Home

User Manual - Support

Contents

1. A larm the suspend threshold is currently exceeded e C alendar suspended see calendar conf s uspended see qmod S ubordinate d isabled see qmod e D isabled see calendar conf E rror sge execd was unable to locate the sge_shepherd use qmod to fix it o rphaned for queue instances By default the qstat command shows only jobs belonging to the cur rent user i e the command is executed with the option u user To see jobs from other users too the following format is used qstat u Bright Computing Inc 42 SGE 6 4 Deleting A Job A job can be deleted in SGE with the following command qdel jobid The job id is the number assigned by SGE when the job is submitted using qsub Only jobs belonging to the logged in user can be deleted Us ing qdel will delete a user s job regardless of whether the job is running or in the queue Bright Computing Inc PBS Variants Torque And PBS Bright Cluster Manager works with Torque and PBS Pro which are two forks of Portable Batch System PBS PBS was a workload management and job scheduling system first developed to manage computing re sources at NASA in the 1990s Torque and PBS Pro can differ significantly in the output they present when using their GUI visual tools However because of their historical legacy their basic design structure and job submission methods from the command line remain very similar for the us
2. Bright Cluster Manager 7 0 User Manual Revision 6158 Date Thu 23 Apr 2015 2015 Bright Computing Inc All Rights Reserved This manual or parts thereof may not be reproduced in any form unless permitted by contract or by written permission of Bright Computing Inc Trademarks Linux is a registered trademark of Linus Torvalds PathScale is a regis tered trademark of Cray Inc Red Hat and all Red Hat based trademarks are trademarks or registered trademarks of Red Hat Inc SUSE is a reg istered trademark of Novell Inc PGI is a registered trademark of The Portland Group Compiler Technology STMicroelectronics Inc SGE is a trademark of Sun Microsystems Inc FLEXIm is a registered trademark of Globetrotter Software Inc Maui Cluster Scheduler is a trademark of Adaptive Computing Inc ScaleMP is a registered trademark of ScaleMP Inc All other trademarks are the property of their respective owners Rights and Restrictions All statements specifications recommendations and technical informa tion contained herein are current or planned as of the date of publication of this document They are reliable as of the time of this writing and are presented without warranty of any kind expressed or implied Bright Computing Inc shall not be liable for technical or editorial errors or omissions which may occur in this document Bright Computing Inc shall not be liable for any damages resulting from the use of this docu ment
3. 9 3 2 Using MIC Cards With PBS Pro 9 3 3 Using MIC Cards With TORQUE 9 3 4 Using MIC Cards WithSGE 9 3 5 Using MIC Cards With openlava User Portal 19 1 Home Pagers 2 25 ur er ann Fan ee aan 10 2 The WORKLOAD Tab ee 10 3 The NODES Tab oriire aa na oE ee 10 4 The GRAPHS Tabi o anne aaa a e a ee MPI Examples A 1 A 2 A 3 AA AS A 6 A 7 HELLO world en N Be densus MPI Skeleton urn tr ure gene MPI Initialization And Finalization 2 2 What Is The Current Process How Many Processes Are Th re 2 EX hte ege metes ns SendingMessages ee ee Receiving Messages lll Blocking Non Blocking And Persistent Messages A 7 1 Blocking Messages llle A 7 2 Non Blocking Messages 2 2 2 2 2 22 20 A 7 3 Persistent Non Blocking Messages B Compiler Flag Equivalence 59 60 61 61 63 65 65 65 66 67 68 68 69 69 70 70 71 71 72 72 73 77 77 78 80 80 80 81 81 81 82 82 83 Preface Welcome to the User Manual for the Bright Cluster Manager 7 0 0 1 About This Manual This manual is intended for the end users of a cluster running Bright Clus ter Manager and tends to see things from a user perspective It covers the basics of using the Bright Cluster Manager user environment to run com pute jobs on the cluster Although it does cover some aspects of general Linux usage i
4. PI Send amp work 1 MPI INT status MPI SOURCE WORKTAG MPI COMM WORLD work x get next work request x Receive results for outstanding work requests x for rank 1 rank ntasks rank MPI_Recv amp result 1 MPI DOUBLE MPI ANY SOURCE MPI ANY TAG MPI COMM WORLD amp status Tell all the compute nodes to exit x for rank 1 rank ntasks rank MPI Send 0 0 MPI INT rank DIETAG MPI COMM WORLD computenode double result int work MPI_Status status for 4 MPI Recv amp work 1 MPI INT 0 MPI ANY TAG Bright Computing Inc 80 MPI Examples MPI COMM WORLD amp status Check the tag of the received message if status MPI_TAG DIETAG return result do the work x MPI_Send amp result 1 MPI DOUBLE 0 0 MPI COMM WORLD Processes are represented by a unique rank integer and ranks are numbered 0 1 2 N 1 MPI COMM WORLD means all the processes in the MPI application It is called a communicator and it provides all information necessary to do message passing Portable libraries do more with communicators to provide synchronisation protection that most other systems cannot handle A 3 MPI Initialization And Finalization As with other systems two functions are provided to initialize and clean up an MPI process MPI Init amp argc amp argv MPI Finalize A 4 What Is The Current Proce
5. export OMP_NUM_THRI EADS 3 The number of threads specified by the variable can then be run over the hosts specified by the mpi run hosts file fred bright70 mpirun np 2 hostfile mpirun hosts World World World World World World HHH HHH am am am am am am thread thread thread thread thread thread Hello I am Processor 0 on node001 Hello I am Processor 1 on node002 Hybrid Hello Hybrid Hello Hybrid Hello Hybrid Hello Hybrid Hello Hybrid Hello dE db db db ou of 2 of 2 out out out out out HR NO HN out Benefits And Drawbacks Of Using OpenMP The main benefit to using OpenMP is that it can decrease memory re quirements with usually no reduction in performance Other benefits include of of of of of of C CO Q0 O0 1 W WwW hybridhello Potential additional parallelization opportunities besides those ex ploited by MPI Less domain decomposition which can help with load balancing as well as allowing for larger messages and fewer tasks participating in MPI collective operations Bright Computing Inc 18 Using MPI OpenMP is a standard so any modifications introduced into an ap plication are portable and appear as comments on systems not using OpenMP By adding annotations to existing code and using a compiler op tion it is possible to add OpenMP to a code somewhat incremen tally almost on a loop by loop basis The vector loo
6. include lt mpi h gt define WORKTAG define DIETAG main argc argv int argc 1 2 char xargv int MPI Init amp argc MPI Comm rank MPI COMM WORLD amp myrank if myrank amp argv initialize MPI x always use this x process rank 0 thru N 1 myrank 0 head else computenode MPI Finalize cleanup MPI x head int ntasks rank work double result MPI_Status status MPI Comm size Bright Computing Inc A 2 MPI Skeleton 79 MPI COMM WORLD always use this amp ntasks d processes in application x Seed the compute nodes for rank 1 rank lt ntasks rank work x get next work request x MPI Send amp work x message buffer x 15 one data item MPI INT data item is an integer rank destination process rank x WORKTAG user chosen message tag x MPI COMM WORLD always use this Receive a result from any compute node and dispatch a new work request work requests have been exhausted x work x get next work request x while valid new work request PI Recv amp result message buffer x 17 one data item PI_DOUBLE of type double real x PI ANY SOURCE x receive from any sender x PI ANY TAG any type of message x PI COMM WORLD always use this amp status received message info x
7. processor name s n id np processor_name di MPI_Finalize return 0 fred bright70 mpicc hello c o hello fred bright70 hello Hello world from process 000 out of 001 processor name bright70 cm cluster fred bright70 module add openmpi gcc dor as appropriate However it still runs on a single processor unless it is submitted to the system in a special way Running An MPI Executable In Parallel Without A Workload Manager Compute node environment provided by user s bashrc After the relevant module files are chosen section 3 3 for MPI an executable com piled with MPI libraries runs on nodes in parallel when submitted with mpirun The executable running on other nodes loads environmental modules on those other nodes by sourcing the bashrc file of the user section 2 3 3 It is therefore important to ensure that the environmental module stack used on the compute node is clean and consistent Example Supposing the bashrc loads two MPI stacks the mpich stack fol lowed by the Open MPI stack then that can cause errors because the compute node may use parts of the wrong MPI implementation The environment of the user from the interactive shell prompt is not normally carried over automatically to the compute nodes during an mpirun submission That is compiling and running the executable will normally work only on the local node without a special treatment To have the executable run on the compu
8. t if i have a right neighbor PI Send init outbuf count dtype right tag comm amp req count while looping MPI_Startall count req do some work PI Waitall count reg amp statuses do rest of work In the example MPI Send init and MPI Recv init perform a persistent communication initialization Bright Computing Inc Compiler Flag Equivalence The following table is an overview of some of the compiler flags that are equivalent or almost equivalent Bright Computing Inc j dir name o 9 v 3 PGI Pathscale Cray Intel GCC Explanation o fast 03 default default 03 ffast math Produce high level of optimization E zs Activa Pd and m m mp nonuma mp Oomp default openmp fopenmp j a ul diregkives a in the code a Read and write Fortra forma byteswapio byteswapio h convert fconvert swap S Aue ortr n unformatted 9 data files as big endian Q byteswapio big endian Process Fortran source using fixed form Mfixed fixedform f fixed fixed ffixed form Van specifications Mfree freeform f free free ffree form Process Poniran source Using Tee form specifications V dumpversion V version version Dump version N A zerouv h zero N A finit local zero Zero fill all uninitialized variables HEU Creates mod files to hold Fortran90 mod ule information for future compiles Specifies the directory dir name to whi
9. 3 1 The following output will show in the output file Used for debugging echo Got SNSLOTS processors echo Machines cat machines Use MPIRUN to run the application mpirun machinefile machines application The number of available slots can be set by the administrator to an arbitrary value However it is typically set so that it matches the number of cores in the cluster for efficiency reasons More slots can be set than cores but most administrators prefer not to do that In a job script the user can request slots from the available slots Re questing multiple slots therefore typically means requesting multiple cores In the case of an environment that is not set up as a parallel environment the request for slots is done with the np option For jobs that run in a parallel environment the pe option is used Mixed jobs running in both non MPI and parallel environments are also possible if the administrator has allowed it in the complex attribute slots settings Whether the request is granted is decided by the workload manager policy set by the administrator If the request exceeds the number of avail able slots then the request is not granted If the administrator has configured the cluster to use cloud comput ing with cm scale cluster section 7 9 2 of the Administrator Manual then the total number of slots available to a cluster changes over time automatically as nodes are started up and stopped dynamically
10. 6 2 Submitting A Job The SGE module must be loaded first so that SGE commands can be ac cessed module add shared sge Bright Computing Inc 40 SGE With SGE a job can be submitted with qsub The qsub command has the following syntax qsub options jobscript jobscript args After completion either successful or not output is put in the user s current directory appended with the job number which is assigned by SGE By default there is an error and an output file myapp e JOBID myapp o JOBID 6 2 1 Submitting To A Specific Queue Some clusters have specific queues for jobs which run are configured to house a certain type of job long and short duration jobs high resource jobs or a queue for a specific type of node To see which queues are available on the cluster the qst at command can be used qstat q c CLUSTER QUEUF COLOAD USED RES AVAIL TOTAL aoACDS cdsuE long q 0 01 0 0 144 288 0 144 default q 0 01 0 0 144 288 0 144 The job is then submitted for example to the long q queue qsub q long q sleeper sh 6 3 Monitoring A Job The job status can be viewed with qstat In this example the sleeper sh script has been submitted Using qstat without options will only dis play a list of jobs with no queue status options qstat job ID prior name user state submit start at queue slots 249 0 00000 Sleeperl root qw 12 03 2008 07 29 00 I 250 0 00
11. Limitation of Liability and Damages Pertaining to Bright Computing Inc The Bright Cluster Manager product principally consists of free software that is licensed by the Linux authors free of charge Bright Computing Inc shall have no liability nor will Bright Computing Inc provide any warranty for the Bright Cluster Manager to the extent that is permitted by law Unless confirmed in writing the Linux authors and or third par ties provide the program as is without any warranty either expressed or implied including but not limited to marketability or suitability for a specific purpose The user of the Bright Cluster Manager product shall accept the full risk for the quality or performance of the product Should the product malfunction the costs for repair service or correction will be borne by the user of the Bright Cluster Manager product No copyright owner or third party who has modified or distributed the program as permitted in this license shall be held liable for damages including gen eral or specific damages damages caused by side effects or consequential damages resulting from the use of the program or the un usability of the program including but not limited to loss of data incorrect processing of data losses that must be borne by you or others or the inability of the program to work together with any other program even if a copyright owner or third party had been advised about the possibility of such dam ages unless
12. Offload IC 0 CPU Time 0 000000 seconds Offload IC 0 CPU gt MIC Data 64 bytes Offload IC 0 MIC Time 0 000134 seconds Offload IC D MIC gt CPU Data 64 bytes My hostname is node001 micO0 user bright70 More information on building applications in offload mode can be found at http software intel com en us articles the heterogeneous programming model 9 3 Using MIC With Workload Managers When a MIC is configured as a regular node the user can start a native application inside the MIC This can be done by logging into it directly us ing ssh or using a workload manager The workload manager schedules jobs by determining the order in which the jobs will use the MIC cards This is the recommended way to use MIC cards on a multiuser cluster but currently only Slurm supports both the native and offload modes All other workload managers support only offload mode by using a precon figured generic resource 9 3 1 Using MIC Cards With Slurm Offload mode The user creates a job script and specifies a consumable resource mi c For example the following job script runs the dgemm test from the MIC binaries on a host where at least one MIC is available and free bin sh SBATCH partition defq SBATCH gres mic l module load intel cluster runtime intel64 opt intel mic perf bin intel64 dgemm cpu x i 2 1 2048 The job is submitted as usual using the sbatch or salloc srun commands of
13. Possible job states are Job States Description C E 3 H U wo iu Job is completed regardless of success or failure Job is exiting after having run Job is held job is queued eligible to run or routed job is running job is suspend job is being moved to new location job is waiting for its execution time The command qstat q shows what queues are available In the fol lowing example there is one job running in the testq queue and 4 are queued qstat q server master cm cluster Queue Memory CPU Time Walltime Node Run Que Lm State testq 23 59 59 m 1 4 E default 23 59 59 0 0 E 1 4 showq From Maui If the Maui scheduler is running and the Maui module loaded module add maui then Maui s showq command displays a similar output In this example one dual core node is available 1 node 2 processors one job is running and 3 are queued in the 1d1e state showq ACTIVE JOBS JOBNAME USERNAME STATE PROC REMAINING STARTTIME 45 cvsupport Running 2 1 59 57 Tue Jul 14 12 46 20 1 Active Job 2 f 2 Processors Active 100 005 1 of 1 Nodes Active 100 005 IDLE JOBS JOBNAME USERNAME STATE PROC WCLIMIT QUEUETIME 46 cvsupport Idle 2 2 00 00 Tue Jul 14 12 46 20 Bright Computing Inc 7 2 Submitting A Job 53 47 cvsupport Idle 2 2 00 00 Tue Jul 14 12 46 21 48 cvsupp
