOSE E lt REQUIRED nb of IN ports EQUALS nb of OUT ports gt lt in name v type double gt lt out name out type double gt lt inLoop name v_1 type double init v gt lt outLoop name 1 type double final out gt lt WHILE Condition must be written in XQuery language gt lt it can contain ONLY LOOP IN ports gt lt while gt v 1t 100 lt while gt lt DO maps the inLoop ports to the outLoop ports straightforward gt lt do gt l v_1 lt do gt lt loop gt 14 5 Client API 14 5 1 Structure of client program The structure of a client program is very close to the structure of usual DIET client The general algorith
116. nction to diet_service_table_remove If you want to remove another service you need to build its profile description just as if you wanted to create a new service and pass it to diet_service_table_remove_desc 16 2 2 Example The following example present in src examples dyn_add_rem initially declares one service This service receives an integer n as parameter It creates n services and removes the service that has just been called Hence a service can only be called once but it spawns n new services include lt iostream gt include lt sstream gt include lt cstring gt include DIET_server h include DIET_Dagda h begin function prototypes int service diet_profile_t pb int add_service const char service_name end function prototypes static unsigned int NB 1 template lt typename T gt std string toString Tt std ostringstream oss oss lt lt t return oss str Solve Function int service diet_profile_t pb int nb INRIA ENS Lyon UCBL Page 143 DIET User s Manual BOSE E if pb gt pb_name std cout lt lt Executing lt lt pb gt pb_name lt lt std endl else std cout lt lt ERROR No name for the service lt lt std endl return 1 diet_scalar_get diet_parameter pb 0 amp nb NULL std cout lt lt Will create lt lt nb lt lt services lt lt std endl for int i 0 i lt nb i 4
117. nd one batch and one synchronous and one asynchronous client The client is configurable to simply ask for only sequential or explicitly parallel services or to let DIET choose the best by default two processors are used and the scheduling algorithm is Round Robin We consequently give the MPI code which is called from the batch SeD which realizes the concatenation of two files sent by the client Note that the user must change some paths in the SeD codes according to the site where he deploys DIET INRIA ENS Lyon UCBL Page 51 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 52 DIET User s Manual BOSE E Chapter 7 Scheduling in DIET 7 1 Introduction We introduce a plugin scheduling facility designed to allow DIET service developers to define application specific performance measures and to implement corresponding scheduling strate gies This section describes the default scheduling policy in DIET and the interface to the plugin scheduling facility 7 2 Default Scheduling Strategy The DIET scheduling subsystem is based on the notion that for the sake of system efficacy and scalability the work of determining the appropriate schedule for a parallel workload should be distributed across the computational platform When a task in such a parallel workload is submitted to the system for processing each Server Daemon SeD provides a performance estimate a collection of data pertaining to the capabilities
118. nd then copied into lt install_dir gt etc cfgs INRIA ENS Lyon UCBL Page 23 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 24 DIET User s Manual BOSE E Chapter 3 DIET data It is important that DIET can manipulate data to optimize copies and memory allocation to estimate data transfer and computation time etc Therefore the data must be fully described in terms of their data types and various attributes associated with these types 3 1 Data types DIET defines a precise set of data types to be used to describe the arguments of the services on the server side and of the problems on the client side The DIET data types are defined in the file lt install_dir gt include DIET_data h The user will also find in this file various function prototypes to manipulate all DIET data types Please refer to this file for a complete and up to date API description To keep DIET type descriptions generic two main sets are used base and composite types 3 1 1 Base types Base types are defined in an enum type diet_base_type_t and have the following semantics Type Description Size in octets DIET_CHAR Character 1 DIET_SHORT Signed short integer 2 DIET_INT Signed integer 4 DIET_LONGINT Long signed integer 8 DIET_FLOAT Simple precision real 4 DIET_DOUBLE Double precision real 8 DIET_SCOMPLEX Simple precision complex 8 DIET_DCOMPLEX Double precision complex 16 NB DIET_SCOMPLEX an
119. nses in responses e const corba_response_t responses the responses are stored in this argument It is an array of corba_response_t which is a CORBA structure containing a CORBA sequence of corba_server_estimation_t Two functions are defined to simplify the aggregation of the results typedef list lt corba_server_estimation_t gt ServerList ServerList CORBA_to_STL const corba_response_t responses int nb_responses void STL_to_CORBA ServerList amp servers corba_response_t kaggrResp The first function converts the received CORBA sequence into a STL list This function make the first aggregation of the results by marshalling all the sequences into one The second function converts a STL list into a CORBA sequence that can be transfer ed by DIET Then an aggregate function should start by a call to the CORBA_to_STL function The obtained list can then be sorted filtered using all the STL list facilities And to finish the result list is computed by the STL to CORBA function Several macros are defined to simplify the sort of a STL list SORTFUN name metric SORTFUN_NB name metric nb REV_SORTFUN name metric REV_SORTFUN_NB name metric nb These macros allow the developer to automatically define a sort function using a metric value For example to define a sort function using the number of CPUs the developer just has to declare SORTFUN compfun NBCPU The SORTFUN_NB macro is used for the multi values metrics f
120. nt this message check check the platform status stop_check stop the platform status then check its status before exit exit exit GoDIET do not change running platform We will now launch this example note that this example is intentionally very simple with all components running locally to provide initial familiarity with the GODIET run procedure Deployment with GODIET is especially useful when launching components on multiple remote machines GoDIET gt launch Launching DIET platform at Wed Jul 13 09 57 03 CEST 2005 Local scratch directory ready home hdail tmp scratch_godiet Launching element OmniNames on localHost Writing config file omni0RB4 cfg Staging file omni0RB4 cfg to localDisk Executing element OmniNames on resource localHost Waiting for 3 seconds after service launch Launching element MA_O on localHost Writing config file MA_0 cfg Staging file MA_0 cfg to localDisk Executing element MA_0 on resource localHost Waiting for 2 seconds after launch without log service feedback Launching element LA_O on localHost Writing config file LA_0 cfg Staging file LA_O cfg to localDisk Executing element LA_O on resource localHost Waiting for 2 seconds after launch without log service feedback Launching element SeD_O on localHost Writing config file SeD_0 cfg Staging file SeD_0 cfg to localDisk Executing element SeD_0 on resource localHost Waiting for 2 seconds after launch without log service feedbac
121. nternally and corrects the time stamps of incoming messages accordingly The time difference is correcting by using a time difference measurement recorded during the last ping that LogCentral has sent to the LogComponent pings are sent periodically to verify the aliveness of the LogComponent However incoming messages are still unsorted Thus the messages are buffered for a short period of time in order to deliver a sorted stream of messages to the tools Messages that arrive out of order within this time are sorted in the buffer and can thus be properly delivered Although this induces a delivery delay for messages this mechanism guarantees the proper ordering of messages within a certain tolerance As tools should not rely on true real time delivery of messages this short delay is acceptable INRIA ENS Lyon UCBL Page 96 DIET User s Manual BOSE TE The System State Problem A problem that arises in distributed environments is the state of the application This state may for example contain information on con nected servers their relationships the active tasks and many other pieces of infor mation that depend on the application The system state can be constructed from all events that occurred in the application Some tools rely on this state to work properly The problem emerges if those specific tools do not receive all messages This might occur as tools can connect to the monitor after the application has been started I
122. ntf get result d diet_wf_scalar_get n4 out4 amp 1 printf ld n long 1 diet_wf_free profile return 0 14 6 Scheduling The MApac agent may receive many requests to execute workflows from one or several clients and the number of ressources to execute all tasks in parallel may not be sufficient on the grid In this case the choice of a particular workflow scheduler is critical to determine the order of execution of all tasks that are ready to be executed Schedulers provide different online scheduling heuristics that apply different priori tization algorithms to choose the order of execution between tasks of the same DAG intra DAG priority and between tasks of different DAGs inter DAG priority All INRIA ENS Lyon UCBL Page 127 DIET User s Manual BOSE TE heuristics are based on the well known HEFT heuristic that is extended to this case of online multi workflow scheduling 14 6 1 Available schedulers The available MApag workflow schedulers are e A basic scheduler option basic or default choice this scheduler manages the precedence constraints between the tasks The priority between tasks within a dag is set according Heterogeneous Earliest Finish Time HEFT 25 heuristic When a task is ready to be executed 1 e the preceding tasks are completed the ready task with the higher HEFT rank is sent to the client for execution without specifying a ressource Then the client performs
123. o DIET 4 1 Structure of a client program Since the client side of DIET is a library a client program has to define a main function that uses DIET through function calls The complete client side interface is described in the files DIET_data h see Chapter 3 and DIET_client h found in lt install_dir gt include Please refer to these two files for a complete and up to date API description and include at least the latter at the beginning of your source code DIET_client h includes DIET_data h include lt stdio h gt include lt stdlib h gt include DIET_client h int main int argc char argv diet_initialize configuration_file argc argv Successive DIET calls diet_finalize The client program must open its DIET session with a call to diet_initialize which parses the configuration file to set all options and get a reference to the DIET Master Agent The session is closed with a call to diet_finalize which frees all resources associated with this session on the client Note that memory allocated for all INOUT and OUT arguments brought back onto the client during the session is not freed during diet_finalize this allows the user to continue to use the data but also requires that the user explicitly free the memory The user must also free the memory he or she allocated for IN arguments 1 Application programming interface INRIA ENS Lyon UCBL Page 35 DIET User s Manual BOSE TE 4 2 Client
124. o allow clients to send computation requests to the SeD pies The SeD pie receives the requests sent by clients calls the SeD piet that returns the response and then sends the result to the client The general behavior of the SeD z is written below e launch a new JXTA peer e create an input pipe to receive the clients requests launch the SeD piet e process each request by a thread that forwards the packed request received to the SeD pie and waits for a packed response sends the response to the client after having bound an output pipe to its input pipe The Multi MAj The Multi MA is composed of all MAs z running at the same time The MA is able to connect the clients to others running MAs Thus each client knows only one MAJ that is its access to the Multi MA Each MA publishes an advertisement with a lifetime in order to avoid clients or other MA to connect to a stopped MA When it receives a request coming from a client the MA submits the problem description to DIET via the MA piet it has itself launched If the submission returns a DIET failure the MA searches other MAs Then it forwards the client s request to other MAs 7 SeD references thus collected are merged and sent to the client The general algorithm of the MA is as follows e launch a new JXTA Peer e build an input pipe to listen to clients requests or agents forwarded requests INRIA ENS Lyon UCBL Page 107 DIET User
