Multicube Explorer User Manual Version 0.5
Contents
1. Arguments Options Description Shows the variable in the current m3explorer environment and reports basic infor mation about available Databases Example usage gt show_vars Shell Command name help Arguments 5 Options long request a list of the available options and arguments for each command Description Reports general help on m3explorer commands Example usage gt help long 8 2 Plugins Commands M3Explorer is organized in modular structure and enables the user to dynamically load precompiled plugins within the environment These plugins are DoE modules adopted for initial experimental design and optimization modules for the definition of opti mization algorithm to be adopted Shell variables can be used for passing additional parameters to the modules Shell Command name doe define doe Arguments Name of the Design Of Experiments to be loaded string Options 2 Description Instantiates the Design of Experiments module to be used during the optimization Example usage gt doe define doe m3 random doe Shell Command name doe show info Arguments E Options Description Displays info message from the DoE module currently loaded in the m3explorer en vironment if any If no doe module is defined then an error message is shown Example usage gt doe_show_info 25 Multicube Explorer User2. opt define optimizer m3 nsga II 14 Multicube Explorer User Manual Version 0 5 6 Interfaces for the integration of new simulators This section describe how it is possible to integrate M3Explorer with a system simulator and how to define to M3Explorer the design space to explore 6 1 Design Space Definition The definition of the design space is done by using an XML file that is composed of a preamble which defines the namespace and a version which in this document is 1 3 lt xml version 1 0 encoding UTF 8 gt lt design_space xmlns http www multicube eu version 1 3 gt lt simulator gt lt simulator gt lt parameters gt lt parameters gt lt system_metrics gt lt system_metrics gt lt rules gt lt rules gt lt design_space gt The remaining part of the file describes the simulator invocation method lt simulator gt lt simulator gt the set of pa rameters of the simulator which can be configured lt parameters gt lt parameters gt the system metrics which can be estimated by the simulator lt system_metrics gt lt system_metrics gt and the rules which have to be taken into account by M3Explorer in order to generate feasible configurations 6 1 1 Simulator Invocation The simulator executable marker is used for specifying the complete path name of the executable simulator simulator executable path path my simulator executable gt lt s
3. compiled with a standard GNU C compiler version 4 0 1 or higher The following libraries and programs are needed for correctly compiling the software The Bison YACC parser generator http www gnu org software bison bison html The Flex Lex lexical analyzer http flex sourceforge net The libxml2 library and libxml2 devel The latest versions of libxml2 can be found on the xmlsoft org server The Gnuplot software In most cases including common Linux distributions these tools are already available on the system 3 2 Installation Procedure M3Explorer is distributed in source form In order to be executed it must be compiled and installed into a standard directory Download the compressed file containing the release of M3Explorer tar gz The current release of the tool is 0 5 Uncompress the downloaded archive gt tar zxvf m3explorer release O 5 tgz This will create a directory m3explorer the complete path name to this directory will be referred to as sourcedir Create a build dir and run the configure command into it gt mkdir build gt cd build gt lt sourcedir gt configure image lt installdir gt where installdir is the final installation directory of the software If installdir is not specified the software will be installed in image Runmake and make install to finish the installation gt make gt make install Optional To delete all the temporary files generated during th
4. derived minimal size gt lt greater equal gt lt parameter name 12_cache_size gt lt constant value 2048 gt lt greater equal gt lt rule gt lt rules gt Describes the rule 12_cache_block_size gt 11_dcache_block_size AND 12_cache_size gt 2048 Available operators The following operators markers can be used lt greater gt lt greater equal gt lt less gt lt less equal gt lt equal gt lt not equal gt lt expr gt The lt expr gt marker can be used for introducing generic expressions e g rule lt greater equal gt lt parameter name 12_cache_size gt lt expr operator x gt lt constant value 2 gt lt parameter name 11_cache_size gt lt expr gt lt greater equal gt lt rule gt The previous set of rules is represents 12_cache_size gt 2 11_cache_size The operators supported by M3Explorer are Combining rules For combining complex expressions the following markers operators can be used lt and gt lt or gt lt not gt For example the following rules are AND ed together lt rules gt lt rule name overall memory subsystem integrity gt lt and gt lt greater equal gt lt parameter name 12 cache block size lt parameter name l11 dcache block size lt greater equal gt lt greater equal gt lt parameter name 12_cache_size gt lt constant value 2048 gt lt greater equal gt lt and gt lt rule gt