14. example for CentOS 6 x Example Bright Computing Inc 10 Cluster Usage fred bright70 which gcc gcc version head 1 usr bin gcc gcc GCC 4 4 7 20120313 Red Hat 4 4 7 3 2 Thelatest version suitable for general use that is packaged as a mod ule by Bright Computing Example fred bright70 module load gcc fred bright70 which gcc gcc version head 1 cm shared apps gcc 4 8 1 bin gcc gcc GCC 4 8 1 To use the latest version of GCC the gcc module must be loaded To revert to the version of GCC that comes natively with the Linux distribu tion the gcc module must be unloaded The compilers in the preceding table are ordinarily used for applica tions that run on a single node However the applications used may fork thread and run across as many nodes and processors as they can access if the application is designed that way The standard structured way of running applications in parallel is to use the MPI based libraries which link to the underlying compilers in the preceding table The underlying compilers are automatically made avail able after choosing the parallel environment MPICH MVAPICH Open MPI etc via the following compiler commands Language C C Fortran77 Fortran90 Fortran95 Command mpicc mpiCC mpif77 mpif90 mpif95 2 4 4 Open MPI And Mixing Compilers Bright Cluster Manager comes with multiple Open MPI packages corre sponding to the different a
15. examples workload slurm jobscripts 9 3 2 Using MIC Cards With PBS Pro PBS Pro version 12 0 and higher allows a special virtual node to be cre ated which represents a MIC coprocessor This virtual node is a con sumable resource which can be allocated and released like any other re sources for example like a GPU A MIC coprocessor is represented as a virtual node altough users cannot actually start a native job via PBS Pro All nodes that show the property resources available mic id as an output to the command pbsnodes av are MICs Users can request a number of MICs for offload jobs as follows user bright70 qsub 1 select 1 ncpus 2 mic_id 1 ncpus 0 mic offload job At the time of writing PBS Pro version 12 1 is not pinning tasks to a specific MIC device That is the oFFLOAD DEVICES environment variable is not set for a job 9 3 3 Using MIC Cards With TORQUE TORQUE version 4 2 and higher detects MIC cards automatically and sets the OFFLOAD DEVICES environment variable for a job To find out how many MIC cards are detected by TORQUE and to find out their detected properties the pbsnodes command can be run For example user bright70 pbsnodes node001 grep mic mics 1 mic_status mic 0 mic_id 8796 num_cores 61 num_threads 244 phys Bright Computing Inc 70 Using MICs mem 8071106560 free physmem 7796191232 swap 0 free swap 0 max frequency y 1181 isa COI_ISA_KNC loa
16. for loading some output elided Bright Computing Inc Cluster Usage Example fred bright70 module avail cm local modulefiles cluster tools 7 0dot module info use own cmd freeipmi 1 2 6 null version cmsh ipmitool 1 8 12 openldap cmsub 7 0 module git shared cm shared modulefiles acml gcc 64 5 3 1 hwloc 1 7 acml gcc fma4 5 3 1 intel cluster checker 2 0 acml gcc mp 64 5 3 1 intel cluster runtime ia32 3 5 acml gcc mp fma4 5 3 1 intel cluster runtime intel64 3 5 acml gcc int64 64 5 3 1 intel cluster runtime mic 3 5 acml gcc int64 fma4 5 3 1 intel tbb oss ia32 41 201303140ss In the list there are two kinds of modules local modules which are specific to the node or head node only shared modules which are made available from a shared storage and which only become available for loading after the shared mod ule is loaded The shared module is obviously a useful local module and is there fore usually configured to be loaded for the user by default Although version numbers are shown in the module avail out put it is not necessary to specify version numbers unless multiple ver sions are available for a module To remove one or more modules the module unload or module rm command is used To remove all modules from the user s environment the module purge command is used The user should be aware that some loaded modules can conflict with others loaded at t
17. in the CUDA GDB NVIDIA CUDA Debugger man ual in the doc directory It is based on GDB the GNU Project debugger and requires the use of the g or G options compiling Example nvcc g G testcode cu o testcod Bright Computing Inc 60 Using GPUs 8 5 2 nvidia smi The NVIDIA System Management Interface command nvidia smi can be used to allow exclusive access to the GPU This means only one appli cation can run on a GPU By default a GPU will allow multiple running applications Syntax nvidia smi OPTION1 ARG1 OPTION2 ARG2 The steps for making a GPU exclusive List GPUs Select a GPU Lock GPU to a compute mode After use release the GPU After setting the compute rule on the GPU the first application which executes on the GPU will block out all others attempting to run This application does not necessarily have to be the one started by the user that set the exclusivity lock on the GPU To list the GPUs the L argument can be used nvidia smi L GPU 0 05E710DE 068F10DE Tesla T10 Processor S N 706539258209 GPU 1 05E710DE 068F10DE Tesla T10 Processor S N 2486719292433 To set the ruleset on the GPU nvidia smi i 0 c 1 The ruleset may be one of the following 0 Default mode multiple applications allowed on the GPU 1 Exclusive thread mode only one compute context is allowed to run on the GPU usable from one thread at a time 2 Proh
18. introduction because of its interactive behavior Ina default Bright Cluster Manager configuration Slurm auto detects the cores available and by default spreads the tasks across the cores that are part of the allocation request To change how Slurm spreads the executable across nodes is typically determined by the options in the following table Short Long Option Option Description N nodes Request this many nodes on the cluster Use all cores on each node by default n ntasks Request this many tasks on the cluster Defaults to 1 task per node c cpus per task request this many CPUS per task not implemented by Open MPI yet none ntasks per node request this number of tasks per node The full options list and syntax for salloc can be viewed with man salloc The requirement of specified options to salloc must be met before the executable is allowed to run So for example if nodes 4 and the cluster only has 3 nodes then the executable does not run 5 2 4 Node Allocation Examples The following session illustrates and explains some node allocation op tions and issues for Slurm using a cluster with just 1 compute node and 4 CPU cores Bright Computing Inc 5 2 Running The Executable With salloc Default settings The hello MPI executable with default settings of Slurm runs successfully over the first and in this case the only node that it finds fred bright52
19. job ID is printed to the terminal when the job is submitted To get the job ID of a job if it has been forgotten the following can be used qstat Or showq 7 2 6 Monitoring Nodes In Torque And PBS Pro The nodes that the workload manager knows about can be viewed using the pbsnodes command The following output is from a cluster made up of 2 core nodes as indicated by the value of 2 for ncpu for Torque and PBS Pro If the node is available to run scripts then its state is free or time shared When a node is used exclusively section 8 5 2 by one script the state is job exclusive For Torque the display resembles some output elided Bright Computing Inc 7 2 Submitting A Job 55 fred bright52 pbsnodes a node001 cm cluster state free np 3 ntype cluster status rectime 1317911358 varattr jobs 96 opus 1 node002 cm cluster state free For PBS Pro the display resembles some output elided fred bright52 pbsnodes a node001 cm cluster Mom node001 cm cluster ntype PBS state free pcpus 3 resources_available arch linux resources available host node001 sharing default shared node002 cm cluster Mom node002 cm cluster ntype PBS state free Bright Computing Inc ncpus 2 Using GPUs GPUs Graphics Processing Units are chips that provide specialized par allel processing power Originally GPUs were designed to handle graph ics process
20. nodes to improve the reliability of the cluster In most clusters the login node is also the head node from where the cluster is monitored and installed On the login node applications can be developed code can be compiled and debugged applications can be submitted to the cluster for execution e running applications can be monitored Bap To carry out an ssh login to the cluster a terminal session can be started from Unix like operating systems Example ssh myname cluster hostname On a Windows operating system an 55H client such as for PuTTY http www putty org can be downloaded Another standard possibility is to run a Unix like environment such as Cygwin http www cywin com within the Windows operating system and then run the SSH client from within it A Mac OS X user can use the Terminal application from the Finder or under Application Utilities Terminal app X11 must be in stalled from the Mac OS X medium or alternatively XQuartz can be used instead XQuartz is an alternative to the official X11 package and is usu ally more up to date and less buggy When using the SSH connection the cluster s address must be added When the connection is made a username and password must be entered at the prompt If the administrator has changed the default SSH port from 22 to some thing else the port can be specified with the p port option ssh X p port lt user gt lt cluster gt Bright Compu
21. same as the salloc options of section 5 2 The various settings and variables that go with the application can also be adjusted 5 3 1 Slurm Job Script Structure A job script submission for the Slurm batch job script format is illustrated by the following fred bright52 cat slurmhello sh bin sh SBATCH o my stdout SBATCH time 30 time limit to batch job SBATCH ntasks 1 SBATCH ntasks per node 4 module add shared openmpi gcc 64 1 4 2 slurm mpirun hello Bright Computing Inc 5 3 Running The Executable As A Slurm Job Script 31 The structure is shebangline shell definition line SBATCHlines optional job script directives section 5 3 2 shell commands optional shell commands such as loading necessary modules application execution line execution of the MPI application using sbatch the Slurm submission wrapper In SBATCH lines SBATCH is used to submit options The various meanings of lines starting with are Line Starts With Treated As Comment in shell and Slurm SBATCH Comment in shell option in Slurm SBATCH Comment in shell and Slurm After the Slurm job script is run with the sbatch command Sec tion 5 3 4 the output goes into file my stdout as specified by the o command If the output file is not specified then the file takes a name of the form slurm jobnumber gt out where lt jobnumber gt is a number starti
22. system also known as a queueing system job scheduler or batch submission system manages the available resources such as CPUs GPUs and memory for jobs submitted to the system by users Jobs are submitted by the users using job scripts Job scripts are con structed by users and include requests for resources How resources are allocated depends upon policies that the system administrator sets up for the workload manager 4 2 Why Use A Workload Manager Workload managers are used so that users do not manually have to keep track of node usage in a cluster in order to plan efficient and fair use of cluster resources Users may still perhaps run jobs on the compute nodes outside of the workload manager if that is administratively permitted However running jobs outside a workload manager tends to eventually lead to an abuse of the cluster resources as more people use the cluster and thus inefficient use of available resources It is therefore usually forbidden as a policy by the system administrator on production clusters 4 3 How Does A Workload Manager Function A workload manager uses policies to ensure that the resources of a cluster are used efficiently and must therefore track cluster resources and jobs A workload manager is therefore generally able to Monitor the node status up down load average all available resources available cores memory on the nodes the jobs state queued on hold deleted done M
23. the documentation for the cluster contact information This typically shows how to contact technical support anoverview of the cluster state displaying some cluster parameters a workload overview This is a table displaying a summary of queues and their associated jobs 10 2 The WORKLOAD Tab The page opened by clicking on the WORKLOAD tab allows a user to see workload related information for the cluster figure 10 2 Bright Cluster Manager User Portal Chromium eoo 275 Bright Cluster Manag Q fk amp bieps 10 1 24 230 work ph wa For quick access place your bookmarks here on the bookmarks bar Import bookmarks now 3 other Bookmarks Bright Computing HOME WORKLOAD NODES GRAPHS Bright Cluster Manager User Portal OVERVIEW Queue Scheduler Slots Nodes Running Queued Failed Completed Avg duration Est delay JOBS RUNNING JobID Scheduler Queue Jobname Processes Username Status Run time Figure 10 2 User Portal Workload Page The following two tables are displayed A workload overview table the same as the table in the home page e A table displaying the current jobs running on the cluster 10 3 The NODES Tab The page opened by clicking on the NODES tab shows a list of nodes on the cluster figure 10 3 Bright Computing Inc 10 4 The GRAPHS Tab 73 Bright Cluster Manager User Portal Chromium ooo 2 5 Bright Cluster Manag e Q dfi amp biepsy1
24. tracks processes that the workload manager launches When it sees processes that the workload manager is unable to clean up after the job is over it carries out the cleanup instead Using cm launcher is recommended if jobs that do not get cleaned up correctly are an issue for the user or administrator 7 1 4 Example Batch Submission Scripts Node Availability The following job script tests which out of 4 nodes requested with 1 nodes are made available to the job in the workload manager Example bin bash PBS 1 walltime 1 00 PBS 1 nodes 4 echo n I am on hostnane echo finding ssh accessible nodes for node in cat S PBS_NODEFILE do echo n running on usr bin ssh node hostname done The directive specifying walltime means the script runs at most for 1 minute The PBS_NODEFILE array used by the script is created and appended with hosts by the queueing system The script illustrates how the workload manager generates a PBS_NODEFILE array based on the requested number of nodes and which can be used in a job script to spawn child processes When the script is submitted the output from the log will look like I am on node001 finding ssh accessible nodes running on node001 running on node002 running on node003 running on node004 This illustrates that the job starts up on a node and that no more than the number of nodes that were asked for in the res
25. up the environment easier Bright Cluster Manager provides preconfigured environment modules section 2 3 2 3 Environment Modules It can be quite hard to set up the correct environment to use a particular software package and version For instance managing several MPI software packages on the same system or even different versions of the same MPI software package is quite difficult for most users on a standard SUSE or Red Hat system be cause many software packages use the same names for executables and libraries A user could end up with the problem of never being quite sure which libraries have been used for the compilation of a program as multiple li braries with the same name may be installed Very often a user would like to test new versions of a software package before permanently installing the package Within a Red Hat or SuSE setup without special utilities this would be quite a complex task to achieve Environment modules using the module command are a special utility to make this task much easier Bright Computing Inc 2 3 Environment Modules 2 3 1 Available commands module help Modules Release 3 2 10 2012 12 21 Copyright GNU GPL v2 1991 Usage module switches subcommand subcommand args Switches H help this usage info V version modules version amp configuration options f force force active dependency resolution t terse terse format avail and list format 1 l
26. value of the letter in b egin the parentheses e nd n do not send email Queue is destination PBS q destination Login shell path is lt shellpath gt PBS S lt shellpath gt Bright Computing Inc 7 1 Components Of A Job Script Resource List Directives Examples Examples of how requests for resource list directives work are shown in the following table Resource Example Description PBS 1 Specification Request 500MB memory mem 500mb Set a maximum runtime of 3 walltime 03 10 30 hours 10 minutes and 30 seconds 8 nodes anywhere on the cluster nodes 8 8 nodes anywhere on the cluster select 8 2 nodes 1 processor per node nodes 2 ppn 1 3 nodes 8 processors per node nodes 3 ppn 8 5 nodes 2 processors per node nodes 5 ppn 2 gpus 1 and 1 GPU per node 5 nodes 2 processors per node select 5 ncpus 2 ngpus 1 and 1 GPU per node 5 nodes 2 processors per node 3 select 5 ncpus 2 mpiprocs 3 virtual processors for MPI code 5 nodes 2 processors per node select 5 ncpus 2 ngpus 1 using any GPU on the nodes 5 nodes 2 processors per node select 5 ncpus 2 gpu id 0 using a GPU with ID 0 from nodes For Torque 2 5 5 For PBS Pro 11 Some of the examples illustrate requests for GPU resource usage GPUs and the CUDA utilities for NVIDIA are introduced in Chapter 8 In the Torque and PBS Pro workload managers GPU usage is treated like the attr