125. o associated with a latency which is time between the end of a FindRequest and the beginning of a SolveRequest VizDIET offers the possiblity to visualize all of these requests from either the point of view of the DIET platform in which case you will see the DIETRequests or in the point of view of the Agents or SeDs in which case you will see respectively the FindRequest and the SolveRequest The different kinds of requests are represented in different types of graphics such as a Gantt chart taskflow chart or bar chart VizDIET also computes some other statistics for the platform such as average time for scheduling for solving or latency This information can be see for the whole service in the platform or for one specific service VizDIET has one other interesting feature the possibility to export all data collected by VizDIET into a file using a format that you specify Finally VizDIET is quite useful for understanding the behavior of the DIET hierarchy INRIA ENS Lyon UCBL Page 99 DIET User s Manual BOSE E and quite simple to use You have to keep in mind that VizDIET bases its information upon log information that is forwarded by LogCentral from DIET components Therefore the information displayed and computed in VizDIET is limited to the DIET hierarchy e g there is no information about clients Future development of VizDIET will depend on new developments in DIET For example a new integration between DIET and
126. ology to contact the data managers Among the data managers having one replicate of the data DAGDA chooses the best source to transfer it To make this choice DAGDA uses some statistics collected from previous data transfers between the nodes By not using dynamic infor INRIA ENS Lyon UCBL Page 131 DIET User s Manual BOSE TE mation it is unsure that DAGDA really chose the best nodes for the transfers In a future version we will introduce some facilities to estimate the time needed to transfer a data and to improve the choice of a data stored on the grid To do the data transfers DAGDA uses the pull model It is the destination node that ask for the data transfer Be SO DAGDA Client Selected SeD s 2 component 5 client request 1 D DAGDA client inpu component data ob submission 3 L Y SeD output data DAGDA component CR y DAGDA DAGDA persistent input component component data from another T ooa component input data update Figure 15 1 DAGDA architecture in DIET Figure 15 1 presents how DAGDA manages the data when a client submit a job In this example the client wants to use some data stored on the grid and some personal data He wants to obtain some results and to store some others on the grid Some of these output data are already stored on the platform and they should be updated after the job execution 1 The client sends a request to the Maste
127. on of performance information and the treatment of these data in the context of the simulated database search Finally it should be noted that the existence of a plugin scheduler is completely transparent to the client and thus client code need not be changed 7 5 Scheduler at agents level In this section we introduce a new way to define a scheduling policy in DIET Some scheduling strategies could not be developed using only the DIETSeDs plugins The schedulers at agents INRIA ENS Lyon UCBL Page 59 DIET User s Manual BOSE E level allow the developer to design every scheduler strategies even the centralized ones The first two sections explain precisely how DIET performs the scheduling The third section enters in the DIET source code and can be ignored by most of the users The fourth section presents the tools provided to make an agent scheduler easily The fifth section deals with the scheduler module compilation and usage The last section presents some scheduler examples 7 5 1 Scheduling from the agents side In DIET the scheduling works as follows see Figure 7 1 for a representation of each step e A request is submitted to the Master Agent step 1 e The Master Agent forwards the request to the Local Agents and SeDs that it manages step 2 e The SeDs which dispose of the asked service return a CORBA response structure which contains an estimation metric vector step 3 e According to a default policy or a u
128. or example the CPU cache for each CPU The nb value designates which value has to be used to sort the list The REV_ functions are used to sort in ascending order To see all the metrics available for the SORTFUN macro see Section 7 5 4 When a sort function has been defined the developer can use the SORT macro to sort the STL list For example with our compfun function SORT serverList compfun This call sorts the server STL list in decreasing order of the number of CPU INRIA ENS Lyon UCBL Page 67 DIET User s Manual BOSE E An example of aggregate method definition We will now present an example of an aggregate method using the functions and macro defined in the UserScheduler hh file SORTFUN compCPU NBCPU SORTFUN_NB compCache CACHECPU 0 REV_SORTFUN compDiskRead DISKACCESSREAD int MyScheduler aggregate corba_response_t aggrResp size_t max_srv const size_t nb_responses const corba_response_t responses ServerList candidates CORBA_to_STL responses nb_responses SORT candidates compCache SORT candidates compCPU SORT candidates compDiskRead STL_to_CORBA candidates aggrResp return 0 This function returns a list sorted by increasing disk access for first criteria and by decreasing CPU number and decreasing CPU cache Access the metric values through macros To simplify the access to some specific values defined inside the SeD you can use these macros TOTALTIME S
129. ori function on the SeD For more information about CoRI see Section 8 3 Then on the agents scheduler side these information are accessed using one of the previously presented macro You also can obtain the user defined information by using the USERDEFINED SeD nb macro These information have been defined on the SeDs metric function using the diet_est_set est Vector t ev int nb double value For more information on how to get performance prediction values please consult Chapter 8 7 5 7 A complete example of scheduler This example source code is available on the src examples agent_scheduler directory The scheduler performs a Round Robin on the SeDs using their hostname to evaluate the number of executions For example if the agent is connected to three SeDs with two launched on the same machine the number of jobs executed on the machine with two SeDs will be at most one more than the number of executed jobs on the other machine Hostname based Round Robin plugin scheduler include GlobalSchedulers hh include UserScheduler hh include est_internal hh include lt map gt std map lt std string unsigned int gt hostCounter class HostnameRR public UserScheduler public static const char stName HostnameRR INRIA ENS Lyon UCBL Page 71 DIET User s Manual BOSE E HostnameRR void init static char serialize HostnameRR GS static HostnameRR deserialize const char serializedSche
130. oshi Matsuoka Jack Dongarra Craig Lee and Henri Casanova Overview of GridRPC A Remote Procedure Call API for Grid Computing In Manish Parashar editor Grid Computing GRID 2002 Third International Workshop Baltimore MD USA November 18 2002 Proceedings volume 2536 of LNCS pages 274 278 Springer 2002 Haluk Topcuouglu Salim Hariri and Min you Wu Performance effective and low complexity task scheduling for heterogeneous computing IEEE Trans Parallel Dis trib Syst 13 3 260 274 2002 INRIA ENS Lyon UCBL Page 158 DIET User s Manual BOSE TE 26 R Wolski N T Spring and J Hayes The Network Weather Service A Distributed Resource Performance Forecasting Service for Metacomputing Future Generation Computing Systems Metacomputing Issue 15 5 6 757 768 Oct 1999 INRIA ENS Lyon UCBL Page 159
131. oy DIET_JXTA architectures Note that this is to be opposed with DIET_WITH_MULTI_MA see 2 2 5 below which offers similar functionalities but based on CORBA e DIET_USE_WORKFLOW enables the support of workflow For the support of workflows in side DIET Xerces and Xqilla libraries are mandatory see http xerces apache org xerces c and http xqilla sourceforge net HomePage For more details about the workflow support in DIET see chapter 14 Note that setting the DIET_USE_WORKFLOW will force the option DIET_USE_DAGDA to be set e DIET_WITH_MULTI_MA activates the so called MULTI Master Agent support which allows the user to connect several MA for them to act as bounded When this option is activated such a bounded MA is allowed to search for a SeD into the MA hierarchies it is connected to Note that MULTI Master Agent support is based on the CORBA layer which is to be opposed with DIET_USE_JXTA which offers similar functionalities but based on JXTA e DIET_WITH_STATISTICS enables the generation of statistics logs 2 2 6 DIET s advanced configuration flags Eventually some configuration flags control the general result of the compilation or some developers extensions e BUILD_TESTING is a conventional variable which is not a cmake internal variable which specifies that the regression tests should also be compiled e BUILD_SHARED_LIBS is a cmake internal variable which specifies whether the libraries should be dynamics as opposed to s
132. pers this directories tree simply corresponds to the directory GRAAL devel diet diet of a cvs checkout of the DIET sources hierarchy Additionally assume we created a build tree directory and cd to it in the example below we chose CVS_DIET_HOME Bin as build tree but feel free to follow your conventions e cd CVS_DIET_HOME Bin e ccmake to enter the GUI press c equivalent of bootstrap sh of the autotools toggle the desired options e g DIET_BUILD_EXAMPLES or DIET_USE_JXTA specify the CMAKE_INSTALL_PREFIX parameter if you wish to install in a directory different from usr local press c again for checking required dependencies check all the parameters preceded with the star character whose value was au tomatically retrieved by cmake provide the required information i e fill in the proper values for all parameters whose value is terminated by NOT FOUND iterate the above process of parameter checking toggle specification and configura tion until all configuration information is satisfied press g to generate the makefile press q to exit ccmake e make in order to classically launch the compilation process e make install when installation is required 2 2 4 DIET s main configuration flags Here are the main configuration flags e OMNIORB4_DIR is the path to the omniORB4 installation directory only relevant when omniORB4 was not installed in usr local Example cmake DO
133. ption The maximum memory space used by DAGDA to store the data If set to 0 DAGDA will not take care of the memory usage By default no maximum memory usage is set Same effect than to choose 0 cacheAlgorithm Component All Mode DAGDA Type String Description The cache replacement algorithm used when DAGDA needs more space to store a data Possible values are LRU LFU FIFO By default no cache replacement algorithm DAGDA never replace a data by another one shareFiles e Component Agent e Mode DAGDA INRIA ENS Lyon UCBL Page 154 DIET User s Manual BOSE TE e Type Boolean e Description The DAGDA component shares its file data with all its children when the path is accessible by them for example if the storage directory is on a NFS partition Value can be 0 or 1 By default no file sharing 0 dataBackupFile e Component Agent and SeD e Mode DAGDA e Type String e Description The path to the file that will be used when DAGDA save all its stored data data path when asked by the user Checkpointing By default no checkpointing is possible restoreOnStart e Component Agent and SeD e Mode DAGDA e Type Boolean e Description DAGDA will load the dataBackupFile file at start and restore all the data recorded at the last checkpointing event Possible values are 0 or 1 By default no file loading on start 0 storageDirectory e Component All e Mode DAGDA or Batch e Type String