5. m n3 a 613 System Metrics Definition ea eroac ca x 614 Feasibility rules 42229 siria a XU 6 2 M3Explorer Simulator Interface o o 624 Simulator ipt ile oe vee eee ae eee ne o ee ee ees 823 Simulator Oa DUE PIS lt lt issue kde e ees 6 23 Simulator Error Management 7 Example of exploration with a simple simulator 8 Shell Command List Bi Basie Commands suo c memi trum eee eS eR lj ws ek de Sw a a A ok a Sak a fa GU PDC wet gegeae e eat ee kn E eel oe oe a ee ek eas el ne Voce a cele Bek ou oz bored ded qe EA ee eo Se eis Ce OE ee ae we Or FW W A W o coco NNN Multicube Explorer User Manual Version 0 5 8 2 8 3 show vars BB e Aw a ha uev Plugins Commands doe_define_doe doe_show_info drv_define_driver drv_show_info opt_define_optimizer opt_show_info opt tune 2 24 ses a Database Commands db change current db export db write db report co db_filter_pareto db plot objectives db show optimum db compute ADRS 9 Authors 10 Acknowledgments Multicube Explorer User Manual Version 0 5 1 Overview of Multicube Explorer Multicube explorer is an interactive program that lets the designer explore a design space of configurations for a param eterized architecture for which an executable model use case simulator exists Multicube explorer is an advanced multi objective optimization fram
6. or generated by means of the doxygen document production system use make doc to generate the guide in the installation image lt installdir gt Multicube Explorer User Manual Version 0 5 4 The Shell of Multicube Explorer To run M3Explorer you need to launch the following command gt lt installdir gt bin m3explorer a liga el e bless NJ M EDNA aN NZ A LUE NS INM ___ _ NL LLLI Multicube Explorer Version release O 5 Send bug reports to zaccariaQelet polimi it gpalermo elet polimi it m3 shell and the M3Explorer Shell for the user interaction starts To exit the program just type exit followed by a return Typing the help command all the available M3Explorer commands with short descriptions are shown m3 shell help db change current change the current database db export exports the db into a csv file db filter pareto filters the current database for pareto points db insert point insert a point in the current database db read reads a database from disk db report reports the contents of a database db show optimum db write doe define doe drv define driver drv show info exit help opt define optimizer opt show info opt tune quit read script set show_vars m3 shell Shows the optimal point of a database writes a database on disk define the doe module define the driver module shows information about current driver quit the current m3explorer session general help on m3expl
7. simulator_input_interface xmlns http ww multicube eu version 1 3 gt lt parameter name seed value 1 gt lt simulator_input_interface gt The number of lt parameter gt sections and the name of the parameters should be the same as defined in the XML Design Space description file The value of the each parameter section should correspond to one of the possible values as defined in the XML Design Space description file Concerning variable vector parameters the actual parameter instances are specified with an itemized list For example an on off mask instance value for the active processors parameter is described in the sim ulator input file as the following list 19 Multicube Explorer User Manual Version 0 5 lt parameter name active processors gt lt item index 1 value 0 gt lt item index 2 value 1 gt lt item index 3 value 0 gt lt item index 4 value 1 gt lt parameter gt In the case of a permutation vector the index attribute is substituted with the position attribute lt parameter name thread_assignment gt lt item position 1 value 2 gt lt item position 2 value 3 gt lt item position 3 value 1 gt lt parameter gt Index and position attributes start from 1 up to the dimension associated to the variable vector 6 2 2 Simulator Output File The simulator output file contains a preamble and a sequence of lt system_metric gt sections where for ea