27. workload manager launches When it sees processes that the workload manager is unable to clean up after a job is over it carries out the cleanup instead Using cm launcher is recommended if jobs that do not get cleaned up correctly are an issue for the user or administrator 6 1 5 Job Script Examples Some job script examples are given in this section Each job script can use a number of variables and directives Single Node Example Script An example script for SGE noob sgecluster cat application bin sh N sleep S bin sh Make sure that the e and o file arrive in th working directory cwd Merge the standard out and standard error to one file F sleep 60 echo Now it is date Parallel Example Script For parallel jobs the pe parallel environment option must be assigned to the script Depending on the interconnect there may be a choice be tween a number of parallel environments such as MPICH Ethernet or MVAPICH InfiniBand bin sh Your job name 5 N My Job Bright Computing Inc 6 2 Submitting A Job Use current working directory cwd Join stdout and stderr j y pe Parallel environment request Set your number of requested slots here pe mpich 2 Run job through bash shell S bin bash If modules are needed source modules environment etc profile d modules sh Add any modules you might require module add shared mpich ge gcc 64
28. 0 1 24 230 nodes p wa n the bookmarks bar Import bookmarks now Other Bookmarks For quick access place your bookmarks hi HOME WORKLOAD NODES GRAPHS Bright Cluster Manager User Portal DEVICE INFORMATION Hostname State Memory Cores CPU Speed GPU NICs IB Category Figure 10 3 User Portal Nodes Page The following information about the head or regular nodes is pre sented Hostname the node name State For example UP DOWN INSTALLING Memory RAM on the node Cores Number of cores on the node CPU Type of CPU for example Dual Core AMD Opteron Speed Processor speed GPU Number of GPUs on the node if any NICs Number of network interface cards on the node if any IB Number of InfiniBand interconnects on the node if any Category The node category that the node has been allocated by the administrator by default it is default 10 4 The GRAPHS Tab By default the GRAPHS tab displays the cluster occupation rate for the last hour figure 10 4 Bright Computing Inc 74 User Portal Bright Cluster Manager User Portal Chromium eoo 215 Bright Cluster Manag Q fi amp btep 10 1 24 230 graphs php YA For quick access place your bookmarks here on the bookmarks bar Import bookmarks now other Bookmarks Bright Computing HOME WORKLOAD NODES GRAPHS Bright Cluster Manager User Portal CLUSTER OCCUPATION RATE OccupationRate Y Datapoints 60 Inter
29. 000 Sleeperl root qw 12 03 2008 07 29 01 1 251 0 00000 Sleeperl root qw 12 03 2008 07 29 02 1 252 0 00000 Sleeperl root qw 12 03 2008 07 29 02 1 253 0 00000 Sleeperl root qw 12 03 2008 07 29 03 1 More details are visible when using the for full option The Queuetype qt ype can be Batch B or Interactive I The used tot or used free column is the count of used free slots in the queue The states column is the state of the queue qstat f queuename qtype used tot load avg arch states all q8node001 cm cluster BI 0 16 NA 1x26 amd64 au Bright Computing Inc 6 3 Monitoring A Job 41 all q node002 cm cluster BI 0 16 NA 1x26 amd64 au FE FEAE FE AE AE FE AE HEE FE FE AE FE HEE FE AE EE HE FE FE FE EE EE EEE FE AE AE FE AE AE FE AE AE FE AE AE FE AE FE FE AE EE FE E FE EE PENDING JOBS PENDING JOBS PENDING JOBS PENDING JOBS PENDING JOBS FE FE EE TE FE FE E HE FE FE TE FE FE HE TE FE FE E TE FE FE E EE HE TE FE FE RE EE E TE FE FE HE TE FE HE EEE E TE HE EH FE E TE FE E E EE E E E E E E E HE 2 2 249 0 55500 Sleeperl root qw 12 03 2008 07 29 00 1 250 0 55500 Sleeperl root qw 12 03 2008 07 29 01 1 Job state can be e d eletion E rror h old running R estarted s uspended S uspended t ransfering T hreshold w aiting The queue state can be e u nknown if the corresponding sge execd cannot be contacted a larm the load threshold is currently exceeded
30. And Software Description On the hardware side commodity hardware is generally used in Be owulf clusters to keep costs down These components are usually x86 compatible processors produced at the Intel and AMD chip foundries standard Ethernet adapters InfiniBand interconnects and switches On the software side free and open source software is generally used in Beowulf clusters to keep costs down For example the Linux operating system the GNU C compiler collection and open source implementations of the Message Passing Interface MPI standard The head node controls the whole cluster and serves files and infor mation to the nodes It is also the cluster s console and gateway to the Bright Computing Inc Introduction outside world Large Beowulf clusters might have more than one head node and possibly other nodes dedicated to particular tasks for exam ple consoles or monitoring stations In most cases compute nodes in a Beowulf system are dumb in general the dumber the better with the focus on the processing capability of the node within the cluster rather than other abilities a computer might generally have A node may there fore have one or more processing elements The processors may be standard CPUs as well as GPUs FPGAs MICs and so on enough local memory memory contained in a single node to deal with the processes passed on to the node aconnection to the rest of the cluster Nodes are config
31. I SOURCE Another function not used in the sample code returns the number of data type elements received It is used when the number of elements received might be smaller than maxcount MPI Get count amp status datatype amp nelements A 7 Blocking Non Blocking And Persistent Messages MPI Send and MPI Receive cause the running program to wait for non local communication from a network Most communication networks function at least an order of magnitude slower than local computations When an MPI process has to wait for non local communication CPU cy cles are lost because the operating system has to block the process then has to wait for communication and then resume the process An optimal efficiency is usually best achieved by overlapping com munication and computation Blocking messaging functions only allow one communication to occur at a time Non blocking messaging func tions allow the application to initiate multiple communication operations enabling the MPI implementation to proceed simultaneously Persistent non blocking messaging functions allow a communication state to per sist so that the MPI implementation does not waste time on initializing or terminating a communication A 7 1 Blocking Messages In the following example the communication implementation executes in a sequential fashion causing each process MPI Recv then MPI Send to block while waiting for its neighbor Example while looping if i_
32. READ FUNNELED serialized MPI THREAD SERIALIZED multiple MPI THREAD MULTIPLE 3 3 6 Further Recommendations Users face various challenges with running and scaling large scale jobs on peta scale production systems For example certain applications may not have enough memory per core the default environment variables may need to be adjusted or I O may dominate run time Possible ways to deal with these are Bright Computing Inc 3 3 Example MPI Run 19 Trying out various compilers and compiler flags and finding out which options are best for particular applications Changing the default MPIrank ordering This is a simple yet some times effective runtime tuning option that requires no source code modification recompilation or re linking The default MPI rank placement on the compute nodes is SMP style However other choices are round robin folded rank and custom ranking Using fewer cores per node is helpful when more memory per pro cess than the default is needed Having fewer processes to share the memory and interconnect bandwidth is also helpful in this case For NUMA nodes extra care must be taken Hybrid MPI OpenMP reduces the memory footprint Overlapping communication with computation in hybrid MPI OpenMP can be considered Some applications may perform better when large memory pages are used Bright Computing Inc Workload Management 4 1 What Is A Workload Manager A workload management
33. Slurm chapter 5 Native mode non distributed job The user creates a job script and sets a constraint miccard For example the following job script runs the dgemm test directly inside the MIC bin sh SBATCH partition micq SBATCH constraint miccard module load intel cluster runtime intel mic opt intel mic perf bin mic dgemm mic x i 2 1 2048 Bright Computing Inc 9 3 Using MIC With Workload Managers 69 Native mode MPI job The user creates a job script in the same way as for non distributed job but the nodes parameter is specified For example the next job script executes the Intel IMB MPI1 benchmark on two MICs using RDMA calls bin sh SBATCH partition micq SBATCH constraint miccard SBATCH nodes 2 SLURM_BIN cm shared apps slurm current klom arch bin MPI_DIR cm shared apps intel mpi current mic MPI RUN S MPI DIR bin mpirun APP MPI_DIR bin IMB MPIl APP_ARGS PingPong MPI_ARGS genv I MPI DAPL PROVIDER ofa v2 scif0 genv I_MPI_FABRICS shm dapl perhost 1 export LD LIBRARY PATH 1ib64 MPI DIR lib LD LIBRARY PATH export PATH SLURM BIN MPI DIR bin PATH S MPI RUN SMPI_ARGS SAPP APP ARGS The value of DAPL provider the argument 1 MPI DAPL PROVIDER should be set to ofa v2 scif0 when an application needs MIC to MIC or MIC to HOST RDMA communication A Slurm job examples given here can be found on a cluster in the following directory cm shared
34. ame to lt job SBATCH o lt jobname log gt name log gt Mail lt user address gt SBATCH mail user lt user address gt Mail on any event SBATCH mail type ALL Mail on job end SBATCH mail type END Run job in partition SBATCH p destination Run job using GPU with ID num SBATCH ber as described in section 8 5 2 gres gpu lt number gt By default both standard output and standard error go to a file slurm lt 3j gt out where lt j gt is the job number 5 3 3 Slurm Environment Variables Available environment variables include SLURM CPUS ON NODE processors available to the job on this node SLURM JOB ID job ID of executing job SLURM LAUNCH NODE IPADDR IP address of node where job launched SLURM NNODES total number of nodes SLURM NODEID relative node ID of current node SLURM NODELIST list of nodes allocated to job SLURM NTASKS total number of processes in current job SLURM PROCID MPI rank or relative process ID of the current process SLURM SUBMIT DIR directory from with job was launched SLURM TASK PID process ID of task started SLURM TASKS PER NODE number of task to be run on each node CUDA VISIBLE DEVICES which GPUs are available for use Typically end users use SLURM PROCID in a program so that an in put of a parallel calculation depends on it The calculation is thus spread across
35. ble the job is queued until resources become available For example if nodes only have 2000MB to spare and 4000MB is re quested then the job is queued indefinitely and it is up to the user to fix the problem Resource list directives also allow for example the number of nodes 7 1 nodes and the virtual processor cores per nodes 1 ppn to be specified If no value is specified the default is 1 core per node If 8 cores are wanted and it does not matter how the cores are allo cated e g 8 per node or 1 on 8 nodes the directive used in Torque is PBS 1 nodes 8 For PBS Pro v11 this also works but is deprecated and the form PBS 1 select 8 is recommended instead Bright Computing Inc 46 PBS Variants Torque And PBS Pro Further examples of node resource specification are given in a table on page 47 Job Directives Job Name Logs And IDs If the name of the job script file is jobname then by default the output and error streams are logged to jobname o lt number gt and jobname e lt number gt respectively where number indicates the asso ciated job number The default paths for the logs can be changed by us ing the o and e directives respectively while the base name jobname here can be changed using the N directive Often a user may simply merge both logs together into one of the two streams using the j directive Thus in the preceding example j oe merges the logs to the output log
36. ch mod files are written when the e moption is specified 84 Bright Computing Inc
37. cm cluster Running the executable with mpirun outside the workload manager as shown does not take the resources of the cluster into account To han dle running jobs with cluster resources is of course what workload man agers such as Slurm are designed to do Workload managers also typ ically take care of what environment modules should be loaded on the compute nodes for a job via additions that the user makes to a job script Running an application through a workload manager via a job script is introduced in Chapter 4 Appendix A contains a number of simple MPI programs 3 3 4 Hybridization OpenMP is an implementation of multi threading This is a method of parallelizing whereby a parent thread a series of instructions executed consecutively forks a specified number of child threads and a task is di vided among them The threads then run concurrently with the runtime environment allocating threads to different processors and accessing the shared memory of an SMP system MPI can be mixed with OpenMP to achieve high performance on a cluster supercomputer of multi core nodes or servers MPI creates pro cesses that reside on the level of node while OpenMP forks threads on the level of a core within an SMP node Each process executes a portion of the overall computation while inside each process a team of threads is created through OpenMP directives to further divide the problem This kind of execution makes sense due to the eas
38. com resources docs Bright Computing Inc Pro 44 PBS Variants Torque And PBS Pro 7 1 Components Of A Job Script To use Torque or PBS Pro a batch job script is created by the user The job script is a shell script containing the set of commands that the user wants to run It also contains the resource requirement directives and other spec ifications for the job After preparation the job script is submitted to the workload manager using the qsub command The workload manager then tries to make the job run according to the job script specifications A job script can be resubmitted with different parameters e g differ ent sets of data or variables 7 1 1 Sample Script Structure A job script in PBS Pro or Torque has a structure illustrated by the follow ing basic example Example bin bash PBS 1 walltime 1 00 00 PBS 1 nodes 4 PBS 1 mem 500mb PBS j oe cd HOME myprogs mpirun myprog a b c The first line is the standard shebang line used for scripts The lines that start with PBS are PBS directive lines described shortly in section 7 1 2 The last two lines are an example of setting remaining options or con figuration settings up for the script to run In this case a change to the directory myprogs is made and then run the executable myprog with ar guments a b c The line that runs the program is called the executable line section 7 1 3 To run the executable file in the execu
39. d 0 400000 normalized_load 0 006557 user bright70 However the mics consumable resource can be used only when TORQUE is used together with MOAB otherwise the job is never scheduled This behavior is subject to change but has been verified on MAUI 3 3 1 and pbs_sched 4 22 When MOAB is used then user can submit a job with offload code regions as shown in the following example Example bin sh PBS N TEST MIC OFFLOAD PBS 1 nodes 1 mics 2 module load intel mic runtime module load intel cluster runtime intel mic module load intel cluster runtime intel64 hello mic These examples can be found in cm shared examples workload torque jobscripts 9 3 4 Using MIC Cards With SGE Bright Cluster Manager distributes a branch of SGE called Open Grid Scheduler OGE OGE for which the current version is 2011 11p1 does not have native support of MIC devices and the oFFLOAD DEVICES envi ronment variable will not be set for a job by default However the support can be emulated using consumable resources 9 3 5 Using MIC Cards With openlava openlava does not have a native support of MIC devices and the OFFLOAD DEVICES environment variable is not set for a job by default However the support can be emulated using a consumable resource Bright Computing Inc 10 User Portal The user portal allows users to login via a browser and view the state of the cluster themselve
40. e MPI COMM WORLD amp numprocs find out how big the SPMD world is x MPI Comm rank MPI COMM WORLD amp myid and this processes rank is At this point all the programs are running equivalently the rank is used to distinguish the roles of the programs in the SPMD model with rank 0 often used specially if myid 0 printf d We have d processors n myid for i 1 i lt numprocs i t sprintf buff Hello d i MPI_Send buff BUFSIZE MPI_CHAR i TAG for i 1 i lt numprocs i t Bright Computing Inc numprocs MPI COMM WORLD 78 MPI Examples MPI Recv buff printf d s n else receive from rank BUFSIZI myid buff MPI COMM WORLD amp stat E MPI CHAR i TAG 0 MPI Recv buff BUFSIZI sprintf idstr strcat buff idstr strcat buff send to rank 0 MPI Send buff BUFSIZI MPI Programs end wit Processor d MPI CHAR 0 TAG MPI COMM WORLD amp stat myid reporting for duty n E MPI CHAR 0 TAG MPI COMM WORLD h MPI Finalize this is a weak synchronization poin MPI Finalize return 0 A 2 MPISkeleton The sample code below contains the complete communications skeleton for a dynamically load balanced head compute node application Fol lowing the code is a description of some of the functions necessary for writing typical parallel applications