134. r run make install command in src examples directory INRIA ENS Lyon UCBL Page 85 DIET User s Manual BOSE TE Tracing API traceLevel default 1 This option controls debugging trace output The following levels are defined level 0 Print only errors level lt 5 Print errors and messages for the main steps such as Got a request default level lt 10 Print errors and messages for all steps level 10 Print errors all steps and some important structures such as the list of offered services level gt 10 Print all DIET messages AND omniORB messages corresponding to an omniORB traceLevel of level 10 Client parameters MAName default none This is a mandatory parameter that specifies the name of the Master Agent to connect to The MA must have registered with this same name to the CORBA name server Agent parameters agentType default none As DIET offers only one executable for both types of agent it is mandatory to specify which kind of agent must be launched Two values are available DIET_MASTER_AGENT and DIET_LOCAL_AGENT They have aliases respectively MA and LA name default none This is a mandatory parameter that specifies the name with which the agent will register to the CORBA name server LA and SeD parameters parentName default none This is a mandatory parameter for Local Agents and SeDs but not for the MA It indicates the name of the parent an LA or the MA to register to
135. r Agent 2 The Master agent returns one or more SeD references 3 The client sends its request to the chosen node The parameters data are identified by a unique ID and the problem profile contains a reference to the client s data manager 4 Receiving the request the SeD asks the client to transfer the data of the user and it asks to the DAGDA architecture to obtain the persistent data already stored on the platform INRIA ENS Lyon UCBL Page 132 DIET User s Manual eae E 5 The SeD executes the job After the execution the SeD stores the output data and it informs the client that the data are ready to be downloaded It also asks the architecture to update the modified output data 6 The client upload its results and the data are updated on the nodes 15 2 The DAGDA configuration options DAGDA introduces new configuration options that can be defined for all the DAGDA components None of these options are mandatory to use DAGDA Figure 15 2 presents all the DAGDA available options their meaning and default values Option Description Default value Q Ppl Y og 2 J JE storageDirectory 2e directory on which DAGDA will The tmp directory ViVviv The maximum size of a CORBA mes The omniORB giopMarMsg maxMsgsize sage sent by DAGDA Size size ViVviv The maximum disk space used by i DAGDA to store the data If set to 0 po erica disk space on the maxDiskSp
136. r the problem has been solved when they will no longer be used NB for files this function removes the file and frees the path since it has been dynamically allocated by DIET in both cases 1 OOO OR lolololok kk int diet_free_data diet_arg_t arg INRIA ENS Lyon UCBL Page 30 DIET User s Manual BOSE E 3 6 Problem description For DIET to match the client problem with a service servers and clients must speak the same language ie they must use the same problem description A unified way to describe problems is to use a name and define its profile with the type diet_profile_t typedef struct char pb_name int last_in last_inout last_out diet_arg_t parameters diet_profile_t The field parameters consists of a diet_arg_t array of size last_out 1 Arguments can be IN The data are sent to the server The memory is allocated by the user INOUT The data are allocated by the user as for the IN arguments then sent to the server and brought back into the same memory zone after the computation has completed without any copy Thus freeing this memory at the client while the computation is performed on the server would result in a segmentation fault when the data are brought back onto the client OUT The data are created on the server and brought back into a newly allocated zone on the client This allocation is performed by DIET After the call has returned the user can find the resul
137. r thread It only releases this reference after a call to the corresponding waiting function The client developer should always use these functions that s why a data ID is only returned by the dagda_wait_ and dagda_wait_data_ID functions e Without end of transfer control The data is loaded from to the DAGDA hierarchy without the possibility to wait for the end of the transfer These functions should only be called from an agent plugin scheduler a SeD plugin scheduler or a SeD if the data transfer without usage of the data is one of the objectives of the called service The data adding functions without control should be used very carefully because there is no way to be sure the data transfer is achieved or even started With asynchronous transfers the user should take care of the data lifetime because DAGDA does not duplicate the data pointed by the passed pointer For example if the program uses a local variable reference to add a data to the DAGDA hierarchy and goes out of the variable scope a crash could occured because the data pointer could be freed by the system before DAGDA has finished to read it DAGDA asynchronous put macros functions The arguments to these functions are the same than for the synchronous ones See Section 15 4 2 for more details All of these functions return a reference to the data transfer which is an unsigned int This value will be passed to the dagda_wait_data_ID function INRIA ENS Lyon
138. rator gt gt can be ignored too e CORBA ULong reqID this field contains the ID of the request e CORBA Long myID this field is for DIET internal usage The developer should ignore it e SeqServerEstimation_t servers this field is a sequence of corba_server_estimation_t It is used to store the SeDs references returned by the aggregate method This is the field that has to be sorted filtered The corba_server_estimation_t is defined as follows struct corba_server_estimation_t typedef _CORBA_ConstrType_Variable_Var lt corba_server_estimation_t gt _var_type corba_server_t loc corba_estimation_t estim void operator gt gt cdrStream amp const void operator lt lt cdrStream amp e corba_server_t loc this field is used to designate a particular SeD e corba_estimation_t estim this field contains the estimation vector for the designated SeD The corba_server_t loc structure is defined as follows struct corba_server_t typedef _CORBA_ConstrType_Variable_Var lt corba_server_t gt _var_type _CORBA_ObjRef_Member lt _objref_SeD SeD_Helper gt ior CORBA String_member hostName CORBA Long port void operator gt gt cdrStream amp const void operator lt lt cdrStream amp INRIA ENS Lyon UCBL Page 63 DIET User s Manual BOSE E The two interesting fields are e ior which is a CORBA reference to the SeD e hostName which is the hostname of the SeD The corba_estimation_
139. rid Forum Advanced Programming Models Working Group whitepaper Nagios http www nagios org H Nakada S Matsuoka K Seymour J Dongarra C Lee and H Casanova GridRPC A Remote Procedure Call API for Grid Computing In Grid 2002 Work shop on Grid Computing number 2536 in Lecture Notes in Computer Science pages 274 278 Baltimore MD USA November 2002 H Nakada M Sato and S Sekiguchi Design and Implementations of Ninf to wards a Global Computing Infrastructure Future Generation Computing Systems Metacomputing Issue 15 5 6 649 658 1999 http ninf apgrid org papers papers shtml OMNIORB http www uk research att com omniORB Andy Oram editor Peer to Peer Harnessing the Benefits of a Disruptive Technol ogy O Reilly 2001 The JuxMem project http juxmem gforge inria fr The JXTA project http www jxta org M Quinson Dynamic Performance Forecasting for Network Enabled Servers in a Metacomputing Environment In International Workshop on Performance Modeling Evaluation and Optimization of Parallel and Distributed Systems PMEO PDS 02 April 15 19 2002 Federico D Sacerdoti Mason J Katz Matthew L Massie and David E Culler Wide area cluster monitoring with ganglia 2003 M Sato M Hirano Y Tanaka and S Sekiguchi OmniRPC A Grid RPC Facility for Cluster and Global Computing in OpenMP Lecture Notes in Computer Science 2104 130 136 2001 Keith Seymour Hidemoto Nakada Sat
140. ro defined in UserScheduler hh automates this declaration You can simply define your class as a scheduler class by calling SCHEDULER_CLASS MyScheduler where MyScheduler is the name of the class which inherits of the UserScheduler class 7 5 4 Easy definition of a new scheduler class The previous section presented how the scheduler class loader is working Many things presented before can be automated The UserScheduler hh file defines some useful functions and macros to make a new scheduler class easily In this section we will present how to create a new scheduler class using these functions and macros The new class definition Every scheduler class has to inherit from the UserScheduler class The only redefinition needed is the aggregate function But the init serialize and deserialize functions have to be declared conforming to the C standard but not defined the inherited functions are sufficient The following example shows a simple scheduler class implementation INRIA ENS Lyon UCBL Page 65 DIET User s Manual BOSE TE class MyScheduler public UserScheduler public static const char stName MyScheduler MyScheduler void init static char serialize MyScheduler GS static MyScheduler deserialize const char serializedScheduler Overriden UserScheduler class aggregate method int aggregate corba_response_t aggrResp size_t max_srv const size_t nb_responses const corba_response_
141. s 16 2 1 5 Software dependencies o ss su Sb daa da 4 4 4 eee we a 16 2 2 Compiling the platform s o s s 6c 0 424344 4x a ee ee de 16 2 2 1 Obtaining and installing cmake per Se 16 2 2 2 Configuring DIET s compilation cmake quick introduction 16 2 2 3 A ccmake walk through for the impatients 18 2 2 4 DIET s main configuration flags sse ee ee ee 18 2 2 5 DIET s extensions configuration flags 19 2 2 6 DIET s advanced configuration flags 20 22 7 Compiling and installing s 4 4 4 2 4 cos o um ees 21 2 3 Diet client server examples 22 2 3 1 Compiling the examples 23 3 DIET data 25 3 1 Datatypes 2 5520 Luka da a des de a AGED ee wea 25 SLI Basetwpes 2 ias hehe eb ee fae se tes a a e G 25 3 1 2 Composite types o s si sos bebe eee ee pos eue 25 21 3 Persistenes made 2s 25k a nie wR ea a R A Oe ia 26 22 Data deecriptioi og acca AM ga aaa a Oe ROR a ke as 26 So Data Hana cementos oats ok Ak OR weak ee Roe he Soe Hat ee RE 26 INRIA ENS Lyon UCBL Page 3 DIET User s Manual BOSE E dl Data identiber s esas Le a ae a eg 26 Gam Data ble sk fate bo ew hee a Re te ee Pee Fa ee a 27 34 Manipulating DIET structures sos e m sacs 4 4 44 De a aoa Ca a ss 27 SA SOG MURS
142. s on Computations On PLatency e Feasibility e Space needed e Time needed Low Level Software Figure 8 1 FAST overview Of course you also need to install the FAST SDK itself It is important to basically understand how FAST works and the role of its dependencies to deactivate the ones that are not needed by the user 8 2 2 Using FAST in the plug in scheduler FAST and NWS based performance estimation metrics are stored in an estimation metric vector see Chapter 7 for more details by calling int diet_estimate_fast estVector_t ev const diet profile t const profilePtr with an appropriate value for ev and the profilePtr corresponding to the current DIET request Attention this option it not available when compiling with the option DDIET_USE_CORI set to OFF To access to this information use CoRI see Section 8 3 8 2 3 Building a server application with FAST Since performance prediction is performed only in the DIET SeD no modification is needed to the client code On the other hand at the SeD level the code must sometimes be adapted In the next subsection we explain convertors and show how they can be used in an example Using convertors The service profiles offered by DIET are sometimes not understandable by the service imple mentations To solve this problem a convertor processes each profile before it is passed to the implementation This is mainly used to hi
143. s C aAB C it can also compute a matrix matrix product a sum of square matrices etc All these services are offered by the BLAS server Two clients are designed to use these services one dgemm_client c is designed to use the dgemm_ function only and the other one client c to use all BLAS functions but currently only dgemm_ and sub services such as MatPROD e dmat_manips the server offers matrix manipulation routines transposition T product MatPROD and sum MatSUM SqMatSUM for square matrices and SqMatSUM_opt for square matrices but re using the memory space of the second operand for the result Any subset of these operations can be specified on the command line The last two of them are given for compatibility with a BLAS server as explained below e file_transfer the server computes the sizes of two input files and returns them A third output parameter may be returned the server decides randomly whether to send back the first file This is to show how to manage a variable number of arguments the profile declares all arguments that may be filled even if they might not be all filled at each request computation INRIA ENS Lyon UCBL Page 22 DIET User s Manual BOSE E e ScaLAPACK the server is designed to offer all ScaLAPACK parallel version of the LA PACK library functions but only manages the pdgemm_ function so far The pdgemm_ routine is the parallel version of the dgemm_ function so that the server also