8. small big value of a specific system metric This attribute is optional and may be taken into account by M3Explorer in the formalization of the optimization problem which however it is not part of this specification M3Explorer expects to find the system metrics defined in this section in the output file ofthe simulator The output file name of the simulator is the second argument passed to the simulator executable file 6 1 4 Feasibility rules The lt rules gt section is used by M3Explorer in order to not generate invalid or not feasible solutions during the automated exploration process The behavior of the simulator when these rules are not met is undefined Each rule is a boolean expres sion which should evaluate to true for a feasible configuration of the design space It is up to M3Explorer tool to check for the rules and generate feasible configurations Each boolean expression can be an operator acting on either a parameter or constant leafs or other boolean expressions This allows creating complex expression trees of rules Rules are AND ed by default by M3Explorer Each rule is identified by a lt rule gt marker and it has an optional name attribute As an example rules rule lt greater equal gt lt parameter name 12_cache_block_size gt 17 Multicube Explorer User Manual Version 0 5 lt parameter name l11 dcache block size lt greater equal gt lt rule gt lt rule name application
9. whatever dimension of the objective space thus it can be used even when graphical investigation towards db plot objectives cannot be performed ADRS computation needs a reference set that is provided to the command by passing a db name parameters where the reference set is stored ADRS result is stored in a variable called last_ADRS Example usage gt set objectives energy cycles performing a full search exploration doe define doe m3 full doe db change current full DB opt define optimizer m3 pareto doe opt tune VV VV Vv performing a APRS exploration db_change_current aprs_DB doe define doe m3 random doe opt define optimizer m3 aprs opt tune V VM MM Vv computing aprs quality db compute adrs full DB Vv 29 Multicube Explorer User Manual Version 0 5 9 Authors Vittorio Zaccaria Politecnico di Milano Gianluca Palermo Politecnico di Milano Giovanni Mariani ALaRI Universit della Svizzera italiana 10 Acknowledgments Much of what M3Explorer is today was also defined by the users of the tool We would like to acknowledge the contributions of the following people for their early adoption ofthe tool their feedback on the tool and on the interfaces their contributions to the tool and their comments on the manual Cristina Silvano Politecnico di Milano William Fornaciari Politecnico di Milano Alessandro Sivieri Politecnico di Milano Al Hi
10. Manual Version 0 5 Shell Command name drv define driver Arguments Name of the driver to be loaded string Options Description Load a driver module into the m3explorer environment On doing so design space is defined and parameters representing design parameter boundaries are organized into the m3explorer environment as shell variables Example usage gt set xml design space file m3 mpeg use case xml gt drv define driver m3 xml driver gt show vars Shell variables Name Value cbs 16 128 dcs 2048 65536 dcw 1 16 ds parameters ics pn ics 2048 65536 icw 1 16 iwidth 1 8 12cs 32768 1048576 12cw 1 16 pn 1 16 Databases in memory Name size root available O Shell Command name drv show info Arguments Options Description Shows information about the current driver loaded Example usage gt drv show info Shell Command name opt define optimizer Arguments Name of the optimizer to be loaded string Options E Description Loads the optimizer plug in Example usage gt opt_define_optimizer m3_aprs Shell Command name opt_show_info Arguments Options Description Shows information about the current optimizer Example usage gt opt_show_info 26 Multicube Exp
11. Multicube Explorer User Manual Version 0 5 Politecnico di Milano Italy and Universit della Svizzera italiana Switzerland January 12 2009 Multicube Explorer User Manual Version 0 5 Contents 1 Overview of Multicube Explorer 11 GoalsofMulticubeExplorer llle 11 Automatic design space exploration Li POI a s Ego RR NEL ue 1 1 3 Modular composition 4 ose o ee ok mo 12 Jnhitectureofthet ol 0 0 4 bb oor Rr eee ee i 13 Interaction with the simulator eee eee eee 2 License 3 Installation Requirements and Procedure 31 Installation Requirements o 2 3 Installation Procedure o A AS eS 33 Testingthe installation 34 Uninstall procedure loce rh rm Rr iw eee 35 DOcwmentaon osc bce kb Rd go kem ER y Ew Re UR s 4 The Shell of Multicube Explorer 5 Available Plugins nl Ge uc eleme exe ea E RUPEE RR SO O E A ues AY BOR Wl a SUR Tondon uod cuim Gk a PS REG 6 X 513 TwolevelsFullFactorial S OODUMEUEOES ina ii da Eq 5 41 Pareto DoE 22999 a om ox dre 3 24 APRS uu QAO 4 X XA XR op e URS UE AN S Ban MODE ga aa ae A AR A e ea Aulo MOPSO oor aaie dq Re Rex EO e oS os rw BAe dog Dog IO sai AR RR 6 Interfaces for the integration of new simulators amp l Design space Definifiofi b ee o m dne eee n 6 1 1 Simulator Invocation os co sooo o o ro gn 6 12 ParametersDefinition o 0 o