41. e of programming that OpenMP provides OpenMP might not require copies of data structure which allows for designs that overlap computation and communication overcoming the limits of parallelism within the SMP node is of course still possible by using the power of other nodes via MPI Example include lt mpi h gt include lt omp h gt include lt stdio h gt include lt stdlib h gt Bright Computing Inc 3 3 Example MPI Run 17 int main int argc int size MPI_Init amp argc amp argv PI Comm rank MPI COMM WORLD amp myrank PI Comm size MPI COMM WORLD amp size PI Get processor name processor name amp namelength charx argv myrank namelength char processor name MPI MAX PROC ESSOR NAME printf Hello I am Processor d on s of dWMn myrank processor name size int tid 0 pragma omp parallel default shared if defined OP ENMP n of threads omp get num threads int n of threads 1 tid omp get thread num endif printf Hybrid Hello World f_threads MPI Finalize return 0 To compile the program fred brig ht 70 private tid n of threads I am thread 4 d out of d n tid n o mpicc o hybridhello omphello c fopenmp To specify the number of OpenMP threads per MPI task the environment variable O Example fred brig P_N UM_THREADS must be set ht 70
42. ecked using module whatis module whatis sge sge Adds sge to your environment The man pages for module gives further details on usage 2 4 Compiling Applications Compiling an application is usually done on the head node or login node Typically there are several compilers available on the head node which provide different levels of optimization standards conformance and sup port for accelerators For example GNU compiler collection Open64 compiler Intel compilers Portland Group compilers The following table summarizes the available compiler commands on the cluster Language GNU Open64 Portland Intel C gcc opencc pgcc icc C g opencc pgcc icc Fortran77 gfortran openf90 ff77 pgf77 ifort Fortran90 gfortran openf90 pgf90 ifort Fortran95 gfortran openf95 pgf95 ifort GNU compilers are the de facto standard on Linux and are installed by default They are provided under the terms of the GNU General Public License AMD s Open 4 is also installed by default on Bright Cluster Manager Commercial compilers by Portland and Intel are available as packages via the Bright Cluster Manager YUM repository and require the purchase of a license to use them To make a compiler available to be used in a user s shell commands the appropriate environment module section 2 3 must be loaded first On most clusters two versions of GCC are available 1 The version of GCC that comes along with the Linux distribution For
43. er Both Torque and PBS Pro are therefore covered in this chapter The possible Torque schedulers Torque s built in scheduler Maui or Moab are also covered when dis cussing Torque Torque and PBS Pro both offer a graphical interface and command line tools for submitting monitoring modifying and deleting jobs For submission and execution of jobs both workload managers use PBS job scripts The user puts values into a job script for the resources being requested such as the number of processors memory Other values are also set for the runtime parameters and application specific variables The steps for running a job through a PBS job script are Creating an application to be run via the job script Creating the job script adding directives applications runtime pa rameters and application specific variables to the script Submitting the script to the workload management system This chapter covers the using the workload managers and job scripts with the PBS variants so that users can get a basic understanding of how they are used and can get started with typical cluster usage In this chapter section 7 1 covers the components of a job script and job script ex amples section 7 2 1 covers submitting monitoring and deleting a job with a job script More depth on using these workload managers is to be found in the PBS Professional User Guide and in the online Torque documentation at http www adaptivecomputing
44. es can be set using the o and e options respectively The error and output files can usefully be concatenated into one file with the j oeor j eooptions More details on this can be found in the qsub man page 7 2 4 Monitoring A Job To use the commands in this section the appropriate workload manager module must be loaded For example for Torque torque module needs to be loaded module add torque qstat Basics The main component is qstat which has several options In this exam ple the most frequently used options are discussed In PBS Torque the command qstat an shows what jobs are cur rently submitted or running on the queuing system An example output 1S fred bright52 qstat an bright52 cm cluster User Req d Req d Elap Job ID name Queue Jobname SessID NDS TSK Memory Time S Time 78 bright52 fred shortq tjob 10476 1 1 555mb 00 01 R 00 00 79 bright52 fred shortq tjob 1 1 555mb 00 01 Q gt Bright Computing Inc 52 PBS Variants Torque And PBS Pro The output shows the Job ID the user who owns the job the queue the job name the session ID for a running job the number of nodes re quested the number of CPUs or tasks requested the time requested 1 walltime the job state S and the elapsed time In this example one job is seen to be running R and one is still queued Q The n parame ter causes nodes that are in use by a running job to display at the end of that line
45. finiBand is not available 3 3 Example MPI Run This example covers an MPI run which can be run inside and outside of a queuing system To use mpirun the relevant environment modules must be loaded For example to use the mpich over Gigabit Ethernet ge GCC imple mentation module add mpich ge gcc or to use the openmpi Open MPI GCC implementation module add openmpi gcc Similarly to use the mvapi ch InfiniBand Open64 implementation module add mvapich open64 Bright Computing Inc 3 3 Example MPI Run 13 Depending on the libraries and compilers installed on the system the availability of these packages might differ To see a full list on the system the command module avail can be typed 3 3 4 Compiling And Preparing The Application The code must be compiled with MPI libraries and an underlying com piler The correct library command can be found in the following table Language C C Fortran77 Fortran90 Fortran95 Command mpicc mpiCC mpif77 mpif90 mpif95 An MPI application myapp c built in C could then be compiled as mpicc myapp c The a out binary that is created can then be executed using the mpirun command section 3 3 3 3 3 2 Creating A Machine File A machine file contains a list of nodes which can be used by MPI pro grams The workload management system creates a machine file based on the nodes allocated for a job when the job is submitted with the workload mana
46. for each option 5 option parameter Available options and their descriptions can be seen with the output of qsub help Table 6 1 3 SGE Job Script Options Option and parameter Description a date time ac context list ar ar id A account string b y es In o binding env pe set exp lin str c ckpt selector ckpt ckpt name clear cwd C directive prefix dc simple context list dl date time e path list h hard help hold jid job identifier list hold jid ad job identifier list i file list j yles n o js job share request a start time add context variables bind job to advance reservation account string in accounting record handle command as binary binds job to processor cores define type of checkpointing for job request checkpoint method skip previous definitions for job use current working directory define command prefix for job script delete context variable s request a deadline initiation time specify standard error stream path s place user hold on job consider following requests hard print this help define jobnet interdependencies define jobnet array interdependencies specify standard input stream file s merge stdout and stderr stream of job share tree or functional job share continued Bright Computing Inc 6 1 Writing A Job Script 37 Table 6 1 3 SGE Job Script Options continued Opti
47. forms and their limitations The nvcc command by default compiles code and links the objects for both the host system and the GPU The nvcc command distinguishes be tween the two and it can hide the details from the developer To compile the host code nvcc will use gcc automatically nvcc options lt inputfile gt A simple example to compile CUDA code to an executable is nvcc testcode cu o testcod The most used options are g or debug level This generates debug able code for the host e Gor device debug level This generates debug able code for the GPU e oor output file file This creates an executable with the name file arch sm 13 This can be enabled if the CUDA device supports compute capability 1 3 which includes double precision If double precision floating point is not supported or the flag is not set warnings such as the following will come up warning Double is not supported Demoting to float The nvcc documentation manual The CUDA Compiler Driver NVCC has more information on compiler options The CUDA SDK has more programming examples and information accessible from the file SCUDA_SDK C Samples html For OpenCL code compilation can be done by linking against the OpenCL library gcc test c lOpenCL g test cpp lOpenCL nvcc test c lOpenCL 8 5 Available Tools 8 5 1 CUDA gdb The CUDA debugger can be started using cuda gdb Details of how to use it are available
48. g to the CUDA device cudaMemcpy cuda_s s len cudaMemcpyHostToDevice Set the grid and block sizes dim3 is a type and Hello World is 12 characters say 3x4 dim3 dimGrid 3 dim3 dimBlock 4 Invoke the kernel to undo shift13 in GPU helloWorld dimGrid dimBlock cuda s Retrieve the results from the CUDA device cudaMemcpy s cuda s len cudaMemcpyDeviceToHost Free up the allocated memory on the CUDA device cudaFree cuda s printf String decoded on GPU as s n s return 0 The preceding code example may be compiled and run with fred bright52 nvcc hello cu o hello fred bright52 module add shared openmpi gcc 64 1 4 4 slurm fred bright52 salloc n1 gres gpu 1 mpirun hello salloc Granted job allocation 2263 String for encode decode Hello World String encoded on CPU as Uryy dlyq Bright Computing Inc 8 5 Available Tools 63 String decoded on GPU as Hello World alloc Relinquishing job allocation 2263 salloc Job allocation 2263 has been revoked The number of characters displayed in the encoded string appear less than expected because there are unprintable characters in the encoding due to the cipher used being not exactly rot13 8 5 5 OpenACC OpenACC http www openacc standard org is a new open parallel programming standard aiming at simplifying the programma bility of heterogeneous CPU GPU computing systems OpenACC al low
49. ger job submission tool So if the user chooses to have the workload management system allocate nodes for the job then creating a machine file is not needed However if an MPI application is being run by hand outside the workload manager then the user is responsible for creating a machine file manually Depending on the MPI implementation the layout of this file may differ Machine files can generally be created in two ways Listing the same node several times to indicate that more than one process should be started on each node node001 node001 node002 node002 Listing nodes once but with a suffix for the number of CPU cores to use on each node node001 2 node002 2 3 3 8 Running The Application A Simple Parallel Processing Executable A simple hello world program designed for parallel processing can be built with MPI After compiling it it can be used to send a message about how and where it is running Bright Computing Inc Using MPI fred bright70 cat hello c include lt stdio h gt include lt mpi h gt int main int argc char xargv int id np i char processor name MPI MAX PROCESSOR NAME int processor name len MPI Init amp argc amp argv MPI Comm size MPI COMM WORLD amp np MPI Comm rank MPI COMM WORLD amp id MPI Get processor name processor name amp processor name len for i 1 i 2 i printf Hello world from process 03d out of 03d
50. ght Cluster Manager which manages a Beowulf cluster config uration It explains how to use the MPI and batch environments how to submit jobs to the queueing system and how to check job progress The specific combination of hardware and software installed may differ de pending on the specification of the cluster which means that parts of this manual may not be relevant to the user s particular cluster 1 1 What Is A Beowulf Cluster 1 1 1 Background And History In the history of the English language Beowulf is the earliest surviving epic poem written in English It is a story about a hero with the strength of many men who defeated a fearsome monster called Grendel In computing a Beowulf class cluster computer is a multiprocessor ar chitecture used for parallel computations i e it uses many processors to gether so that it has the brute force to defeat certain fearsome number crunching problems The architecture was first popularized in the Linux community when the source code used for the original Beowulf cluster built at NASA was made widely available The Beowulf class cluster computer design usu ally consists of one head node and one or more regular nodes connected together via Ethernet or some other type of network While the origi nal Beowulf software and hardware has long been superseded the name given to this basic design remains Beowulf class cluster computer or less formally Beowulf cluster 1 1 2 Brief Hardware
51. have_a_left_neighbor MPI Recv inbuf count dtype left tag comm amp status if i have a right neighbor MPI Send outbuf count dtype right tag comm do other work MPI also has the potential to allow both communications to occur si multaneously as in the following communication implementation exam ple Bright Computing Inc amp status 82 MPI Examples A 7 2 Non Blocking Messages Example while looping count 0 if i_have_a_left_neighbor MPI Irecv inbuf count dtype left tag comm amp req count if i have a right neighbor MPI Isend outbuf count dtype right tag comm amp req count MPI Waitall count req amp statuses do other work In the example MPI Waitall potentially allows both communica tions to occur simultaneously However the process as show is blocked until both communications are complete A 7 3 Persistent Non Blocking Messages A more efficient use of the waiting time means to carry out some other work in the meantime that does not depend on that communication If the same buffers and communication parameters are to be used in each iteration then a further optimization is to use the MPI persistent mode The following code instructs MPI to set up the communications once and communicate similar messages every time Example int count 0 if i have a left neighbor MPI Recv init inbuf count dtype left tag comm amp req count
52. he host CPU s Offload statements can also be combined with OpenMP directives The following hello mic example shows how a system call is of floaded to the MIC The example is used in other sections of this chapter Example finclude lt stdio h gt finclude stdlib h include lt limits h gt include lt unistd h gt int main void char hostname HOST NAME MAX pragma offload target mic gethostname hostname HOST NAME MAX printf My hostname is s n hostname exit 0 Standard command line arguments with no MIC related switch re quired compile the code This is because offloading is enabled by default in Intel Compiler version 2013 and higher Example user bright70 module load intel compiler 64 13 0 user bright70 icc 03 hello mic c o hello mic user bright70 module load intel mic runtime user bright70 hello mic My hostname is bright70 micO user bright70 8 To get debug information when an offloaded region is executed the OFFLOAD_REPORT environment variable can be used Possible values in order of increasing verbosity are 1 2 or 3 Setting the empty string dis ables debug messages Example Bright Computing Inc 68 Using MICs user bright70 module load intel mic runtime user bright70 export OFFLOAD_REPORT 2 user bright70 hello_mic Offload IC 0 File hello mic c Offload IC 0 Line 7 Offload IC Tag TagO
53. he same time For example loading openmpi gcc 64 without removing an already loaded openmpi gcc 64 can result in confusion about what compiler opencc is meant to use 2 3 3 Changing The Default Environment The initial state of modules in the user environment can be set as a de fault using the module initx subcommands The more useful ones of these are module initadd add a module to the initial state module initrm remove a module from the initial state module initlist list all modules loaded initially For multiple versions when no version is specified the alphabetically last version is chosen This usually is the latest but can be an issue when versions move from say 9 to 10 For example the following is sorted in alphabetical order v1 v10 v11 v12 v13 v2 v3 v4 v5 v6 v7 v8 v9 Bright Computing Inc 2 4 Compiling Applications module initclear clear all modules from the list of modules loaded initially Example module initclear module initlist bash initialization file HOME bashrc loads modules null module initadd shared gcc 4 8 1 openmpi gcc sge module initlist bash initialization file HOME bashrc loads modules null shared gcc 4 8 1 openmpi gcc 64 1 6 5 sge 2011 11pl In the preceding example the newly defined initial state module en vironment for the user is loaded from the next login onwards If the user is unsure about what the module does it can be ch