144. s are varied and often exhibit performance behavior specific to the do main from which they arise Consequently application specific scheduling approaches are often necessary to achieve high performance execution We propose an extensible framework to build plugin schedulers enabling application developers to specify performance estimation metrics that are tailored to their individual needs 7 3 1 Estimation Metric Vector The new type estVector_t represents an estimation vector logically a structure that can man age a dynamic collection of performance estimation values It contains values that represent the performance profile provided by a SeD in response to a DIET service request This collection of values may include either standard performance measures that are available through DIET or developer defined values that are meaningful solely in the context of the application being developed 7 3 2 Standard Estimation Tags To access to the different fields of the estVector_t it is necessary to specify the tag that correspond to a specific information type Table 7 1 describes this correspondence Some tags represent a list of values one has to use the diet_est_array_ functions to have access to them In Table 7 1 the second column marks these multi value tags The tag ALLINFOS is a special his field is always empty but it allows to fill the vector with all known tags by the particular collector Standard Performance Metrics To acces
145. s return at the end of the transfer or if an error occured They all return an integer value 0 if the operation succeeds another value if it fails DAGDA put data macros functions The following functions put a data on the DAGDA hierarchy to be used later The last parameter is always a pointer to a C string which will be initialized with a pointer to the ID string of the data This string is allocated by DAGDA and can be freed when the user does not need it anymore The first parameter is always a pointer to the data For a scalar value a pointer on the data for a vector matrix or string a pointer on the INRIA ENS Lyon UCBL Page 134 DIET User s Manual BOSE E first element of the data The value argument for a file is a C string containing the path of this file The persistence mode for a data managed by DAGDA should allways be DIET_PERSISTENT or DIET_STICKY The VOLATILE and RETURN modes do not make sense in this data management context dagda_put_scalar void value diet_base_type_t base_type diet_persistence_mode_t mode char ID This macro adds to the platform the scalar data of type base_type pointed by value with the persistence mode mode DIET_PERSISTENT or DIET STICKY and initializes D with the ID of the data dagda_put_vector void value diet_base_type_t base_type diet_persistent_mode_t mode size_t size char ID This macro adds to the platform the vector of size
146. s to the existing default performance estimation routines as described in Chapter 8 the following functions are available to facilitate the construction of custom performance estimation INRIA ENS Lyon UCBL Page 54 DIET User s Manual BOSE TE Information tag multi Explication starts with EST_ value TCOMP the predicted time to solve a problem TIMESINCELASTSOLVE time since last solve has been made sec FREECPU amount of free CPU power between 0 and 1 FREEMEM amount of free memory Mb NBCPU number of available processors CPUSPEED x frequency of CPUs MHz TOTALMEM total memory size Mb AVGFREECPU average amount of free CPU power in 0 1 BOGOMIPS x CPUs bogomips CACHECPU x cache size CPUs Kb TOTALSIZEDISK size of the partition Mb FREESIZEDISK amount of free place on partition Mb DISKACCESREAD average time to read on disk Mb sec DISKACCESWRITE average time to write to disk sec ALLINFOS x empty fill all possible fields PARAL_NB_FREE_RESOURCES_IN_DEFAULT_QUEUE number of idle resources Table 7 1 Explication of the estimation tags functions e FAST and NWS based performance estimation metrics can be used in the plugin sched uler See the Section 8 2 2 for information on how to use them e The time elapsed since the last execution to enable the round robin scheduler is stored in an estimation metric vector by calling int diet_estimate_laste
147. sages too MAName e Component Client e Mode All e Type String e Description Master Agent name agent Type e Component Agent MA and LA e Mode All e Type Agent type e Description Master Agent or Local Agent As there is only one executable for both agent types it is COMPULSORY to specify the type of this agent DIET MASTER AGENT or MA or DIET LOCAL AGENT or LA INRIA ENS Lyon UCBL Page 147 DIET User s Manual BOSE TE diet Port e Component All e Mode All e Type Integer e Description the listening port of the agent If not specified let the ORB get a port from the system if the default 2809 was busy dietHost Name e Component All e Mode All e Type String e Description the listening interface of the agent If not specified let the ORB get the hostname from the system the first one if several one are available name e Component Agent and SeD e Mode All e Type String e Description The name of the element The ORB configuration files of the clients and the children of this MA LAs and SeDs must point at the same CORBA Naming Service as the one pointed at by the ORB configuration file of this agent parent Name e Component LA and SeD e Mode All e Type String e Description the name of the agent to which the element will register This agent must have registered at the same CORBA Naming Service that is pointed to by your ORB configuration fast Use e Compon
148. scheduler module The first step is to compile DIET activating the USERSCHED option With this option you ll find a subdirectory scheduler in the include directory of the DIET installation This directory contains all the headers needed to develop the basis class of the scheduler module A scheduler module needs to be linked with some libraries to compile e omniORB4 The basis omniORB library e omnithread The omniORB thread library e DIET libraries CorbaCommon The basis DIET Corba library UtilsCommon amp UtilsNodes The DIET utilities libraries IDLAgent amp IDLCommon The IDL DIET libraries UtilsVector The vector library internally used in DIET IDLLA amp IDLMA The agents libraries When using g as compiler the option shared has to be used to compile the module under Linux and dynamiclib under Mac OS X The fPIC has to be used for both operating systems How to configure the agent and the SeD to use a scheduler module On the agent side the parameter schedulerModule has to be set to the path of the module scheduler in the agent configuration file This option uses the same syntax than the other agents and ORB options schedulerModule lt path to module gt On the SeD side the developer has to choose DIET_AGG_USER as aggregator diet_aggregator_desc_t agg diet_service_table_init 1 profile diet_profile_desc_alloc serviceName diet_generic_desc_s
149. se the LogService for monitoring IsOutbuffersize e Component Agent and SeD e Mode All e Type Integer e Description the size of the buffer for outgoing messages IsFlushinterval e Component Agent and SeD e Mode All e Type Integer e Description the flush interval for the outgoing message buffer neighbours e Component MA e Mode Multi MA e Type String e Description A list of Master Agent that must be contacted to build a federa tion The format is a list of host port minimumNeighbours INRIA ENS Lyon UCBL Page 150 DIET User s Manual BOSE E e Component MA e Mode Multi MA e Type Integer e Description Minimum number of connected neighbours If the agent has less that this number of connected neighbours is going to find some new connections maximumNeighbours e Component MA e Mode Integer e Type Multi MA e Description maximum number of connected neighbours The agent does not accept a greater number of connection to build the federation than maximum Neighbours updateLinkPeriod e Component MA e Mode Multi MA e Type Integer e Description The agent check at a regular time basis that all it s neighbours are still alive and try to connect to a new one if the number of connections is less than minimumNeighbours updateLinkPeriod indicate the period in second between two checks bindServicePort e Component MA e Mode All e Type Integer e Description port us
150. ser defined one the responses from the SeDs are aggregated Then the responses sequence is sent to the parent agent which aggregates all the results of its children step 4 e When the aggregated responses reach the Master Agent it returns the aggregated list of all responses to the client step 5 e Finally the client chooses the better server according to the chosen aggregation method step 6 cpu3 gt cpui gt cpu2 u mem ti response 3 response 1 ime response 2 Sorted SeD list Figure 7 1 Scheduling steps in DIET INRIA ENS Lyon UCBL Page 60 eae E DIET User s Manual 7 5 2 Aggregation methods overloading To aggregate the responses of the SeDs DIET uses an aggregation method which is called by the agents This method is chosen from the SeDs by defining the aggregator type see Section 7 3 2 By default two aggregator types are proposed by DIET DIET_AGG_DEFAULT and DIET_AGG_PRIORITY In the last versions of DIET we introduced a new aggregator type DIET_AGG_USER Using this aggregator the user can define its own aggregation method to be used by the agents Figure 7 2 presents the global schedulers classes organization in DIET By choosing the DIET_AGG_USER aggregator the user commands the GlobalScheduler class to load an external
151. services available in the subtree rooted at the LA for each service LAs store a list of children agents or servers that can be contacted to find the service Depending on the underlying network topology a hierarchy of LAs may be deployed between an MA and the servers Of course the function of an LA is to doa partial scheduling on its subtree which reduces the workload at the MA Server Daemon SeD A SeD encapsulates a computational server For instance it can be located on the entry point of a parallel computer The information stored on a SeD is a list of the data available locally i e on the server the list of problems that can be solved on it and performance related information such as the amount of available memory or the number of resources available When it registers a SeD declares the problems it can solve to its parent LA or MA A SeD can give perfomance and hardware information by using the module CoRI or performance predictions for some types of problems by using the module FAST Both modules are described in Chapter 8 1 2 Communications layer NES environments can be implemented using a classic socket communication layer Several problems to this approach have been pointed out such as the lack of portability or limits on the number of sockets that can be opened concurrently Our aim is to implement and deploy a distributed NES environment that works at a wider scale Distributed object environments such as Java DCOM or C
152. setting configuration parameters in order to generate the makefiles in the form of two commands ccmake and cmake the first one has an extra c character ccmake options lt path to source gt in order to specify the parameters interactively through a GUI interface cmake options lt path to source gt D lt var gt lt type gt lt value gt in order to define the parameters with the D flag directly from the command line In the above syntax description of both commands lt path to source gt specifies a path to the top level of the source tree 1 e the directory where the top level CMakeLists txt file is to be encountered Also the current working directory will be used as the root of the build tree for the project out of source building is generally encouraged especially when working on a CVS tree Here is a short list of cmake internal parameters that are worth mentioning e CMAKE_BUILD_TYPE controls the type of build mode among which Debug will produce binaries and libraries with the debugging information e CMAKE_VERBOSE_MAKEFILE is a Boolean parameter which when set to ON will generate makefiles without the SILENT directive This is useful for watching the invoked com mands and their arguments in case things go wrong e CMAKE_C XX _FLAGS is a family of parameters used for the setting and the customization of various C C compiler options e CMAKE_INSTALL_PREFIX variable defines the location of the install direc