12. Redistributions in binary form must reproduce the above copyright notice this list of conditions and the following disclaimer in the documentation and or other materials provided with the distribution Neither the name of the Politecnico di Milano and Universit della Svizzera italiana nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS AS IS AND ANY EXPRESS OR IMPLIED WARRANTIES INCLUDING BUT NOT LIMITED TO THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT INDIRECT INCIDENTAL SPECIAL EXEMPLARY OR CONSEQUENTIAL DAMAGES INCLUDING BUT NOT LIMITED TO PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES LOSS OF USE DATA OR PROFITS OR BUSINESS INTERRUPTION HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY WHETHER IN CONTRACT STRICT LIABILITY OR TORT INCLUD ING NEGLIGENCE OR OTHERWISE ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE Multicube Explorer User Manual Version 0 5 3 3 1 Installation Requirements and Procedure Installation Requirements Multicube Explorer m3explorer in short has been designed to be compatible with LINUX and BSD MAC OSX platforms M3Explorer is written in C and can be
13. ar structure of M3Explorer Database commands used to handle one or more database where visited design points are stored 8 1 Basic Commands These commands support the user on performing simple operations as load scripts display help messages and set variables value and close m3explorer Shell Command name exit Arguments n Options Description Exits from the m3explorer shell Example usage gt exit Shell Command name quit Arguments Options Description Same as exit command Example usage gt quit Shell Command name read_script Arguments Name of the script to be read string Options Description All commands within the given script are sequentially executed in the m3explorer environment Example usage gt read_script my_script scr Shell Command name set Arguments Name of the variable to be set identifier Value to be inserted into the variable object Options 2 Description Set an environment variable Set command cares about definition of the variable if this was never used before or modification of variable content if variable was exist ing Example usage gt set my_string I m a string gt set my_number 13 gt set my_list my_string my_number 0 9233 24 Multicube Explorer User Manual Version 0 5 Shell Command name show vars
14. cessor system Permutation Variable size permutations are used for example in the case of thread to processor mapping problems In this case a task identifier should be generated for each active processor lt parameter type permutation name thread assignment dimension Onumber of threads A permutation contains a non repeatable sequence of values from 1 to the actual dimension of the vector For example the variable vector parameter lt parameter type permutation name example dimension 2 P yp P P can assume the following values 1 2 or 2 1 6 1 3 System Metrics Definition The lt system_metrics gt section is used by the use case and simulator provider to specify the names the types and the units of the system metrics that can be estimated by the simulator lt system_metrics gt lt system_metric name cycles type integer unit cycles desired small gt system metric name instructions type integer unit insts description gt system metric name powerconsumption type float unit W description system metric name area type float unit mm2 desired small gt lt system_metrics gt Feasible system metric name attributes are identified by the following regular expression A Za z A Za z0 9_ The optional description attribute is a generic string describing the nature of the system metric The desired attribute indicates whether it is desirable to have a
15. ch metric the name and the value is specified lt xml version 1 0 encoding UTF 8 gt lt simulator_output_interface xmlns http www multicube eu version 1 3 gt lt system_metric name cycles value 3000 gt lt system_metric name instructions value 1500 gt lt system_metric name power _consumption value 2 5 gt lt system_metric name area value 25 gt lt simulator_output_interface gt The number of lt system_metric gt sections and the name of the system metrics should be the same as defined in the XML Design Space description file 6 2 3 Simulator Error Management In the case of errors during the simulator execution the simulator output file should contain a single lt error gt marker indi cating the error reason lt xml version 1 0 encoding UTF 8 gt lt simulator_output_interface xmlns http www multicube eu version 1 3 gt lt error reason memory full kind fatal gt lt simulator_output_interface gt The attribute reason is a generic string that can contain a report about the error cause Overall the error strings of the simulator are meant to be related to memory full or disk full problems file system permissions problems license problems internal exceptions other consistency or feasibility violation if checked by the simulator The kind can be fatal non fatal Fatal errors should block the overall exploration process while non fatal err