54. host For example user bright70 module load intel compiler 64 13 0 user bright70 icc mmic test c o test user bright70 If the GNU autoconf application is used instead then the environment variables are like those defined earlier in section 9 1 1 Detailed information on building a native application for the Intel Xeon Phi coprocessor using the Intel compilers can be found at http software intel com en us articles building a native application for intel xeon phi coprocessors Bright Computing Inc 9 2 Compiling Code In Offload Mode 67 9 2 Compiling Code In Offload Mode MICs can also build user applications in offload heterogeneous mode With this method the application starts on a host platform The Intel Compiler should be used since the current stable version of the GNU Compiler does not support this mode A special statement pragma should be added to the C C or Fortran code to mark regions of code that are to be offloaded to a MIC and run there This directive approach ressembles that used by the PGI compiler CAPS HMPP or OpenACC when these specify the offloading of code to GPUs In this case all data transfer and synchronization are managed by the compiler and runtime When an offload code region is encountered and a MIC is found on the host the offload code and data are transferred and run on the MIC coprocessor If no available MIC devices are found then the offload code is run on t
55. ibited mode no compute contexts are allowed to run on the GPU 3 Exclusive process mode only one compute context is allowed to run on the GPU usable from multiple threads at a time To check the state of the GPU nvidia smi i 0 q COMPUTE mode rules for GPU 0 1 In this example GPUO is locked and there is a running application using GPUO A second application attempting to run on this GPU will not be able to run on this GPU histogram device 0 main cpp 101 cudaSafeCall Runtime API error no CUDA capable device is available After use the GPU can be unlocked to allow multiple users nvidia smi i 0 c 0 Bright Computing Inc 8 5 Available Tools 61 8 5 3 CUDA Utility Library CUTIL is a simple utility library designed for use in the CUDA SDK sam ples There are 2 parts for CUDA and OpenCL The locations are CUDA_SDK C lib CUDA SDK OpenCL common lib Other applications may also refer to them and the toolkit libraries have already been pre configured accordingly However they need to be com piled prior to use Depending on the cluster this might have already have been done fred demo cd fred demo cp r SCUDA_SDK fred demo cd basename CUDA_SDK cd C fred demo C make fred demo C cd basename CUDA SDK cd OpenCL fred8 demo OpenCL make CUTIL provides functions for parsing command line arguments read and writing binary files and PPM format i
56. ibutes of a resource which the cluster administrator will have pre configured according to local requirements For further details on resource list directives the Torque and PBS Pro user documentation should be consulted 7 1 3 The Executable Line In the job script structure section 7 1 1 the executable line is launched with the job launcher command after the directives lines have been dealt with and after any other shell commands have been carried out to set up the execution environment Using mpirun In The Executable Line The mpirun command is used for executables compiled with MPI li braries Executables that have not been compiled with MPI libraries or which are launched without any specified number of nodes run on a sin gle free node chosen by the workload manager The executable line to run a program myprog that has been compiled with MPI libraries is run by placing the job launcher command mpi run before it as follows mpirun myprog Bright Computing Inc 48 PBS Variants Torque And PBS Pro Using cm launcher With mpirun In The Executable Line For Torque for some MPI implementations jobs sometimes leave pro cesses behind after they have ended A default Bright Cluster Manager installation provides a cleanup utility that removes such processes To use it the user simply runs the executable line using the cm 1auncher wrapper before the mpirun job launcher command cm launcher mpirun myprog The wrapper
57. ing as part of the video processor but their ability to handle non graphics tasks in a similar manner has become important for gen eral computing GPUs designed for general purpose computing task are commonly called General Purpose GPUs or GPGPUs A GPU is suited for processing an algorithm that naturally breaks down into a process requiring many similar calculations running in paral lel GPUs cores are able to rapidly apply the instruction on multiple data points organized in a 2 D and more recently 3 D image The image is placed in a framebuffer In the original chips the data points held in the framebuffer were intended for output to a display thereby accelerating image generation The similarity between multicore CPU chips and modern GPUs makes it at first sight attractive to use GPUs for general purpose computing However the instruction set on GPGPUs is used in a component called the shader pipeline This is as the name suggests to do with a limited set of graphics operations and so is by its nature rather limited Using the instruction set for problems unrelated to shader pipeline manipulations requires that the problems being processed map over to a similar manip ulation This works best for algorithms that naturally break down into a process requiring an operation to be applied in the same way on many independent vertices and pixels In practice this means that 1 D vector operations are an order of magnitude less efficient on GPU
58. ke an ordinary shell script and certain commands and variables can be put in there that are needed for the job The exact composition of a job script depends on the workload manager used but normally includes commands to load relevant modules or set environment variables directives for the workload manager to request resources control the output set email addresses for messages to go to anexecution job submission line When running a job script the workload manager is normally respon sible for generating a machine file based on the requested number of pro cessor cores np as well as being responsible for the allocation any other requested resources The executable submission line in a job script is the line where the job is submitted to the workload manager This can take various forms Example For the Slurm workload manager the line might look like srun mpi mpichl p4 a out Example For Torque or PBS Pro it may simply be Bright Computing Inc 4 6 Running Jobs On A Workload Manager 23 mpirun a out Example For SGE it may look like mpirun np 4 machinefile TMP machines a out 4 6 Running Jobs On A Workload Manager The details of running jobs through the following workload managers is discussed later on for Slurm Chapter 5 e SGE Chapter 6 Torque with Maui or Moab and PBS Pro Chapter 7 4 7 Running Jobs In Cluster Extension Cloud Nodes Using cmsub Extra comp
59. m cluster PBS O HOST mascm4 cm cluster PBS O WORKDIR home cvsupport test package PBS O QUEUE default etime Tue Jul 14 12 35 31 2009 submit args pbs job q default Variable Lis Job Details With checkjob The checkjob command only for Maui is particularly useful for checking why a job has not yet executed For a job that has an excessive memory requirement the output looks some thing like Bright Computing Inc 54 PBS Variants Torque And PBS Pro fred bright52 checkjob 65 checking job 65 State Idle Creds user fred group fred class shortq qos DEFAULT WallTime 00 00 00 of 00 01 00 SubmitTime Tue Sep 13 15 22 44 Time Queued Total 2 53 41 Eligible 2 53 41 Total Tasks 1 Req 0 TaskCount 1 Partition ALL Network NONE Memory gt 0 Disk gt 0 Swap gt 0 Opsys NONE Arch NONE Features NONE Dedicated Resources Per Task PROCS 1 MEM 495M IWD NONE Executable NONE Bypass 0 StartCount 0 PartitionMask ALL Flags RESTARTABLE PE 1 01 StartPriority 173 job cannot run in partition DEFAULT idle procs do not meet requirement S 0 of 1 procs found idle procs 3 feasible procs 0 Rejection Reasons CPU 3 The v option gives slightly more detail 7 2 5 Deleting A Job An already submitted job can be deleted using the qde1 command gdel jobid The
60. mages comparing data arrays typically used for comparing GPU results with CPU results timers macros for checking error codes checking for shared memory bank conflicts 8 5 4 CUDA Hello world Example A hello world example code using CUDA is Example CUDA example Hello World using shift13 a rotl3 like function Encoded on CPU decoded on GPU rot13 cycles between 26 normal alphabet characters shift13 shifts 13 steps along the normal alphabet characters So it translates half the alphabet into non alphabet characters shift13 is used because it is simpler than rot13 in c So we can focus on the point c Bright Computing Taras Shapovalov lt taras shapovalov brightcomputing com gt Bright Computing Inc 62 Using GPUs include lt cuda h gt x remove this line in CUDA 6 onwards x include cutil inline h remove this line in CUDA 6 onwards include lt stdio h gt CUDA kernel definition undo shift13 __global__ void helloWorld char str int idx blockIdx x blockDim x threadIdx x str idx 13 int main int argc charx argv char s Hello World printf String for encode decode s n s CPU shift13 int len sizeof s for int i 0 i lt len i s i 13 printf String encoded on CPU as s n s Allocate memory on the CUDA device char xcuda S cudaMalloc void amp cuda_s len Copy the strin
61. mmand mpirun application Soptions Run the parallel MPI executable nodes ppn echo Running mpirun command eval Smpirun_command Bright Computing Inc If modules are needed by the script then source modules environment 50 PBS Variants Torque And PBS Pro In the preceding script no machine file is needed since it is automat ically built by the workload manager and passed on to the mpi run par allel job launcher utility The job is given a unique ID and run in parallel on the nodes based on the resource specification 7 1 5 Links To Other Resources About Job Scripts In Torque And PBS Pro A number of useful links are Torque examples http bmi cchmc org resources software torque examples PBS Pro script files http www ccs tulane edu computing pbs pbs phtml 7 2 Submitting A Job 7 2 1 Preliminaries Loading The Modules Environment To submit a job to the workload management system the user must en sure that the following environment modules are loaded If using Torque with no external scheduler module add shared torque If using Torque with Maui module add shared torque maui If using Torque with Moab module add shared torque moab f using PBS Pro module add shared pbspro Users can pre load particular environment modules as their default using the module init commands section 2 3 3 7 2 2 Using qsub The command qsub is used to submit jobs to the workload manage
62. mple Script Structure 22222220 7 152 Directves aaa seine he a 71 3 The Executable Line 7 1 4 Example Batch Submission Scripts 7 4 5 Links To Other Resources About Job Scripts In Torque And PBS Pro 004 7 2 Submitting AJob eee 7 2 1 Preliminaries Loading The Modules Environment 722 2 Using asubiri mes ee Mar ana 7 23 Job Output 26 ou 22282 zv EY 724 Monitoring AJob 22 222000 725 DeletingAJob es ea a e aa a 7 2 6 Monitoring Nodes In Torque And PBS Pro Using GPUs 8 1 Packages cem Rcg 35 2 Ser De ee E 8 2 Using C UDA 6 wae rar Bee ae aaa 8 3 Using OpenCL i vice as abe Pa eem aes e 6 8 4 Compiling Code 0 0 0 0 0 0000 85 Available Tools 27 27 28 28 30 30 31 32 33 33 35 35 35 36 36 37 38 39 40 40 42 43 47 48 50 50 50 50 51 51 54 54 Table of Contents iii 85 CUDA gab 2 24 0 21 2 2 00 eee ie 8 52 mVvidia SM eum tert RE d nuptum 853 CUDA Utility Library sss 85 4 CUDA Hello world Example 8 55 Open AGG icem el b rte atra 9 Using MICs 10 9 1 9 2 9 3 Compiling Code In Native Mode 9 1 1 Using The GNU Compiler 9 1 2 Using Intel Compilers Compiling Code In Offload Mode Using MIC With Workload Managers 9 3 1 Using MIC Cards WithSlurm
63. nc Workload Management All cmsub command line options can also be specified in a job directive style format in the job script itself using the CMSUB tag to indicate an option Example cat myscript2 bin sh CMSUB regions us west 2 eu west 1 CMSUB input list home user myjob in CMSUB output list home user myjob out CMSUB remote output list home user file which will be created CMSUB input home user onemoreinput dat CMSUB input home user myexec myexec cmsub myscript2 Upload job id 4 User job id 5 Download job id 6 4 8 Configuring Passwordless Login To Cloud Nodes Logging in to a cloud node is possible in a similar way to regular nodes There is however a minor initial complication passwordless ssh login to the cloud node or cloud director is not configured by default This is because passwordless ssh uses certificate based authentication and the cloud director thus needs to have the ssh public key certificate of the user in the home directory of that user in the cloud director Since the home directory of the user in the cloud director is only created after the first login of the user to the cloud director there can be no ssh public key certificate in that location to begin with This existential issue can be dealt with with the following one time actions The user after logging in generates an ssh key pair on the head node with an empty passphrase The u
64. ng from 1 The command sbatch usage lists possible options that can be used on the command line or in the job script Command line values override script provided values 5 3 2 Slurm Job Script Options Options sometimes called directives can be set in the job script file using this line format for each option SBATCH option parameter Directives are used to specify the resource allocation for a job so that Slurm can manage the job optimally Available options and their de scriptions can be seen with the output of sbatch help The more overviewable usage output from sbatch usage may also be helpful Some of the more useful ones are listed in the following table Bright Computing Inc 32 Slurm Directive Description Specified As Name the job lt jobname gt SBATCH J lt jobname gt Request at least lt minnodes gt nodes SBATCH N lt minnodes gt Request lt minnodes gt to lt maxn SBATCH N odes gt nodes lt minnodes gt lt maxnodes gt Request at least lt MB gt amount of SBATCH tmp lt MB gt temporary disk space Run the job for a time of lt wall SBATCH t lt walltime gt time gt Run the job at lt time gt SBATCH begin lt time gt Set the working directory to lt di SBATCH D rectorypath gt lt directorypath gt Set error log name to lt job SBATCH e lt jobname err gt name err gt Set output log n
65. ng the first line the 1 line to the preferred shell Any shell specified by the first line can also be overridden by using the PBS S directive to set the shell path Walltime Directive The workload manager typically has default walltime limits per queue with a value limit set by the administrator The user sets walltime limit by setting the PBS 1 walltime directive to a specific time The time specified is the maximum time that the user expects the job should run for and it allows the workload manager to work out an optimum time to run the job The job can then run sooner than it would by default If the walltime limit is exceeded by a job then the job is stopped and an error message like the following is displayed PBS job killed walltime lt runningtime gt exceeded limit settime Here runningtime indicates the time for which the job actually went on to run while settime indicates the time that the user set as the wall time resource limit Resource List Directives Resource list directives specify arguments to the 1 directive of the job script and allow users to specify values to use instead of the system de faults For example in the sample script structure earlier a job walltime of one hour and a memory space of at least 500MB are requested the script requires the size of the space be spelled in lower case so 500mb is used If a requested resource list value exceeds what is availa
66. nodes to NUMA non Uniform Memory Ac cess systems and interconnected cluster nodes The available MPI implementation for the variant MPI 3 is MPICH version 3 Open MPI supports both variants These MPI libaries can be compiled with GCC Open64 Intel or PGI Depending on the cluster hardware the interconnect available may be Ethernet GE InfiniBand IB or Myrinet MX Depending on the cluster configuration MPI implementations for dif ferent compilers can be loaded By default MPI implementations that are installed are compiled and made available using both GCC and Open64 The interconnect and compiler implementation can be worked out from looking at the module and package name The modules available can be searched through for the compiler variant and then the package providing it can be found Example fred bright70 search for modules starting with the name openmpi fred bright70 module 1 avail 2 gt amp 1 grep openmpi openmpi gcc 64 1 6 5 2013 09 05 22 01 44 openmpi intel 64 1 6 5 2013 09 05 21 23 57 openmpi open64 64 1 6 5 2013 09 05 22 28 37 fred bright70 rpm qa grep openmpi openmpi geib open64 64 1 6 5 165 cm7 0 x86 64 openmpi geib gcc 64 1 6 5 165 cm7 0 x86 64 openmpi ge intel 64 1 6 5 165 cm7 0 x86 64 Here for example openmpi geib open64 64 1 6 5 165 cm7 0 x86 64 implies Open MPI version 1 6 5 compiled for both Gigabit Ethernet ge and InfiniBand ib with the O