153. shboard 95 MOM Lor ermita iia da Ans a eb a UE ee ek ee BA ek RE 95 WO IET ere ie k ete Bee Pe ee Bee ok ee ee GR ee ee 97 11 Multi MA extension 101 11 1 Function of the Multi MA extension 101 11 2 Deployment example o cuida 4 da ma ane a bad cl a 101 11 3 o o 2 ci dacris HA Lis PO e ee DR Ag dan ia 103 12 P2P DIET extension DIET 105 12 1 PDP and IXTA cae 6 Oh dan De den en wee DRED a os 105 12 2 Description of the current architecture developed with JXTA 105 12 2 1 The JXTA components 106 12 2 2 Interfacing JATA and DIET with INT 2 24421 0 goe Lune 108 INRIA ENS Lyon UCBL Page 5 DIET User s Manual BOSE E 12 3 The future of DIETA os osos a a a ee 109 1231 Remsinino problems sos caia ne pile ha Ne Beg dans eng ae 6 109 12 4 Working with a DIET y platform s sae soca sensei sra N G 109 124 1 Installation and configuration 65 44 28 da bee ee nue BED 109 12 4 2 Deploying a DIET platform 110 13 JuxMem extension 113 13 1 Introduction c e sosa a p aaa ak aea a an eee eee ph eee 113 18 2 Overview of dim Mens ot bab ae Pe dun ee e Se Ruth 113 13 3 How to configure DIET to use JuxMem 113 13 41 AWE 000 ok ws GS A ee RE eed bee Ee eee eS 114 13 5 Troubleshooting o es L 4 48 de ewe Oa a eee ae Owe oe be ee A 114 14 Workflow management in Diet 115 A II 11
154. stVector_t diet_perfmetric_t diet_service_use_perfmetric diet_perfmetric_t perfmetric_fn Thus the type diet_perfmetric_t is a function pointer that takes as arguments a performance estimation represented by the estVector_t object and a DIET service request profile The application developer can associate such a function or performance estimation routine with DIET services via the diet service use perfmetric interface This interface returns the pre viously registered performance estimation routine if one was defined and NULL otherwise At this point a service added using the diet service table add function will be associated with INRIA ENS Lyon UCBL Page 56 DIET User s Manual BOSE TE the declared performance estimation routine Additionally a performance estimation routine so specified will be associated with all services added into the service table until another call to the diet_service_use_perfmetric interface is made In the performance estimation routine the SeD developer should store in the provided estimation vector any performance data used in the server response aggregation methods described in the next section 7 3 4 Aggregation Methods At the time a DIET service is defined an aggregation method the logical mechanism by which SeD responses are sorted is associated with the service the default behavior was described in Section 7 2 If application specific data are supplied 1 e the estimation
155. stalled on the node EST_CPUSPEED The frequencies of the CPUs of the node This value is accessed using the diet_est_array_get_internal function INRIA ENS Lyon UCBL Page 64 DIET User s Manual BOSE E EST_TOTALMEM The total memory of the node EST_AVGFREEMEM The average free memory on the node EST_AVGFREECPU The average free CPU computation power on the node EST_BOGOMIPS The computation power of the nodes CPUs given in bogomips EST_TOTALTIME The total time to execute the request evaluated by FAST FAST must be activated at the compilation time EST_TOTALSIZEDISK The total disk size on the node EST_FREESIZEDISK The available disk space on the node EST_DISKACCESREAD An evaluation of the disk read access performance EST_DISKACCESWRITE An evaluation of the disk write access performance EST_USERDEFINED The first user defined value EST_USERDEFINED n The nt user defined value To make the new scheduler class loadable by the GlobalScheduler class the developer has to define these two functions outside the class definition extern C GlobalScheduler x constructor return new MyScheduler extern C void destructor UserScheduler scheduler 4 delete scheduler No C implementation of dynamic class loading are defined in the C standard So the UserScheduler class has to use C functions to load an external module containing the new scheduler class A mac
156. ster Agent MA_example started Launching an LA For such a platform an LA configuration file could be INRIA ENS Lyon UCBL Page 89 DIET User s Manual BOSE E file LA example cfg configuration file for an LA agentType DIET LOCAL AGENT name LA example parentName MA example traceLevel 1 default dietPort lt port gt not needed dietHostname lt hostname IP gt not needed fastUse 1 ttldapUse 0 default nwsUse 1 nwsNameserver nwshost 9001 useLogService 0 default 1lsOutbuffersize 0 default 1sFlushinterval 10000 default This configuration file is the only argument to the executable dietAgent which is installed in lt install dir gt bin This LA will register as a child of MA example Run gt dietAgent LA_example cfg Local Agent LA_example started Launching a server For such a platform a SeD configuration file could be file SeD_example cfg configuration file for a SeD parentName LA example traceLevel 1 default dietPort lt port gt not needed dietHostname lt hostname IP gt not needed fastUse 1 ldapUse 1 ldapBase Idaphost 9000 1dapMask dc LIP dc ens lyon dc fr nwsUse 1 nwsNameserver nwshost 9001 useLogService 0 default lsOutbuffersize 0 default lsFlushinterval 10000 default The SeD will register as a child of LA example Run the executable that you linked with the DIET SeD library and do not forg
157. stribution to illustrate the usage of the plugin scheduler functionality this code is available in the directory src examples plugin_example A DIET server and client corresponding to a simulation of a database research application are provided If the construction of examples was enabled during DIET configuration two binaries server and client will be built in this directory Having deployed a DIET agent hierarchy the server may be instantiated server lt SeD_config gt lt DB gt lt DB gt where lt DB gt are string s that represent the existence of a particular database at the SeD s site A client would pose a query against a set of databases client lt client_config gt lt DB gt lt DB gt The application uses the plugin scheduling facility to prioritize the existence of databases in selecting a server and thus the expected result is that one of the SeDs with the fewest number of database mismatches will be selected In the main function of the server c file the following block of code a specifies the use of the priority aggregator for this service b declares a performance estimation function to supply the necessary data at request time and c defines the order of precedence of the performance values 7 e minimizing the number of database mismatches and then maximizing the elapsed execution time new section of the profile aggregator diet_aggregator_desc_t agg agg diet_profile
158. t responses const char MyScheduler stName UserGS MyScheduler MyScheduler MyScheduler MyScheduler this gt name this gt stName this gt nameLength strlen this gt name int MyScheduler aggregate corba_response_t aggrResp size_t max_srv const size_t nb_responses const corba_response_t responses SCHEDULER _CLASS MyScheduler After defining the scheduler class the developer just has to use the SCHEDULER_CLASS macro to define it as a scheduler class loadable from an agent In our example the call to SCHEDULER_CLASS MyScheduler after the class declaration makes the class loadable by a DIET agent The aggregation method redefinition The aggregate function has the following prototype int MyScheduler aggregate corba_response_t aggrResp size_t max_srv const size_t nb_responses const corba_response_t responses INRIA ENS Lyon UCBL Page 66 DIET User s Manual BOSE E The aggregate method takes 4 arguments e corba_response_t aggrResp the result of the aggregation has to be set in this argument aggrResp is an array of corba_server_estimation_t objects e size_t max_srv this argument gives the maximum number of responses to return in aggrResp This value can be ignored without any risk and it is sometimes useful to ignore it because this parameter is hard coded in the DIET sources e const size_t nb_responses this argument gives the number of respo
159. t the MA will not accept incoming connection You can find the full set of DIET configuration file options in the chapter A INRIA ENS Lyon UCBL Page 88 DIET User s Manual BOSE E 9 1 3 Example As shown in Section 1 3 the hierarchy is built from top to bottom children register to their parent Here is an example of a complete platform deployment Let us assume that DIET was compiled with FAST on all machines used the LDAP server is launched on the machine ldaphost and listens on the port 9000 the NWS name server is launched on the machine nwshost and listens on the port 9001 the NWS forecaster is launched on the machine nwshost and listens on the port 9002 the NWS sensors are launched on every machine we use Launching the MA For such a platform the MA configuration file could be file MA_example cfg configuration file for an MA agentType DIET_MASTER_AGENT name MA_example traceLevel 1 default dietPort lt port gt not needed dietHostname lt hostname IP gt not needed fastUse 1 ldapUse 0 default nwsUse 1 nwsNameserver nwshost 9001 useLogService 0 default lsOutbuffersize 0 default lsFlushinterval 10000 default This configuration file is the only argument to the executable dietAgent which is installed in lt install_dir gt bin Provided lt install_dir gt bin is in your PATH envi ronment variable run gt dietAgent MA_example cfg Ma
160. t in the zone pointed at by the value field Of course DIET cannot guess how long the user will need these data so the user must free the memory him herself with diet_free_data The fields last_in last_inout and last_out of the diet_profile_t structure respectively point at the indexes in the parameters array of the last IN INOUT and OUT arguments Functions to create and destroy such profiles are defined with the prototypes below diet_profile_t diet_profile_alloc char pb_name int last_in int last_inout int last_out int diet_profile_free diet_profile_t profile The values of last_in last inout and last_out are respectively last_in 1 number of input data last_inout last_in number of inout data last_out last_inout number of out data 3 7 Examples 3 7 1 Example 1 without persistency Let us consider the product of a scalar by a matrix the matrix must be multiplied in place and the computation time must be returned This problem has one IN argument the scalar factor one INOUT argument the matrix and one OUT argument the computation time so its profile will be built as follows INRIA ENS Lyon UCBL Page 31 DIET User s Manual eae E profile last_in 0 last_inout 1 last_out 2 type DIET_SCALAR base_type DIET_DOUBLE type DIET_MATRIX base_type DIET_DOUBLE type DIET_SCALAR base_type DIET_FLOAT value amp factor nbr 5 va
161. t is possible to declare several services for one single SeD INRIA ENS Lyon UCBL Page 42 DIET User s Manual BOSE TE The difference is that the prototype is described with the generic parts of diet_data_desc only whereas the client description uses full diet_data file DIET_data h struct diet_data_generic diet_data_type_t type diet_base_type_t base_type file DIET_server h typedef struct diet_data_generic diet_arg_desc_t typedef struct char path int last_in last_inout last_out diet_arg_desc_t param_desc diet_profile_desc_t diet_profile_desc_t diet_profile_desc_alloc const char path int last_in int last_inout int last_out int diet_profile_desc_free diet_profile_desc_t desc diet_profile_desc_t diet_profile_desc_alloc int last_in int last_inout int last_out int diet_profile_desc_free diet_profile_desc_t desc Each profile can be allocated with diet profile desc_alloc with the same semantics as for diet_profile_alloc Every argument of the profile will then be set with diet_generic_desc_set defined in DIET_server h Declaring services Every service must be added in the service table before the server is launched The complete service table API is defined in DIET_server h typedef int diet_solve_t diet_profile_t int diet_service_table_init int max_size int diet_service_table_add diet_profile_desc_t profile diet_convertor_t cvt diet_solve_t solve_func