16. db change current Arguments Name of the db to be set as current string Options 5 Description The current db is the one actually used for storing exploration results and it is the target of the commands of the shell All operations performed in m3explorer which act on database has the current db as target db_change_current allow to define which database have to be used as current taking its name as parameter In the case no database with the specified name exists in the m3eplorer environment such database is automatically generated and set as current Example usage gt db change current MY EXPLORATION DB Shell Command name db export Arguments File name into which the current database should be exported string Options Description This command can be used to export the current database in CSV format to the file specified as parameter Example usage gt db export MY EXPLORATION DB csv Shell Command name db write Arguments File name into which the current database should be written string Options Description Writes the current db on the disk to a specified file in a format readable from m3explorer Example usage gt db write MY EXPLORATION DB db 27 Multicube Explorer User Manual Version 0 5 Shell Command name db report Arguments Options Description Reports all the architectural confi
17. e Start the exploration m3 shell opt tune Write the results of the exploration in an internal M3Explorer db format and in a standard csv format m3 shell db write my full db m3 shell db write my full csv Set up the objective functions of the problem to enable Pareto filtering of visited points perform Pareto filtering of the current database by eliminating dominated points and report the results m3 shell set objectives sum difference product m3 shell db filter pareto m3 shell db report Write the Pareto points in internal and csv format m3 shell db write my full db m3 shell db write my full csv Exit from the shell m3 shell exit The example can be automated by using an M3Explorer script simple sim scr scr gt lt installdir gt bin m3explorer x simple sim ds xml f simple sim scr scr Figure 5 shows the output of such script 22 Multicube Explorer User Manual Version 0 5 bin m3explorer x simple sim ds xml f simple sim scr scr Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information Information In
18. e compilation use the following command gt make dist clean This command deletes everything that has been built with configure and make Multicube Explorer User Manual Version 0 5 3 3 Testing the installation To test that the installation has been performed correctly first run the application executable m3explorer which is present in the installation dir gt lt installdir gt bin m3explorer Multicube Explorer Version release_0_5 Send bug reports to zaccariaQelet polimi it gpalermo elet polimi it m3 shell To exit the program just type exit followed by a return A more comprehensive test can be run with the do test script which is present in the test directory of the installation image lt installdir gt A succesfull execution produces the following output installdir tests do tests 1 Tested XML design space construction PASSED 2 Tested XML rules PASSED 3 Tested XML error reporting PASSED 4 Tested XML input output interface PASSED 5 Tested full factorial features PASSED 3 4 Uninstall procedure If you installed m3explorer into a dedicated directory it can be removed by simply deleting the directory Otherwise the specific executable and the associated installation files should be removed manually 3 5 Documentation The documentation of M3Explorer is mainly composed by two documents The present User Guide The Developer Guide This guide can be browsed either on the M3Explorer website
19. e functions The term simulated annealing derives from the analogy with the physical process of heating and then slowly cooling a substance to obtain a strong crystalline structure In the Simulated Annealing algorithm a new configuration is constructed by imposing a random displacement If the cost function of this new state is less than the previous one the change is accepted unconditionally and the system is updated If the cost function is greater the new configuration is accepted probabilistically the acceptance possibility decreases with the temperature optimization time This procedure allows the system to move consistently towards lower cost function states thus jumping out of local minima due to the probabilistic acceptance of some upward moves This optimizers implemented in M3Explorer is called Multi Objective Simulated Annealing MOSA and it is derived by Smith K I Everson R M Fieldsend J E Murphy C Misra R Dominance Based Multiobjective Simulated Annealing IEEE Transaction on Evolutionary Computation 12 3 323 342 2008 m3 shell opt define optimizer m3 mosa 5 2 4 MOPSO More in general Particle Swarm Optimization PSO is a heuristic search methodology that tries to mimic the movements of a flock of birds aiming at finding food PSO is based on a population of particles flying through an hyper dimensional search space Each particle possesses a position and a velocity both variables are changed to emulat