67. o cation from the cluster for the executable In the current cluster example it can be 1 to 4 tasks The default resources available on a cluster are the number of available processor cores A resource allocation request for 5 tasks with the ntasks option causes an error because it exceeds the default resources available on the 4 core cluster fred bright52 salloc n5 mpirun hello salloc error Failed to allocate resources More processors requested t han permitted Adding and configuring just one more node to the current cluster would allows the resource allocation to succeed since an added node would provide at least one more processor to the cluster Number of tasks requested per node The value assigned to the ntasks per node option is the number of tasks that are requested for allocation from each node on the cluster In the current cluster exam ple it can be 1 to 4 tasks A resource allocation request for 5 tasks per Bright Computing Inc 30 Slurm node with ntasks per node fails on this 4 core cluster giving an output like fred bright52 salloc ntasks per node 5 mpirun hello salloc error Failed to allocate resources More processors requested t han permitted Adding and configuring another 4 core node to the current cluster would still not allow resource allocation to succeed because the request is for at least 5 cores per node rather than per cluster Restricting the number
68. odify the status of jobs freeze hold the job resume the job delete the job Bright Computing Inc 22 Workload Management the priority and execution order for jobs the run status of a job For example by adding checkpoints to freeze a job optional how related tasks in a job are handled according to their resource requirements For example a job with two tasks may have a greater need for disk I O resources for the first task and a greater need for CPU resources during the second task Some workload managers can adapt to external triggers such as hard ware failure and send alerts or attempt automatic recovery 4 4 Job Submission Process Whenever a job is submitted the workload management system checks on the resources requested by the job script It assigns cores accelerators local disk space and memory to the job and sends the job to the nodes for computation If the required number of cores or memory are not yet available it queues the job until these resources become available If the job requests resources that are always going to exceed those that can be come available then the job accordingly remains queued indefinitely The workload management system keeps track of the status of the job and returns the resources to the available pool when a job has finished that is been deleted has crashed or successfully completed 4 5 What Do Job Scripts Look Like A job script looks very much li
69. oes A Workload Manager Function 44 Job Submission Process 2 2222 4 5 What Do Job Scripts Look Like 2 2 2 4 6 Running Jobs On A Workload Manager 47 Running Jobs In Cluster Extension Cloud Nodes Using EMSUB HE 44 dase anced desde der eoe p psu lt NFP RP RP m Oo CONN DD UO GI A ce 11 12 12 12 12 12 13 13 13 16 18 18 Table of Contents 48 Configuring Passwordless Login To Cloud Nodes Slurm 5 1 Loading Slurm Modules And Compiling The Executable 52 Running The Executable With salloc 5 2 1 Node Allocation Examples 53 Running The Executable As A Slurm Job Script 5 3 1 Slurm Job Script Structure 2 2 0 0004 53 2 Slurm Job Script Options 5 3 3 Slurm Environment Variables 2 22 5 3 4 Submitting The Slurm Job Script 5 3 5 Checking And Changing Queued Job Status SGE 6 1 Writing A Job Script o 2 00 sms ua 6 11 Directives 3 2 zu doque Baar aan S 6 1 2 SGE Environment Variables 613 Job Script Options 2 22 22 6 1 4 The Executable Line 615 Job Script Examples lesen 62 Submitting A Job 2 0 000000000000 6 2 1 Submitting To A Specific Queue 6 3 Monitoring A Job aa 64 Deleting AJob eee PBS Variants Torque And PBS Pro 71 Components Of AJobScript llle 711 Sa
70. of tasks that can run per node A resource al location request for 2 tasks per node with the ntasks per node op tion and simultaneously an allocation request for 1 task to run on the cluster using the ntasks option runs successfully although it use lessly ties up resources for 1 task per node fre bright52 salloc ntasks per node 2 ntasks 1 mpirun hello salloc Granted job allocation 574 Hello world from process 0 out of 1 host name node005 salloc Relinquishing job allocation 574 The other way round that is a resource allocation request for 1 task per node with the ntasks per node option and simultaneously an allocation request for 2 tasks to run on the cluster using the ntasks option fails because on the 1 cluster node only 1 task can be allocated resources on the single node while resources for 2 tasks are being asked for on the cluster fred bright52 salloc ntasks per node 1 ntasks 3 mpirun hello salloc error Failed to allocate resources Requested node configuratio n is not available salloc Job allocation 575 has been revoked 5 3 Running The Executable As A Slurm Job Script Instead of using options appended to the salloc command line as in section 5 2 it is usually more convenient to send jobs to Slurm with the sbatch command acting on a job script A job script is also sometimes called a batch file In a job script the user can add and adjust the Slurm options which are the
71. on and parameter Description jsv jsv url l resource list m mail options masterq wc queue list notify now y es n o M mail 1 N name o path 1l P projec p priori pe pe na q wc queue 1l R yles r yles sc context shell yl soft ist ist CY n o n o es c name me slot range ist list n o sync y es n o S path 1l ist t task id range tc max running tasks terse v variable list verify M w e w n vlp wd working directory file job submission verification script to be used request the given resources define mail notification events bind master task to queue s notify job before killing suspending it start job immediately or not at all notify these e mail addresses specify job name specify standard output stream path s set job s project define job s relative priority request slot range for parallel jobs bind job to queue s reservation desired define job as not restartable set job context replaces old context start command with or without wrapping lt loginshell gt c consider following requests as soft wait for job to end and return exit code command interpreter to be used create a job array with these tasks throttle the number of concurrent tasks experimental tersed output print only the job id export these environment variables do not submit just verify expo
72. ong long format avail and list format h human readable format avail and list format v verbose nabl verbose messages s silent disable verbose messages c create create caches for avail and apropos i icase case insensitive u userlvl lt lvl gt set user level to nov ice exp ert adv anced Available SubCommands and Args add load modulefile modulefile rm unload modulefile modulefile switch swap modulefilel modulefile2 display show modulefile modulefile avail modulefile modulefile use a append dir PAIE zu unuse dir dir se update refresh purge list clear help modulefile modulefile whatis modulefile modulefile apropos keyword string initadd modulefile modulefile initprepend modulefile modulefile initrm modulefile modulefile initswitch modulefilel modulefile2 initlist initclear 2 3 2 Changing The Current Environment The modules loaded into the user s environment can be seen with module list Modules can be loaded using the add or load options A list of mod ules can be added by spacing them module add shared open64 openmpi open64 The shared module is special If itis to be loaded it is usually placed first in a list before the other modules because the other modules often depend on it The shared module is described further shortly The module avail command lists all modules that are available
73. orized keys to make sure we haven t added extra keys that you weren t expecting galt bright70 ssh cnode001 Last login Tue May 27 08 24 07 2014 from master cm cluster galt cnode001 If the user is able to login to a cloud node cnode001 in the preceding example then it implies that the user can login to the cloud director any way so the user should now be correctly configured with passwordless authentication to the cloud nodes Bright Computing Inc Slurm is a workload management system developed originally at the Lawrence Livermore National Laboratory Slurm used to stand for Sim ple Linux Utility for Resource Management However Slurm has evolved since then and its advanced state nowadays means that the acronym is obsolete Slurm has both a graphical interface and command line tools for sub mitting monitoring modifying and deleting jobs It is normally used with job scripts to submit and execute jobs Various settings can be put in the job script such as number of processors resource usage and application specific variables The steps for running a job through Slurm are to Create the script or executable that will be handled as a job Create a job script that sets the resources for the script executable Submit the job script to the workload management system The details of Slurm usage depends upon the MPI implementation used The description in this chapter will cover using Slurm s Open MPI implemen
74. ort Idle 2 2 00 00 Tue Jul 14 12 46 22 3 Idle Jobs BLOCK JOBNAME USERNAME STATE PROC WCLIMIT QUEUETIME GI Oo C UJ P Total Jobs 4 Active Jobs 1 Idle Jobs 3 Blocked Jobs 0 Viewing Job Details With qstat And check job Job Details With gstat With qstat f the full output of the job is displayed The output shows what the jobname is where the error and output files are stored and various other settings and variables qstat f Job Id 19 mascm4 cm cluster Job Name TestJobPBS Job Owner cvsupport mascm4 cm cluster job state Q queue testq Server mascm4 cm cluster Checkpoint u ctime Tue Jul 14 12 35 31 2009 Error Path mascm4 cm cluster home cvsupport test package TestJobPBS e19 Hold Types n Join Path n Keep Files n ail Points a mtime Tue Jul 14 12 35 31 2009 Output Path mascm4 cm cluster home cvsupport test package TestJobPB S019 Priority 0 qtime Tue Jul 14 12 35 31 2009 Rerunable Tru Resource_List nodect 1 Resource List nodes 1 ppn 2 Resource List walltime 02 00 00 t PBS O HOME home cvsupport PBS O LANG en US UTF 8 PBS O LOGNAME cvsupport PBS O PATH usr kerberos bin usr local bin bin usr bin sbin usr sbin home cvsupport bin cm shared apps torque 2 3 5 bin cm shar ed apps torque 2 3 5 sbin PBS O MAIL var spool mail cvsupport PBS O SHELL bin bash PBS_SERVER mascm4 c
75. ource specification are provided The list of all nodes for a cluster can be found using the pbsnodes command section 7 2 6 Bright Computing Inc 7 1 Components Of A Job Script 49 Using InfiniBand A sample PBS script for InfiniBand is bin bash l Sample PBS file Name of job PBS N MPI Number of nodes in this case 8 nodes with 4 CPUs each The total number of nodes passed to mpirun will be nodesxppn Second entry Total amount of wall clock time true time 02 00 00 indicates 02 hours PBS 1 nodes 8 ppn 4 walltime 02 00 00 Mail to user when job terminates or aborts PBS m ae etc profile d modules sh Add any modules you might require module add shared mvapich gcc torque maui pbspro Full path to application application name application lt application gt Run options for the application options lt options gt Work directory workdir lt work dir gt EEE HEE HE EH EEE EE HEHE HH EHH EE HE HE HHH HE HE HH HEE HE 2 2 2 EH HEE You should not have to change anything below this line EEE HEHE HE EEE EH EEE HH HEHE EH HE HE HH EH HH HH HE HE HEE EE HEE change the working directory default is home directory cd workdir echo Running on host hostname echo Time is date echo Directory is pwd echo PBS job ID is PBS_JOBID echo This job runs on the following machines echo cat SPBS NODEFILE uniq Smpirun_co
76. path while j eo would merge it to error log path The job ID is an identifier based on the job number and the FODN of the login node For a login node called bright52 cm cluster the job ID for a job number with the associated value lt number gt from earlier would by default be lt number gt bright52 cm cluster but it can also simply be abbreviated to number Job Queues Sending a job to a particular job queue is sometimes appropriate An administrator may have set queues up so that some queues are for very long term jobs or some queues are for users that require GPUs Submit ting a job to a particular queue destination is done by using the directive PBS q lt destination gt Directives Summary A summary of the job directives covered with a few extras are shown in the following table Directive Description Specified As Name the job lt jobname gt PBS N lt jobname gt Run the job for a time of lt walltime gt PBS l lt walltime gt Run the job at lt time gt PBS a time Set error log name to lt jobname err gt PBS lt jobname err gt Set output log name to lt jobname log gt PBS o lt jobname log gt Join error messages to output log PBS j eo Join output messages to error log PBS j o Mail to lt user address gt PBS M lt user address gt Mail on lt event gt PBS m lt event gt where lt event gt takes the a bort
77. pen64 open64 compiler for a 64 bit Bright Computing Inc 12 Using MPI architecture packaged as a cluster manager cm package for version 7 0 of Bright Cluster Manager for the x86 64 architecture 3 1 Interconnects Jobs can use particular networks for intra node communication 3 1 1 Gigabit Ethernet Gigabit Ethernet is the interconnect that is most commonly available For Gigabit Ethernet no additional modules or libraries are needed The Open MPI MPICH implementations will work over Gigabit Ethernet 3 1 2 InfiniBand InfiniBand is a high performance switched fabric which is characterized by its high throughput and low latency Open MPI MVAPICH and MVA PICH2 are suitable MPI implementations for InfiniBand 3 2 Selecting An MPI implementation Once the appropriate compiler module has been loaded the MPI imple mentation is selected along with the appropriate library modules The following list compiler indicates a choice of gcc intel open64 or pgi mpich ge lt compiler gt mvapich lt compiler gt mvapich2 lt compiler gt openmpi lt compiler gt After the appropriate MPI module has been added to the user environ ment the user can start compiling applications The mpich and openmpi implementations may be used on Ethernet On InfiniBand mvapich mvapich2 and openmpi may be used Open MPI s openmpi implemen tation will first attempt to use InfiniBand but will revert to Ethernet if In
78. pends on how the system administra tor has configured CUDA 8 2 Using CUDA After installation of the packages for general usage and compilation it is sufficient to load just the CUDA version toolkit module where ver sion is a number 42 or 50 indicating the version module add cuda50 toolkit Also available are several other modules related to CUDA e cuda50 blas Provides paths and settings for the CUBLAS library e cuda50 fft Provides paths and settings for the CUFFT library The toolkit comes with the necessary tools and the NVIDIA compiler wrapper to compile CUDA C code Extensive documentation on how to get started the various tools and how to use the CUDA suite is in the CUDA INSTALL PATH doc direc tory 8 3 Using OpenCL OpenCL functionality is provided with the cuda version toolkit en vironment module where version is a number 42 or 50 Examples of OpenCL code can be found in the CUDA_SDK OpenCL directory 8 4 Compiling Code Both CUDA and OpenCL involve running code on different platforms host with one or more CPUs device with one or more CUDA enabled GPUs Bright Computing Inc 8 5 Available Tools 59 Accordingly both the host and device manage their own memory space and it is possible to copy data between them The CUDA and OpenCL Best Practices Guides in the doc directory provided by the CUDA toolkit package have more information on how to handle both plat
79. performance By default gcc with MIC support is in the following directory Bright Computing Inc 66 Using MICs usr linux klom lt version gt To build a native application on an x86_64 host the compiler tools prefixed by x86 64 k10om linux have to be used Example user bright70 module load intel mic runtime user bright70 x86 64 klom linux gcc test c o test user bright70 If the GNU autoconf tool is used then the following shell commands can be used to build the application MIC_ARCH klom GCC_VERSION 4 7 GCC ROOT usr linux MIC ARCH GCC VERSION configure CXX GCC_ROOT bin x86_64 MIC_ARCH linux gt CXXFLAGS I GCC ROOT linux MIC ARCH usr include CXXCPP GCC ROOT bin x86 64 MIC ARCHj linux cpp CC GCC ROOT bin x86 64 MIC ARCHj linux gcc CFLAGS IS GCC ROOT linux MIC ARCH usr include CPP GCC_ROOT bin x86_64 MIC_ARCH linux cpp LDFLAGS L GCC ROOT linux MIC ARCH usr lib64 LD GCC ROOT bin x86 64 MIC ARCHj linux ld build x86 64 redhat linux host x86_64 MIC_ARCH linux target x86_64 MIC_ARCH linux make 9 1 2 Using Intel Compilers Intel Composer XE version 2013 and higher can also be used to compile a native application on Intel Xeon Phi coprocessors The compiler for this is at opt intel composer_xe_ lt version gt The mmic switch generates code for the MIC on the non MIC