162. t structure is defined as follows struct corba_estimation_t typedef _CORBA_ConstrType_Variable_Var lt corba_estimation_t gt _var_type SeqEstValue_t estValues void operator gt gt cdrStream amp const void operator lt lt cdrStream amp SeqEst Value_t est Values This field is a CORBA sequence of estimation values These estima tion values are accessed through the specific functions diet_est_get_internal and diet_est_array_get_internal defined in scheduler est_internal hh These functions prototypes are double diet_est_get_internal estVectorConst_t ev int tag double errVal double diet_est_array_get_internal estVectorConst_t ev int tag int idx double errVal e ev the estimation vector to evaluate e tag the estimation tag e idx the index of the value when available For example to obtain the frequency of the second processor we have to set idx to 1 e errVal the value returned by the function if an error occurred The tag argument may be assigned one of the following values EST_TCOMP The computation time evaluated by FAST FAST must be activated at the compilation time EST_TIMESINCELASTSOLVE The time elapsed since this SeD solved a request This value is used by the default Round Robin scheduler when available EST_COMMPROXIMITY EST_TRANSFEREFFORT EST_FREECPU The free CPU computation power EST_FREEMEM The free memory on the node EST_NBCPU The number of CPU in
163. t use start option 7 gt omniNames Tue Jun 28 15 58 08 2005 Read log file successfully Root context is IOR 010000002b00000049444c3a6f6d67 2e6f7 267 2f436f734e616d696e672f4e61 6d696e67436f 6e746578744578743a312e300000010000000000000060000000010102000d0000003134 302e37372e31332e34390000f90a0b0000004e616d655365727669636500020000000000000008000000 0100000000545441010000001c0000000100000001000100010000000100010509010100010000000901 Checkpointing Phase 1 Prepare Checkpointing Phase 2 Commit Checkpointing completed CORBA usage for DIET Every DIET entity must connect to the CORBA name server it is the way services dis cover each others The reference to the omniORB name server is written in a CORBA INRIA ENS Lyon UCBL Page 84 DIET User s Manual BOSE E configuration file whose path is given to omniORB through the environment variable OMNIORB_CONFIG or with omniORB 4 at compile time through the configure script option with omniORB config An example of such a configuration file is given in the directory src examples cfgs of the DIET source tree and installed in lt install_dir gt etc The lines concerning the name server in the omniORB configuration file are built as fol lows omniORB 3 ORBInitialHost lt name server hostname gt ORBInitialPort lt name server port gt omniORB 4 InitRef NameService corbaname lt name server hostname gt lt name server port gt The name server port is the port
164. t_scalar_desc_set diet_parameter pb 2 amp time Free IN data diet_free_data diet_parameter pb 0 return res int main int argc char argv diet_profile_desc_t profile Initialize table with maximum 1 service diet_service_table_init 1 Define smprod profile profile diet_profile_desc_alloc smprod 0 1 2 diet_generic_desc_set diet_param_desc profile 0 DIET_SCALAR DIET_DOUBLE diet_generic_desc_set diet_param_desc profile 1 DIET_MATRIX DIET_DOUBLE diet_generic_desc_set diet_param_desc profile 2 DIET_SCALAR DIET_FLOAT Add the service the profile descriptor is deep copied diet_service_table_add profile NULL solve_smprod Free the profile descriptor since it was deep copied diet_profile_desc_free profile Launch the SeD no return call diet_SeD SeD cfg argc argv Dead code return 0 INRIA ENS Lyon UCBL Page 44 DIET User s Manual BOSE TE 5 4 Compilation After compiling her his server program the user must link it with the DIET and CORBA libraries This process is very similar to the one described for the client in section 4 4 Please refer to this section for details INRIA ENS Lyon UCBL Page 45 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 46 DIET User s Manual BOSE TE Chapter 6 Batch and parallel submissions 6 1 Introduction Most of resources in a grid are parallel eith
165. tatic on Mac system this option is automatically set to ON as static compilation of binaries seems to be forbidden on these systems e DIET_USE_DART enables DART reporting system refer to http public kitware com Dart which is used for constructing DIET s dashboard see http graal ens lyon fr DietDashboard Note that setting the DIET_USE_DART will force the option BUILD_TESTING to be set INRIA ENS Lyon UCBL Page 20 DIET User s Manual BASE E e Maintainer By default cmake offers four different build modes that one toggles by po sitioning CMAKE_BUILD_TYPE built in variable to Debug Release RelWithDebInfo and MinSizeRel Maintainer is an additional mode which fulfills two basic needs of the task of the maintainer of DIET The first preventive task is to provide code free from any compilation and link warnings The second corresponds to the snafu stage which is to debug the code For reaching those goals the Maintainer build type sets the compilers flags respectively the linker flags with all the possible warning flags activated resp with the additional debug flags 2 2 7 Compiling and installing Summarizing the configuration choices Once the configuration is properly made one can check the choices made by looking the little summary proposed by cmake This summary should look like denotes eluded portions DIET gt cmake XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
166. te with a C MA pie A similar interface appears in the SeD communication process JXTA pipe JNl interface DIET Figure 12 1 DIET architecture 12 2 1 The JXTA components The client Only one component the client is fully written in java Since it communicates only with JXTA components it doesn t need the DIET client library JXTA pipes do not allow all types of data to be sent through The description of the problem and the problem itself have to be packed to be sent through JXTA pipes These messages are unpacked inside the MA pie and SeD piet The behaviour of the JXTA client is e launch a new JXTA peer e get MA y advertisements JXTA messages travelling through the network identifying a JXTA object by sending a JXTA discovery query e extract the reference of the input pipe of the first MA advertisement discovered INRIA ENS Lyon UCBL Page 106 DIET User s Manual BOSE E e create an output pipe to bind the input pipe extracted e create and send the description of the problem via the pipe created and wait for the response of the MA bound including references of SeDs able to solve the problem e Try to create an output pipe to bind the input pipe of one of the SeDs found e Send the packed problem including data needed for the computation to the SeD bound and wait for its response e Extract results of the response received The SeD The role of the SeD is t
167. the site INRIA ENS Lyon UCBL Page 50 DIET User s Manual BOSE E In order for the service to be system independent the SeD API provides some meta variables which can be used in the script e DIET_NAME_FRONTALE frontale name e DIET_USER_NBPROCS number of processors e DIET_BATCH_NODESLIST list of reserved nodes e DIET_BATCH_NBNODES number of reserved nodes e DIET_BATCH_NODESFILE name of the file containing the identity of the reserved nodes e DIET_BATCH_JOBID batch job ID e DIET_BATCHNAME name of the batch system Once the script written in a string it is given as an argument to the following function int diet_submit_parallel diet_profile_t x pb char script 6 8 4 Example of the client server concatenation problem There are fully commented client server examples in lt diet_src gt src examples Batch direc tory The root directory contains a simple example and TestAl1Batch and Cori_cycle_stealing are more practical the latter being a code to explain the CoRI _batch API The root directory contains a simple basic example on how to use the batch API is given here no IN or INOUT args the client receives as a result the number of processors on which the service has been executed The service only writes to a file with batch independent mnemonics some information on the batch system The lt diet_src gt src examples Batch file_transfer directory contains 3 servers one sequential one parallel a
168. tial or parallel in which case its resolution requires numerous proces sors of a parallel resource a parallel machine or a cluster of workstations If parallel the task can be modeled with the MPI standard or composed of multiple sequential tasks deployed for example with ssh resolving a single service it is often the case with data parallelism problems Note that when dealing with batch reservation systems we will likely speak about jobs rather than about tasks http public eu egee org INRIA ENS Lyon UCBL Page 47 DIET User s Manual BOSE TE 6 3 Configuration for compilation You must enable the batch flag in cmake arguments Typically if you build DIET from the command line you can use the following ccmake diet_src_path a DDIET USE ALT BATCH BOOL ON 6 4 Parallel systems Single parallel systems are surely the less deployed in actual computing grids They are usually composed of a frontal node where clients log in and from which they can log on numerous nodes and execute their parallel jobs without any kind of reservation time and space Some problems occur with such a use of parallel resources multiple parallel tasks can share a single processor hence delaying the execution of all applications using it during the deployment the application must at least check the connectivity of the resources if performance is wanted some monitoring has to be performed by the application 6 5 Batch system
169. til the element is connected to a new parent e long removeElement sends an order to a SeD to kill itself The SeD first unsubscribe from its parent before ending itself properly e long removeElement in boolean recursive same as above but for agents The parameter recursive if true also destroys the underlying hierarchy oth erwise only the agent is killed Now what happens if during a request submission an element receives an order to change its parent Actually nothing will change as whenever a request is received a reference to the parent from which the request originates is locally kept So if the parent changes before the request is sent back to the parent as we keep a local reference on the parent the request will be sent back to the correct parent Hence for a short period of time an element can have multiple parents WARNING currently no control is done on whether or not you are cre ating loops in the hierarchy when changing a parent 16 1 3 Example Two examples on how to call those CORBA methods are present in src examples dynamic_hierarchy e connect cc sends orders to change the parent of an element Usage connect lt SED LA gt lt element name gt lt parent name gt e disconnect cc sends orders to disconnect an element from its parent It does not kill the element but only disconnects it from the DIET hierarchy useful when your platform is not heavily loaded and you want to use only part of t
170. ting argument other functions than DIET_CVT_IDENTITY must be applied The result of this function will then be used as the value for the destination argument Corresponding to the DIET datatypes the following functions exist DIET_CVT_IDENTITY Copy the argument DIET_CVT_VECT_SIZE Get the size of a vector DIET_CVT_MAT_NB_ROW Get the number of rows of a matrix DIET_CVT_MAT_NB_COL Get the number of columns of a matrix DIET_CVT_MAT_ORDER Get the order of a matrix DIET_CVT_STR_LEN Get the length of the string DIET_CVT_FILE_SIZE Get the size of the file Only the DIET_CVT_IDENTITY function can be applied to any argument all other functions only operate on one type of argument 8 2 4 Example with convertors A short example is available below Example 1 Assume we declared a profile INOUT MATRIX with the path solve_T This profile will be called by the client Our implementation expects FAST with the path T_solve We will write a convertor that changes the name and extracts the a profile IN INT IN INT INOUT MATRIX This profile is known to matrix s dimensions declare a new convertor with 2 IN 1 INOUT and O OUT arguments INRIA ENS Lyon UCBL Page 78 DIET User s Manual BOSE E cvt diet_convertor_alloc T_solve 0 1 1 apply the function DIET_CVT_MAT_NB_ROW to determine the Oth argument of the converted profile The function s argument is the Oth argument of th
171. to its parent LA it submits a list of the services it offers The agent then reports the new service offering through its parent agent until the MA If the service was previously unavailable along that arm of the hierarchy the agents update their records Finally clients can access the registered service by contacting the MA 5 to get a reference to the best server available and then directly connect to it 6 to launch the computation The architecture of the hierarchy is described in configuration files see Section 9 1 2 and each component transmits the local configuration to its parent Thus the system administration can also be hierarchical For instance an MA can manage a domain like a university providing prioritary access to users of this domain Then each laboratory can run an LA while each team of the laboratory can run some other LAs to administrate its own servers This hierarchical administration of the system allows local changes in the configuration without interfering with the whole platform 1 4 Solving a problem Assuming that the architecture described in Section 1 1 includes several servers able to solve the same problem the algorithm presented below lets an MA select a server for the computation among those available This decision is made in four steps e The MA propagates the client request through its subtrees down to the capable servers actually the agents only forward the request on those subtrees offering the servic
172. tory defaulted to usr local on Un x This is cmake s portable equivalent of the autotools configure s prefix option Eventually here is a short list of ccmake interface tips e when lost look at the bottom lines of the interface which always summarizes ccmake s most pertinent options corresponding keyboard shortcuts depending on your current context e hitting the h key will direct you ccmake embedded tutorial and a list of keyboard shortcuts as mentioned in the bottom lines hit e to exit e up down navigation among parameter items can be achieved with the up down arrows e when on a parameter item the line in inverted colors close above the bottom of the screen contains a short description of the selected parameter as well as the set of possi ble recommended values e toggling of boolean parameters is made with enter e press enter to edit path variables INRIA ENS Lyon UCBL Page 17 DIET User s Manual BOSE TE e when editing a PATH typed parameter the TAB keyboard shortcut provides an emacs like or bash like automatic path completion e toggling of advanced mode press t reveals hidden parameters 2 2 3 A ccmake walk through for the impatients Assume that CVS_DIET_HOME represents a path to the top level directory of DIET sources This DIET sources directories tree can be obtained by DIET users by expanding the DIET current source level distribution tarball But for the DIET develo
173. ts are stored in an LDAP tree The user API of FAST is composed of a small set of functions that combine resource availabilities and routine needs from low level software to produce ready to use values These results can be combined into analytical models by the parallel extension 3 to forecast execution times of parallel routines FAST clients can access information like the time needed to move a given amount of data between two FAST enabled machines SeDs the time to solve a problem with a given set of computational resources or the combination of these two quantities For more details about FAST please refer to the FAST webpage t 8 2 1 Building FAST The first step is to download and install FAST and its dependent programs FAST depends on e NWS the Network Weather Service e GSL the GNU Scientific Library e OpenLDAP an implementation of the Lightweight Directory Access Protocol Ihttp www loria fr quinson fast html INRIA ENS Lyon UCBL Page 76 DIET User s Manual BOSE TE Client Application Knows about Computer A a K PO A i now e CPU speed o FAST O a s about e Available memory gt Parallel Batch System Extensions gt oo Network 2 5 Bandwid oh lt p Fast API 77 gt e Memory load eLatenc ii a eBatch status Topolosy System ae Network Protocol hid Availabilitie
174. uble matrix 5 The matrix to multiply float time NULL To check that time is set by the server diet_profile_t profile 5 diet_reqID_t rst 5 0 0 0 0 0 m 60 n 100 Initialize a DIET session diet_initialize client cfg argc argv INRIA ENS Lyon UCBL Page 37 DIET User s Manual BOSE E Create the profile as explained in Chapter 3 for i 0 i lt 5 i Allocate the matrix m lines n columns matrix i malloc m n sizeof double Fill in the matrix with dummy values who cares 7 for j 0 j lt m n j matrix i j 1 2 j profile i diet_profile_alloc smprod 0 1 2 last_in last_inout last_out Set profile arguments diet_scalar_set diet_parameter profile i 0 amp factor 0 DIET_DOUBLE diet_matrix_set diet_parameter profile i 1 matrix i 0 DIET_DOUBLE m n DIET_COL_MAJOR diet_scalar_set diet_parameter profile i 2 NULL O DIET_FLOAT Call Diet int rst_call 0 for i 0 i lt 5 i if rst_call diet_call_async profile il amp rst i 0 printf Error in diet_call_async return d n rst_call else printf request ID value d n rst i if rstlil lt 0 4 printf error in request value ID n return 1 rst_call 0 Wait for Diet answers if rst_call diet_wait_and diet_reqID_t amp rst unsigned int 5 0 printf Error in diet
175. uffer is specified by it s size lsOutbuffersize number of messages and the time it is regu larly flushed 1sFlushinterval nanoseconds It is recommended not to change the default parameters if you do not encounter problems The buffer options will be ignored if useLogService is set to 0 FAST options Currently FAST is only used at the SeD level so these parameters will only have an effect in SeD configuration files fastUse default 0 This option activates the requests to FAST It is ignored if DIET was compiled without FAST and defaults to 0 otherwise The following options are ignored if DIET was compiled without FAST or if fastUse is set to 0 LDAP options ldapUse default 0 This option activates the use of LDAP in FAST requests Only SeDs need to connect to the LDAP so the option is ignored at the agent level The following two options are ignored if 1dapUse is set to 0 ldapBase default none Specify the host port address of the LDAP base where FAST gets the results of its benchmarks INRIA ENS Lyon UCBL Page 87 DIET User s Manual BOSE E ldapMask default none Specify the mask used for requests to the LDAP base It must match the one given in the 1dif file of the server that was added to the base NWS options nwsUse default 0 This option activates the use of NWS in FAST requests If 0 FAST will use an internal sensor for the performance of the machine but will not be able to evaluate communica
176. ulti MA extension as shown in Figure 11 2 gamma delta p CR a alpha beta Figure 11 1 Example of a multi MA deployment The following lines appear in the MA configuration file of alpha They tell that the multi MA extension should listen for incoming connection at port 2001 They also tell that the MA should create a link toward the MA of the organization gamma and toward the MA of the organization beta The description of each configuration parameter are available in Section 9 1 2 agentType DIET_MASTER_AGENT dietHostname diet alpha com bindServicePort 2001 neighbours diet beta com 2001 ma gamma com 6000 The following lines appear in the MA configuration file of beta agentType DIET_MASTER_AGENT dietHostname diet beta com bindServicePort 2001 neighbours diet alpha com 2001 ma gamma com 6000 INRIA ENS Lyon UCBL Page 102 DIET User s Manual BOSE TE The following lines appear in the MA configuration file of gamma The neighbours value is empty This means that the gamma s MA will not try to connect itself to other MA However the three others are configured to be connected to gamma So after all the gamma MA is connected to the other three agentType DIET_MASTER_AGENT dietHostname ma gamma com bindServicePort 6000 neighbours Finally the following lines appear in the MA configuration file of delta agentType DIET_MASTER_
177. ur OS name and version and the JuxMem version number However please do understand that this is an on going work and therefore no full support is provided INRIA ENS Lyon UCBL Page 114 DIET User s Manual BOSE TE Chapter 14 Workflow management in Diet 14 1 Overview Workflow applications consists of multiple components tasks related by precedence con straints that usually follow from the data flow between them Data files generated by one task are needed to start another task Although this is the most common situation the precedence constraints may follow from other reasons as well and may be arbitrarily defined by the user This kind of application can be modeled as a DAG Directed Acyclic Graph where each vertex is a task with given input data and service name and each edge can either be a data link between two tasks or a basic precedence constraint The DIET workflow engine can handle that kind of workflow by assigning each task to a SeD in the DIET hierarchy using a DIET service call This assignment is made dynamically when the task is ready to be executed i e all predecessors are done depending on the service performance properties and on available ressources on the grid A specific agent called the Master Agent DAG MApac provides DAG workflow scheduling This agent serves as the entry point to the DIET Hierarchy for a client that wants to submit a DAG workflow The language supported by the MApac is based on
178. urator See http shell jxta org index html for config help TCP Senings Figure 12 2 Configuring JXTA INRIA ENS Lyon UCBL Page 112 DIET User s Manual BOSE E Chapter 13 JuxMem extension 13 1 Introduction With the release of version 2 0 of the DIET toolkit we have introduced the ability to use JuxMem for managing persistent data blocks This section shortly describes how to use JuxMem inside DIET as it is an on going work 13 2 Overview of JuxMem JuxMem stands for Juxtaposed Memory implements the concept of data sharing service for grid based on a compromise between DSM systems and P2P systems JuxMem de couples data management from grid computation by providing location transparency as well as data persistence in a dynamic environnement JuxMem is based on the P2P plat form called JXTA which stands for Juxtaposed For more information about JuxMem please check the available documentation on the web site of JuxMem 19 13 3 How to configure DIET to use JuxMem DIET currently needs JuxMem version 0 3 to work This version can be downloaded on the web site of JuxMem 19 For configuring and building JuxMem please check the README file included in this 0 3 release of JuxMem When the DDIET_USE_JUXMEM option is activated you need to have JuxMem C build so please read the documentation for building JuxMem C Currently for configuring DIET in order to use JuxMem you need to specify the build path of Jux
179. use FAST in the plug in scheduler to obtain advanced performance predictions A second possibility to get performance prediction called CoRI is now available The aim of CoRI is to simplify the access to the information Inside of CoRI FAST can be called but it is only one source of information among other sources for example Cori Easy FAST is described in Section 3 2 CoRI is described in Section 8 3 The default compiling is without FAST and without CoRI Note that if you compile with batch enabled then CoRl is also enabled In the table 8 1 you can see which information is available with each compiling option 8 2 FAST Fast Agent s System Timer This section deals with FAST a performance prediction module that can be used by DIET It is non mandatory but can provide SeDs with improved performance prediction capability You can use FAST in stand alone mode without having compiled with CoRI option FAST 21 is a tool for dynamic performance forecasting in a Grid environment As shown in Figure 8 1 FAST is composed of several layers and relies on a variety of low level software First it uses the Network Weather Service NWS 26 a distributed system that periodically monitors and dynamically forecasts the performances of various network and computational resources The resource availabilities acquisition module of FAST uses and enhances NWS Indeed if there is no direct NWS monitoring between two machines FAST automatically searches
180. utElse port to an input port lt condition name IF_Example gt lt in name i type integer gt lt in name j type integer gt lt outThen name outi type integer gt lt outElse name out2 type integer gt lt IF Condition must be written in XQuery language gt lt if gt i 1t j lt if gt lt then gt out1 i lt then gt lt else gt out2 j lt else gt lt condition gt Note that all the operators and functions defined in the XQuery standard see http www w3 org TR xquery operators can be used to make the boolean expression of the lt if gt tag These can process both numerical and string variables and can also contain XPath expressions to access elements of an array when the input port type is an array for ex the expression contains inlistitem 1 text a tests if the 1st element of the array provided by the in port contains the letter a While loops This structure uses specific port tags lt inLoop gt and lt outLoop gt in ad dition to standard port tags They are used to connect this processor to other processors that will be iterated as long as the while condition is true condition is evaluated before the first iteration The standard lt in gt and lt out gt ports are used to connect this processor to the rest of the workflow The loop initialization is done by mapping data from in ports to inLoop ports using the init attribute Each iteration produces data o