80. processors according to the assigned SLURM PROCID so that each processor handles the parallel part of the calculation with different val ues More information on environment variables is also to be found in the man page for sbatch Bright Computing Inc 5 3 Running The Executable As A Slurm Job Script 33 5 3 4 Submitting The Slurm Job Script Submitting a Slurm job script created like in the previous section is done by executing the job script with sbatch fred bright52 sbatch slurmhello sh Submitted batch job 703 fred bright52 cat slurm 703 out Hello world from process 001 out of 004 processor name node001 Queues in Slurm terminology are called partitions Slurm has a de fault queue called defq The administrator may have removed this or created others If a particular queue is to be used this is typically set in the job script using the p or partition option SBATCH partition bitcoinsq It can also be specified as an option to the sbat ch command during sub mission to Slurm 5 3 5 Checking And Changing Queued Job Status After a job has gone into a queue the queue status can be checked using the squeue command The job number can be specified with the j op tion to avoid seeing other jobs The man page for squeue covers other options Jobs can be canceled with scancel job number The scontrol command allows users to see and change the job di rectives while the job is still que
81. ps in a code that vectorize well are good candidates for OpenMP There are also some potential drawbacks OpenMP can be hard to program and or debug in some cases Effective usage can be complicated on NUMA systems due to local ity considerations If an application is network or memory bandwidth bound then threading it is not going to help In this case it will be OK to leave some cores idle In some cases a serial portion may be essential which can inhibit P y performance In most MPI codes synchronization is implicit and happens when messages are sent and received However with OpenMP much synchronization must be added to the code explicitly The program mer must also explicitly determine which variables can be shared among threads and which ones cannot parallel scoping OpenMP codes that have errors introduced by incomplete or misplaced syn chronization or improper scoping can be difficult to debug because the error can introduce race conditions which cause the error to hap pen only intermittently 3 3 5 Support Thread Levels MPI defines four levels of thread safety The maximum thread support level is returned by the MPI Init thread call in the provided argu ment An environment variable MPICH MAX THREAD SAFETY can be set to different values to increase the thread safety MPICH MAX THREAD SAFETY Supported Thread Level not set MPI THREAD SINGLE single MPI THREAD SINGLE funneled MPI TH
82. r analyze results Therefore it should rarely be necessary for a user to log on to one of the nodes and in some cases node logins are disabled altogether The head login and compute nodes usually communicate with each other through a gigabit Ethernet network capable of transmitting information at a maximum rate of 1000 Mbps In some clusters 10 gigabit Ethernet 10GE 10GBE or 10GigE is used capable of up to 10 Gbps rates Sometimes an additional network is used by the cluster for even faster communication between the compute nodes This particular network is mainly used for programs dedicated to solving large scale computational problems which may require multiple machines and could involve the exchange of vast amounts of information One such network topology is InfiniBand commonly capable of transmitting information at a max imum effective data rate of about 124Gbps and about 1 2js end to end latency on small packets for clusters in 2013 The commonly available maximum transmission rates will increase over the years as the technol ogy advances Applications relying on message passing benefit greatly from lower latency The fast network is usually complementary to a slower Ethernet based network Bright Computing Inc Cluster Usage 2 1 Login To The Cluster Environment The login node is the node where the user logs in and works from Simple clusters have a single login node but large clusters sometimes have mul tiple login
83. r sys tem The command returns a unique job identifier which is used to query and control the job and to identify output The usage format of qsub and some useful options are listed here USAGE qsub lt options gt job script Option Hint Description a at run the job at a certain time Fal list request certain resource s q queue job is run in this queue N name name of job S shell shell to run job under join join output and error files Bright Computing Inc 7 2 Submitting A Job For example a job script called mpirun job with all the relevant di rectives set inside the script may be submitted as follows Example qsub mpirun job A job may be submitted to a specific queue test q as follows Example qsub q testq mpirun job The man page for qsub describes these and other options The options correspond to PBS directives in job scripts section 7 1 1 If a particular item is specified by a qsub option as well as by a PBS directive then the qsub option takes precedence 7 2 3 Job Output By default the output from the job script lt scriptname gt goes into the home directory of the user for Torque or into the current working directory for PBS Pro By default error output is written to lt scriptname gt e jobid and the application output is written to lt scriptname gt o jobid where jobid is a unique number that the workload manager allocates Specific output and error fil
84. rt all environment variables verify mode error warning none just verify poke for jobs use working directory read commandline input from file More detail on these options and their use is found in the man page for qsub 6 1 4 The Executable Line In a job script the executable line is launched with the job launcher com mand after the directives lines have been dealt with and after any other shell commands have been carried out to set up the execution environ ment Bright Computing Inc 38 SGE Using mpirun In The Executable Line The mpi run job launcher command is used for executables compiled with MPI libraries Executables that have not been compiled with MPI li braries or which are launched without any specified number of nodes run on a single free node chosen by the workload manager The executable line to run a program myprog that has been compiled with MPI libraries is run by placing the job launcher command mpirun before it as follows mpirun myprog Using cm launcher With mpi run In The Executable Line For SGE for some MPI implementations jobs sometimes leave processes behind after they have ended A default Bright Cluster Manager installa tion provides a cleanup utility that removes such processes To use it the user simply runs the executable line using the cm 1auncher wrapper before the mpi run job launcher command cm launcher mpirun myprog The wrapper tracks processes that the
85. s It is a read only interface The first time a browser is used to login to the cluster portal a warn ing about the site certificate being untrusted appears in a default Bright Cluster Manager configuration This can safely be accepted 10 1 Home Page The default home page allows a quick glance to convey the most impor tant cluster related information for users figure 10 1 Bright Cluster Manager User Portal Chromium ooo right Cluster Manag Q f amp btep 10 1 24 230 index php ZA For quick access place your bookmarks here on the bookmarks bar Import bookmarks now other Bookmarks Bright Computing HOME WORKLOAD NODES GRAPHS Bright Cluster Manager User Portal MESSAGE OF THE DAY DOCUMENTATION CONTACT ompletely operational and all n Bright Computing website ii Administrator manual User manua CLUSTER OVERVIEW Uptime 3 6 hc 03 Memory 1 1 GiB 3 1 GiB Nodes 5t440 Swap 316 9 MiB out of 81 9 GiB total Devices 0t140 Load 0 496 use Cores 515 1 196 sy Users 0 f 2 total 94 6 idle Phase Load N A ampere 3 9 othe Occupation Rate 3 4 WORKLOAD OVERVIEW Queue Scheduler Slots Nodes Running Queued Failed Completed Avg Duration Est Delay Figure 10 1 User Portal Default Home Page The following items are displayed on a default home page Bright Computing Inc 72 User Portal a Message Of The Day The administrator may put up important messages for users here e links to
86. s parallel programmers to provide OpenACC directives to the compiler identifying which areas of code to accelerate This frees the programmer from carrying out time consuming modifications to the original code it self By pointing out parallelism to the compiler directives get the com piler to carry out the details of mapping the computation onto the accel erator Using OpenACC directives requires a compiler that supports the Ope nACC standard In the following example where 7t is calculated adding the fpragma directive is sufficient for the compiler to produce code for the loop that can run on either the GPU or CPU Example include lt stdio h gt define N 1000000 int main void double pi 0 0f long i pragma acc parallel loop reduction pi for i20 i lt N i double t double i 0 5 N pi 4 0 1 0 t t printf pi 16 15f n pi N return 0 Bright Computing Inc Using MICs The hardware concept of the Intel MIC Many Integrated Cores architec ture is to bundle many x86 like chips into a processor currently imple mented in the Intel Xeon Phi release The MIC implementation is placed on a MIC card which can be hosted inside a node using the PCIe bus In this chapter the word MIC on its own implies the MIC architecture or implementation Bright Cluster Manager deals with MIC cards as if they are regu lar nodes If the Slurm workload manager is used then the MIC cards become compu
87. s than opera tions on triangular matrices Modern GPGPU implementations have matured so that they can now sub divide their resources between independent processes that work on independent data and they provide programmer friendlier ways of data transfer between the host and GPU memory Physically one GPU is typically a built in part of the motherboard of a node or a board in a node and consists of several hundred processing cores There are also dedicated standalone units commonly called GPU Units consisting of several GPUs in one chassis Several of these can be assigned to particular nodes typically via PCI Express connections to increase the density of parallelism even further Bright Cluster Managerhas several tools that can be used to set up and program GPUs for general purpose computations Bright Computing Inc Using GPUs 8 1 Packages A number of different GPU related packages are included in Bright Clus ter Manager For CUDA these are cuda50 driver Provides the GPU driver e cuda50 libs Provides the libraries that come with the driver libcuda etc cuda50 toolkit Provides the compilers cuda gdb and math libraries e cuda50 tools Provides the CUDA tools SDK cuda50 profiler Provides the CUDA visual profiler cuda50 sdk Provides additional tools development files and source examples CUDA versions 4 2 and 5 0 are also provided by Bright Cluster Man ager The exact implementation de
88. salloc mpirun hello salloc Granted job allocation 572 Hello world from process 0 out of 4 host name node001 Hello world from process 1 out of 4 host name node001 Hello world from process 2 out of 4 host name node001 Hello world from process 3 out of 4 host name node001 salloc Relinquishing job allocation 572 The preceding output also displays if N1 indicating 1 node is specified or if n4 indicating 4 tasks is specified The node and task allocation is almost certainly not going to be done by relying on defaults Instead node specifications are supplied to Slurm along with the executable To understand Slurm node specifications the following cases consider and explain where the node specification is valid and invalid Number of nodes requested The value assigned to the N nodes option is the number of nodes from the cluster that is requested for allo cation for the executable In the current cluster example it can only be 1 For a cluster with for example 1000 nodes it could be a number up to 1000 A resource allocation request for 2 nodes with the nodes option causes an error on the current 1 node cluster example fred bright52 salloc N2 mpirun hello salloc error Failed to allocate resources Node count specification in valid salloc Relinquishing job allocation 573 Number of tasks requested per cluster The value assigned to the nl ntasks option is the number of tasks that are requested for all
89. ser runs the ssh copy id command to copy the public key over to the cloud director instance using the usual password for authentication After this the user can carry out a passwordless login from the head node to the cloud director or the cloud node For example after logging into the head node using the standard login with a password a user galt can generate an ssh private and public certificate with ssh keygen as follows Example galt bright70 user galt already exists and has password galt bright70 ssh keygen Generating public private rsa key pair Enter file in which to save the key home galt ssh id rsa Enter passphrase empty for no passphrase Bright Computing Inc 4 8 Configuring Passwordless Login To Cloud Nodes 25 Enter same passphrase again Your identification has been saved in home galt ssh id rsa Your public key has been saved in home galt ssh id rsa pub The key fingerprint is 7a 31 d1 7 93 d5 96 bd 5a 6f al 11 b3 c2 d1 14 galt bright70 The key s randomart image is RSA 2048 o ouem S o ott o 0 o oO The user can then copy the public key over to the cloud director u west 1 director in the following example using ssh copy id Example galt bright70 ssh copy id eu west 1 director galt eu west 1 director s password Now try logging into the machine with ssh eu west l director and check in Ssh auth
90. specific variables The steps for running a job through SGE are to Create a job script Select the directives to use Add the scripts and applications and runtime parameters Submit it to the workload management system 6 1 Writing A Job Script A binary cannot be submitted directly to SGE a job script is needed for that A job script can contain various settings and variables to go with the application A job script format looks like bin bash S Script options Optional script directives shell commands Optional shell commands application Application itself 6 1 1 Directives It is possible to specify options directives to SGE by using in the script The difference in the meaning of lines that start with the character in the job script file should be noted Line Starts With Treated As 4 Comment in shell and SGE LES Comment in shell directive in SGE Comment in shell and SGE Bright Computing Inc 36 SGE 6 1 2 SGE Environment Variables Available environment variables SHO SUSER User ID of job owner S JOB ID Current job ID Dm m S JOB NAME Current job name like th Home directory on execution machine login and qalter SHOSTNAME Name of th xecution host STASK ID Array job task index number 6 1 3 Job Script Options N option in qsub qsh qrsh q Options can be set in the job script file using this line format
91. ss How Many Processes Are There Typically a process in a parallel application needs to know who it is its rank and how many other processes exist A process finds out its own rank by calling MPI Comm rank Int myrank MPI Comm rank MPI COMM WORLD amp myrank The total number of processes is returned by MPT Comm size int nprocs MPI Comm size MPI COMM WORLD amp nprocs A 5 Sending Messages A message is an array of elements of a given data type MPI supports all the basic data types and allows a more elaborate application to construct new data types at runtime A message is sent to a specific process and is marked by a tag integer value specified by the user Tags are used to dis tinguish between different message types a process might send receive In the sample code above the tag is used to distinguish between work and termination messages MPI Send buffer count datatype destination tag MPI COMM WORLD Bright Computing Inc A 6 Receiving Messages A 6 Receiving Messages A receiving process specifies the tag and the rank of the sending process MPI ANY TAGandMPI ANY SOURCE may be used optionally to receive a message of any tag and from any sending process MPI Recv buffer maxcount datatype source tag MPI COMM WORLD Information about the received message is returned in a status vari able The received message tag is status MPI TAG and the rank of the sending process is status MP
92. such copyright owner or third party has signed a writing to the contrary Table of Contents Table of Contents 000000 ce eee ee eee 01 About This Manual 2 N mn 0 2 Getting User Level Support 2 2 2220 Introduction 11 What Is A Beowulf Cluster 22222222200 11 1 Background And History 1 1 2 Brief Hardware And Software Description 12 Brief Network Description lees Cluster Usage 21 Login To The Cluster Environment 22 Setting Up The User Environment 23 Environment Modules ln 2 3 1 Available commands 2 2 2222 22020 23 2 Changing The Current Environment 2 3 3 Changing The Default Environment 24 Compiling Applications 22 22 2020 2 4 1 Open MPI And Mixing Compilers Using MPI 31 Interconnects Comm 311 Gigabit Ethernet 222 22 3 312 InfiniBand 2 2 2 2 0 00000 002 we 3 2 Selecting An MPI implementation 33 Example MPIRun 2 lee 3 3 1 Compiling And Preparing The Application 3 3 2 Creating A Machine File 3 3 3 Running The Application 33 4 Hybridization 2 222222 0 0 0 0 0000 3 3 5 Support Thread Levels 000 4 3 3 6 Further Recommendations Workload Management 41 What Is A Workload Manager 42 Why Use A Workload Manager 43 How D