181. value 1 gt lt cfg options gt lt SeD gt lt master_agent gt lt diet_hierarchy gt lt diet_configuration gt remote_binary server_dyn_add_rem gt remote binary server dyn add rem gt Figure 9 1 Example XML input file for GODIET INRIA ENS Lyon UCBL Page 94 DIET User s Manual BOSE E Chapter 10 DIET dashboard This section discussed monitoring tools that can be used with DIET We are currently working on a tool called DIET Dashboard that will integrate a variety of external tools to provide a single management and monitoring environment for DIET Currently however each of these tools is available separately See Section 10 1 for a description of LogService Section 10 2 for a description of VizDIET and Section 9 2 for a description of GODIET 10 1 LogService The DIET platform can be monitored using a system called LogService This monitoring service offers the capability to be aware of information that you want to relay from the platform As shown in Figure 10 1 LogService is composed of three modules LogCom ponent LogCentral and LogToo l A LogComponent is attached to a component and relays information and messages to LogCentral LogComponents are typically used within components one wants to monitor LogCentral collects messages received from LogComponents then LogCentral stores or sends these messages to Log Tools LogTools connect themselves to LogCentral
182. vide the required data in the estimation vector by implementing a plugin scheduler see chapter 7 The following fields in the estimation vector must be filled in 1 The TCOMP field must contain the estimation of the computation time for the job in milliseconds This can be done using the diet_estimate_comptime estVector_t ev double value method within the performance evaluation function INRIA ENS Lyon UCBL Page 128 DIET User s Manual BOSE TE 2 The EFT field must contain the estimation of the earliest finish time in milliseconds from the time of the current submit request for the job To compute this value the SeD programmer can use the API method diet_estimate_eft to retrieve the estimated value of earliest finish time for a new job INRIA ENS Lyon UCBL Page 129 DIET User s Manual BOSE E INRIA ENS Lyon UCBL Page 130 DIET User s Manual BOSE TE Chapter 15 DAGDA extension DAGDA Data Arrangement for Grid and Distributed Applications is a new data man ager for DIET DAGDA offers to the DIET application developers a simple and efficient way to manage the data It was not designed to replace the JuxMem extension but to be possibly coupled with it In a future work DAGDA will be divided in two parts The DAGDA data manager and the DAGDA data interface The data interface will make interactions between DAGDA JuxMem FTP etc and other data transfer management protocols In this chapter
183. ware DIET uses BLAS to build demonstration examples of clien t server Note that the option DIET_BUILD_BLAS_EXAMPLES can only be effective when DIET_BUILD_EXAMPLES is enabled DIET_BUILD_BLAS_EXAMPLES is disabled by default e DIET_USE_ALT_BATCH enables the transparent submission to batch servers See Chapter 6 for more details e DIET_USE_CORI CoRI which stands for COllector of Resource Information provides a framework for probing hardware and performance information about the SeD CoRI also yields a very basic set of probing resources which are heavily dependent on the system calls available for the considered platform When this option is activated disabled by default the user can either define new collectors or use existing collectors like FAST see the DIET_USE_FAST option through CoRI s interface CoRI thus provides a possible tactical approach for tuning the performance of your favorite plug in scheduler Chapter 8 describes in more details CoRI and its possible usage within DIET e DIET_USE_DAGDA enables the use of the new data management module of DIET DAGDA offers to the DIET application developers a simple and efficient way to manage the data See Chapter15 for more details e DIET_USE_DYNAMICS enables a mode in which you can dynamically modify its shape using CORBA calls See chapter 16 for more details about the option and how it works e DIET_USE_FAST activates DIET support of FAST refer to http www loria fr quinson
184. way to implement the Multi MA is to use peer to peer technology and thus have a distributed Multi Agent system where MAs dynamically discover each others and coop erate in order to give clients the largest possible area of search in a transparent manner JXTA 20 is a technology written with java 11 It aims at allowing the development of destributed applications using peer to peer concepts and the java language JXTA provides functionalities such as passing firewalls and similar network protections dynam ically discovering other peers and other essential tools to develop a Multi Agent system using peer to peer technology 12 2 Description of the current architecture devel oped with JXTA In this chapter we discuss one prototype We plan to update this prototype that will be totally merged in DIET and able to process all requests supported by DIET The DIET architecture is shown Figure 7 1 We can consider that the elements allowing its use are divided in two parts e a JXTA part that includes chent MA and SeD z These components are written in java to be able to communicate together using JX TA e a part of integration of the JXTA part in DIET java JXTA and C DIET must cooperate The technology used to allow this integration is JNI 12 that INRIA ENS Lyon UCBL Page 105 DIET User s Manual BOSE E allows java to call functions written in C JNI is located in the MA and the SeD The MA has to launch and communica
185. will store all the data it s managing When a SeD has a configured and valid path name to a backup file dataBackupFile option in the configuration file a client can ask to the agents or SeDsDAGDA components to save the data The dagda_save_platform function which can only be called from a client records all the data managed by the agents or SeDs DAGDA components that allow it Then the restoreOnStart configuration file option asks to the DAGDA component to restore the data stored on the dataBackupFile file when the component starts This mechanism allows to stop the DIET platform for a while and restart it conserving the same data distribution 15 4 5 Create data ID aliases For many applications using large sets of data shared by several users to use an auto matically generated ID to retrieve a data is difficult or even impossible DAGDA allows the user to define data aliases using human readable and expressive strings to retrieve a data ID Two functions are defined to do it dagda_data_alias const char id const char alias Tries to associate alias to id If the alias is already defined returns a non zero value A data can have several aliases but an alias is always associated to only one data dagda_id_from_alias const char alias char id This function tries to retrieve the data id associated to the alias INRIA ENS Lyon UCBL Page 138 DIET User s Manual BOSE E 15 4 6
186. xec estVector_t ev const diet_profile_t const profilePtr with an appropriate value for ev and the profilePtr corresponding to the current DIET request e The number of waiting jobs when using the maximum concurrent jobs limit is stored in an estimation metric vector by calling int diet_estimate_waiting_jobs estVector_t ev e CoRI allows to access in an easy way to basic performance prediction See Chapter 8 3 to know more about the use of it In the future we plan to expand the suite of default estimation metrics to include dynamic internal DIET system state information e g queue lengths Developer defined Performance Metrics Application developers may also define performance values to be included in a SeD response to a client request For example a DIET SeD that provides a service to query particular databases INRIA ENS Lyon UCBL Page 55 DIET User s Manual BOSE E may need to include information about which databases are currently resident in its disk cache in order that an appropriate server be identified for each client request To store such values the SeD developer should first choose a unique integer identifier referred to as the tag to denote each logical datum to be stored Values are associated with tags using the following interface int diet_est_set estVector_t ev int userTag double value The ev parameter is the estimation vector where the value will be stored the userTag parameter denotes t
187. xtension of a Dynamic Performance Forecasting Tool In Proceedings of the International Symposium on Parallel and Distributed Computing lasi Romania July 2002 4 Eddy Caron and Holly Dail Godiet a tool for managing distributed hierarchies of diet agents and servers Research report 2005 06 Laboratoire de l Informatique du Paralllisme LIP February 2005 Also available as INRIA Research Report RR 5520 5 Condor G http www cs wisc edu condor condorg 6 M C Ferris M P Mesnier and J J Mori NEOS and Condor Solving Optimization Problems Over the Internet ACM Transaction on Mathematical Sofware 26 1 1 18 2000 http www unix mcs anl gov metaneos publications index html 7 C Germain G Fedak V N ri and F Cappello Global computing systems Lecture Notes in Computer Science 2179 218 227 2001 8 Globus http www globus org 9 Sun GridEngine http wwws sun com software gridware 10 GridRPC Working Group https forge gridforum org projects gridrpc wg 11 Java http java sun com INRIA ENS Lyon UCBL Page 157 DIET User s Manual BOSE E 12 13 14 15 16 17 18 19 20 21 22 23 24 25 JNI http java sun com j2se 1 3 docs guide jni S Matsuoka H Nakada M Sato and S Sekiguchi Design Issues of Network Enabled Server Systems for the Grid http www eece unm edu dbader grid WhitePapers satoshi pdf 2000 G
188. y modify its shape using CORBA calls For this you need to compile DIET with the option DIET_USE_DYNAMICS In this mode if a DIET element cannot reach its parent when initializing it won t exit but will wait for an order to connect itself to a new parent Hence you do not need to deploy DIET starting from the MA down to the SeD you can launch all the elements at once and then send the orders for each element to connect to its correct parent you do not even need to follow the shape of the tree you can start from the bottom to the tree up to the root or use a random order the service tables will be correctly initialized You now have access to the following CORBA methods e long bindParent in string parentName sends an order to a SeD or agent to bind to a new parent having the name parentName if this parent can be contacted otherwise the element keeps its old parent If the element already has a parent it unsubscribes itself from the parent so that this latter is able to update its service INRIA ENS Lyon UCBL Page 141 DIET User s Manual BOSE E table and list of children A null value is returned if the change occurred otherwise a value different from 0 is returned if a problem occurred e long disconnect sends an order to disconnect an element from its parent This does not kill the element but merely removes the link between the element and its parent Thus the underlying hierarchy will be unreachable un
189. zation has been written only for these cases One of the first things to do is to write a generic packing and unpacking to be able to process all problems supported by DIET e The client isn t very simple to write because nothing is hidden to the user neither the details of the JXTA communication nor the creation of the problem As for the client pie an API providing all mechanisms needed to communicate with DIET via JXTA pipes should be written The implementation of a Java Client taking in account the JXTA communication seems to be the solution 12 4 Working with a DIET platform 12 4 1 Installation and configuration e You need a JDK1 4 1 or later release from for instance http java sun com javase downloads index jsp Previous JDKs and other java compiler are known to generate errors Ensure that environment variable PATH contains the javac and javah binaries location INRIA ENS Lyon UCBL Page 109 DIET User s Manual BOSE E e Then to configure DIET with the JXTA option switch the DIET_USE_JXTA option to ON inside the ccmake GUI The JXTA client example is compiled if DIET BUILD EXAMPLE is also switched to ON 12 4 2 Deploying a DIET platform Please refer to the previous chapter for more information concerning things to do before deploying the platform e First step launching a MAJ After having set the LD LIBRARY PATH OMNIORB_CONFIG and OMNINAMES_LOGDIR paths DIET is ready to run except the JXTA part

Download Pdf Manuals

Related Search

Related Contents

PARTIAL STAINLESS STEEL GAS BBQ GRILL USER`S MANUAL  USER MANUAL FOR LnT Stores  Sony VAIO SVF1521GCXB notebook  PC接続によるファームウェアアップデート方法  Chief KSA1002B      retention.link user manual  World Marketing of America ELCG240 fireplace  NEU - Eppendorf  

Copyright © All rights reserved. Failed to retrieve file