93. t is by no means comprehensive in this area Readers are expected to have some familiarity with the basics of a Linux environment from the regular user point of view Regularly updated production versions of the 7 0 manuals are avail able on updated clusters by default at cm shared docs cm The latest updates are always online at http support brightcomputing com manuals The manuals constantly evolve to keep up with the development of the Bright Cluster Manager environment and the addition of new hard ware and or applications The manuals also regularly incorporate cus tomer feedback Administrator and user input is greatly valued at Bright Computing So any comments suggestions or corrections will be very gratefully accepted at manuals brightcomputing com 0 2 Getting User Level Support A user is first expected to refer to this manual or other supplementary site documentation when dealing with an issue If that is not enough to re solve the issue then support for an end user is typically provided by the cluster administrator who is often a unix or Linux system administrator with some cluster experience Commonly the administrator has config ured and tested the cluster beforehand and therefore has a good idea of its behavior and quirks The initial step when calling in outside help is thus often to call in the cluster administrator Introduction This manual is intended for cluster users who need a quick introduction to the Bri
94. table line in parallel the job launcher mpirun is placed immediately before the executable file The number of nodes the parallel job is to run on is assumed to have been specified in the PBS directives 7 1 2 Directives Job Script Directives And qsub Options A job script typically has several configurable values called job script di rectives set with job script directive lines These are lines that start with a 4 PBS Any directive lines beyond the first executable line are ignored The lines are comments as far as the shell is concerned because they start with a 4 However at the same time the lines are special com mands when the job script is processed by the qsub command The dif ference is illustrated by the following The following shell comment is only a comment for a job script pro cessed by qsub PBS Bright Computing Inc 7 1 Components Of A Job Script 45 The following shell comment is also a job script directive when pro cessed by qsub PBS Job script directive lines with the PBS part removed are the same as options applied to the qsub command so a look at the man pages of qsub describes the possible directives and how they are used If there is both a job script directive and a qsub command option set for the same item the qsub option takes precedence Since the job script file is a shell script the shell interpreter used can be changed to another shell interpreter by modifyi
95. tation which is quite standard Slurm documentation can be consulted http slurm schedmd com mpi guide html if the implementation the user is using is very different 5 1 Loading Slurm Modules And Compiling The Executable In section 3 3 3 an MPI Hello world executable that can run in parallel is created and run in parallel outside a workload manager The executable can be run in parallel using the Slurm workload man ager For this the Slurm module should first be loaded by the user on top of the chosen MPI implementation in this case Open MPI Example fred bright52 module list Currently Loaded Modulefiles 1 gcc 4 4 6 3 shared 2 openmpi gcc 64 1 4 2 4 cuda40 toolkit 4 0 17 fred bright52 module add slurm module list Bright Computing Inc Slurm Slurm Currently Loaded Modulefiles 1 gcc 4 4 6 3 shared 5 slurm 2 2 4 2 openmpi gcc 64 1 4 2 4 cuda40 toolkit 4 0 17 The hello world executable from section 3 3 3 can then be com piled and run for one task outside the workload manager on the local host as mpicc hello c o hello mpirun np 1 hello 5 2 Running The Executable With sa11oc Running it as a job managed by Slurm can be done interactively with the Slurm allocation command salloc as follows fred bright52 salloc mpirun hello Slurm is more typically run as a batch job section 5 3 However exe cution via salloc uses the same options and it is more convenient as an
96. te nodes the right environment modules for the job must be made available on the compute nodes too as part of the user login process to the compute nodes for that job Usually Bright Computing Inc 3 3 Example MPI Run 15 the system administrator takes care of such matters in the default user configuration by setting up the default user environment section 2 3 3 with reasonable initrm and initadd options Users are then typically allowed to set up their personal default overrides to the default adminis trator settings by placing their own initrmand initadd options to the module command according to their needs Running mpirun outside a workload manager When using mpirun manually outside a workload manager environment the number of pro cesses np as well as the number of hosts machinefile should be specified For example on a cluster with 2 compute nodes and a machine file as specified in section 3 3 2 Example fred bright70 module initclear module initadd openmpi gcc fred bright70 module add openmpi gcc dor as appropriate fred bright70 mpirun np 4 machinefile mpirun hosts hello Hello world from process 002 out of 004 processor name node002 cm Hello world from process 003 out of 004 processor name node001 cm Hello world from process 000 out of 004 processor name node002 cm Hello world from process 001 out of 004 processor name node001 cm The output of the preceding program is actuall
97. te nodes in Slurm because the Slurm compute daemon slurmd can be started inside a MIC card MIC offloading is supported by an appropriate generic resource in the workload managers supported by Bright Cluster Manager Both offloaded and native jobs are supported by Slurm This guide does not give details on how to write code for MIC or what tools and libraries should be used to do so The next two sections give an overview of the native and offload modes under which MIC can be used 9 1 Compiling Code In Native Mode The simplest way to run applications on the Intel Xeon Phi coprocessor is in native mode The native application can be compiled inside a co processor or on a host In the second case the binary can then be copied to the coprocessor and has to be started there Although the MIC and x86 64 architectures are very similar the MIC native application cannot be run on an x86 64 core The MIC assembly instruction set is highly but not completely x86 compatible and also has some additional scalar and vector instructions not found elsewhere Therefore to run a distributed MPI application on MICs and on hosts at the same time two versions of the application binary have to be built 9 1 1 Using The GNU Compiler Bright Cluster Manager provides a patched gcc which can be used to build a native MIC application However Intel recommends using the Intel Compiler section 9 1 2 which can create a more optimized k1om code for a better
98. ting Inc Cluster Usage The X option can be dropped if no X11 forwarding is required X11 forwarding allows a GUI application from the cluster to be displayed lo cally Optionally after logging in the password used can be changed using the passwd command passwd 2 2 Setting Up The User Environment By default each user uses the bash shell interpreter In that case each time a user login takes place a file named bashrc is executed to set up the shell environment for the user The shell and its environment can be customized to suit user preferences For example the prompt can be changed to indicate the current username host and directory for example by setting the prompt string variable PS1 u h w the size of the command history file can be increased for example export HISTSIZE 100 aliases can be added for frequently used command sequences for example alias lart ls alrt environment variables can be created or modified for example EXPORT MYVAR MY STRING the location of software packages and versions that are to be used by a user the path to a package can be set Because there is a huge choice of software packages and versions it can be hard to set up the right environment variables and paths for soft ware that is to be used Collisions between different versions of the same package and non matching dependencies on other packages must also be avoided To make setting
99. ued For example a user may have spec ified a job using the begin directive to start at 10am the next day by mistake To change the job to start at 10pm tonight something like the following session may take place fred bright52 scontrol show jobid 254 grep Time RunTime 00 00 04 TimeLimit UNLIMITED TimeMin N A SubmitTime 2011 10 18T17 41 34 EligibleTime 2011 10 19T10 00 00 StartTime 2011 10 18T17 44 15 EndTime Unknown SuspendTime None SecsPreSuspend 0 The parameter that should be changed is EligibleTime which can be done as follows fred bright52 scontrol update jobid 254 EligibleTime 2011 10 18T22 00 00 An approximate GUI Slurm equivalent to scontrol is the sview tool This allows the job to be viewed under its jobs tab and the job to be edited with a right click menu item It can also carry out many other functions including canceling a job Webbrowser accessible job viewing is possible from the workload tab ofthe User Portal section 10 2 Bright Computing Inc SGE Sun Grid Engine SGE is a workload management and job scheduling system first developed to manage computing resources by Sun Microsys tems SGE has both a graphical interface and command line tools for submitting monitoring modifying and deleting jobs SGE uses job scripts to submit and execute jobs Various settings can be put in the job script such as number of processors resource usage and application
100. ured and controlled by the head node and do only what they are told to do One of the main differences between Beowulf and a Cluster of Workstations COW is the fact that Beowulf behaves more like a single machine rather than many workstations In most cases the nodes do not have keyboards or monitors and are accessed only via remote login or possibly serial terminal Beowulf nodes can be thought of as a CPU memory package which can be plugged into the cluster just like a CPU or memory module can be plugged into a motherboard to form a larger and more powerful machine A significant difference is that the nodes of a cluster have a relatively slower interconnect 1 2 Brief Network Description A Beowulf Cluster consists of a login compile and job submission node called the head and one or more compute nodes often referred to as worker nodes A second fail over head node may be present in order to take control of the cluster in case the main head node fails Furthermore a second fast network may also have been installed for high performance low latency communication between the head and the nodes see fig ure 1 1 Bright Computing Inc 1 2 Brief Network Description Local network a Head node Failover node x Ethernet switch InfiniBand switch Figure 1 1 Cluster layout The login node is used to compile software to submit a parallel or batch program to a job queueing system and to gathe
101. utational power from cloud service providers such as the Ama zon Elastic Compute Cloud EC2 can be used by an appropriately con figured cluster managed by Bright Cluster Manager If the head node is running outside a cloud services provider and at least some of the compute nodes are in the cloud then this hybrid clus ter configuration is called a Cluster Extension cluster with the compute nodes in the cloud being the cloud extension of the cluster For a Cluster Extension cluster job scripts to a workload manager should be submitted using Bright Cluster Manager s cmsub utility This allows the job to be considered for running on the extension the cloud nodes Jobs that are to run on the local regular nodes not in a cloud are not dealt with by cmsub The environment module section 2 3 cmsub is typically configured by the system administrator to load by default on the head node It must be loaded for the cmsub utility to work The basic usage for cmsub is cmsub options script Options details are given in the man page for cmsub 1 Users that are used to running jobs as root should note that the root user cannot usefully run a job with cmsub The user can submit some cloud related values as options to cmsub on the command line followed by the job script Example cat myscriptl bin sh hostname cmsub regions eu west 1 myscriptl Upload job id 1 User job id 2 Download job id 3 Bright Computing I
102. vailable compilers However sometimes mix ing compilers is desirable For example C compilation may be preferred using icc from Intel while Fortran90 compilation may be preferred us ing openf90 from Open64 In such cases it is possible to override the default compiler path environment variable for example fred bright70 module list Currently Loaded Modulefiles 1 null 3 goco 4 4 7 5 sge 2011 11 2 shared 4 openmpi gcc 64 1 4 5 fred bright70 mpicc version showme mpif90 version showm gcc version gfortran version fred bright70 export OMPI CC icc export OMPI FC openf90 fred bright70 mpicc version showme mpif90 version showm icc version openf90 version Variables that may be set are OMPI CC OMPI FC OMPI_F77 and OMPI CXX More on overriding the Open MPI wrapper settings is doc umented in the man pages of mpicc in the environment section Bright Computing Inc Using MPI The Message Passing Interface MPI is a standardized and portable mes sage passing system designed by a group of researchers from academia and industry to function on a wide variety of parallel computers The standard defines the syntax and semantics of a core of library routines useful to a wide range of users writing portable message passing pro grams in Fortran or the C programming language MPI libraries allow the compilation of code so that it can be used over a variety of multi processor systems from SMP
103. val Hours 1 Figure 10 4 User Portal Graphs Page Selecting other values is possible for Workload Management Metrics The following workload manager metrics can be viewed RunningJobs QueuedJobs FailedJobs CompletedJobs EstimatedDelay AvgJobDuration AvgExpFactor Cluster Management Metrics The following metrics can be viewed OccupationRate NetworkBytesRecv NetworkBytesSent DevicesUp NodesUp TotalMemoryUsed TotalswapUsed PhaseLoad CPUCoresAvailable Bright Computing Inc 10 4 The GRAPHS Tab 75 GPUAvailable TotalCPUUser TotalCPUSystem TotalCPUldle Datapoints The number of points used for the graph can be speci fied The points are interpolated if necessary nterval Hours The period over which the data points are dis played The Update button must be clicked to display any changes made Bright Computing Inc A 1 Hello world MPI Examples A quick application to test the MPI libraries and the network Hello World Type MPI Test Program include lt mpi h gt include lt stdio h gt include lt string h gt define BUFSIZE 128 define TAG 0 int main int argc char xargv char idstr 32 char buff BUFSIZI int numprocs LH int myid Ent ub MPI Status stat all MPI programs start with MPI Init all MPI_Init amp argc amp argv N processes exist thereafter MPI Comm siz
104. y printed in random order This can be modified as follows so that only process 0 prints to the standard output and other processes communicate their output to process 0 include mpi h include string h include lt stdio h gt int main int argc char argv int numprocs myrank namelen i char processor name MPI MAX PROCESSOR NAME char greeting MPI MAX PROCESSOR NAME 80 MPI Status status MPI Init amp argc amp argv MPI Comm size MPI COMM WORLD amp numprocs MPI Comm rank MPI COMM WORLD amp myrank MPI Get processor name processor name amp namelen Sprintf greeting Hello world from process d of d on s myrank numprocs processor name if myrank 0 printf s n greeting for i 1 i lt numprocs itt MPI_Recv greeting sizeof greeting MPI_CHAR i 1 MPI COMM WORLD amp status printf s n greeting Bright Computing Inc cluster cluster cluster cluster Using MPI else MPI Send greeting strlen greeting 1 MPI CHAR 0 1 MPI COMM WORLD MPI Finalize return 0 fred bright70 module add mvapich gcc dor as appropriate fred bright70 mpirun np 4 machinefile mpirun hosts hello Hello world from process 0 of 4 on node001 cm cluster Hello world from process 1 of 4 on node002 cm cluster Hello world from process 2 of 4 on node001 cm cluster 3 Hello world from process of 4 on node002

Download Pdf Manuals

Related Search

Related Contents

Manuel d`Utilisation    User`s Manual  Samsung Samsung E3213 User Manual  Bilance a piattaforma  msms User Manual  Instructions de montage / Mode d`emploi  Guide de pose du Duomur  Bedienungsanleitung  ANL 00834514 Funksteckdosen Set.cdr  

Copyright © All rights reserved. Failed to retrieve file