20. e the social psychological tendency to mimic the success of other individuals in the population also called swarm This optimizers implemented in M3Explorer is called Multi Objective Particle Swarm Optimization MOPSO and it is derived by G Palermo C Silvano V Zaccaria Discrete Particle Swarm Optimization for Multi objective Design Space Exploration In Euromicro Proceedings of DSD 08 Conference on Digital System Design September 2008 m3 shell opt define optimizer m3 mopso 13 Multicube Explorer User Manual Version 0 5 5 2 5 NSGA II In a Genetic Algorithm many design alternatives belonging to design space are seen like individuals in a stored population The exploration procedure consists of the simulation of the evolution process of generation of individuals and the improve ment of solutions belonging to next generations is explained by Darwinian theory The evolutionary operators describe how individuals are selected to reproduce how a new generation of individuals is generated from parents by crossover and mutation and how new generation of individuals is inserted into population replacing or not the parents The implemented approach for Multiobjective optimization is the non dominated sorting genetic algorithm NSGA II de scribed in Kalyanmoy Deb Amrit Pratap Sameer Agarwal T Meyarivan A fast and elitist multiobjective genetic algorithm NSGA II IEEE Transactions on Evolutionary Computation 2002 m3 shell
21. ework which is entirely command line script driven and can be retargeted to any configurable platform by writing a suitable XML design space definition file and providing a configurable simulator Multicube Explorer is supported by the EC under grant FP7 216693 MULTICUBE http www multicube eu The tool and the documentation can be currently found at the following address http home dei polimi it zaccaria multicube explorer gt Response Surface Design Modeling y of Experiments M3Explorer Shell 3 3 gt S E g o Optimization x Algorithm go Architecture o z database XML XML Design Space Simulator NERY Figure 1 Structure of Multicube Explorer 1 1 Goals of Multicube Explorer The overall goal of the open source design space exploration framework aims at providing a retargetable tool to drive the designer towards near optimal solutions to the architectural exploration problem with the given multiple constraints The final product of the framework is a Pareto curve of configurations within the design evaluation space of the given architec ture 1 1 1 Automatic design space exploration One of the goals of the open source tool is to provide a command line interface to the exploration kernel that allows the con struction of automated exploration strategies Those strategies are implemented by means of command scripts interpreted by the tool without the need o
22. f manual intervention This structure can easily support the batch execution of complex strategies that are less prone to human intervention due to their execution time 1 1 2 Portability Another goal of the open source tool is to be portable across a wide range of systems This goal will be achieved by not sacrificing the efficiency of the overall exploration engine The standard ANSI C programming language will be used for developing the open source framework The Standard Template Library as well as other open source libraries will be used during the development process Multicube Explorer User Manual Version 0 5 1 1 3 Modular composition One of the strength of the open source tool is the modularity of its components Simulator optimization algorithms and other design space exploration components are dynamically linked at run time without the need of recompiling the entire code base This will be supported by well defined interfaces between the drivers supporting the simulation and the optimization algorithm This will strongly enable the introduction of new modules for both academic and industrial purposes Given the modular decomposition a single optimization algorithm can used for every use case simulator Moreover a single use case architecture can be optimized with a wide range of optimization algorithms 1 2 Architecture of the tool The tool Figure 2 is basically composed by an exploration kernel which orchestrates the functi
23. formation Creating the xml driver Loading the xml driver Assigned value true to clean directory on exit Changing current DB to full db Database not existing Creating a new one Loading the full Current doe has been set to Current optimizer has been set to search doe Full search doe Pareto doe optimizer Starting with the pareto doe optimization process Evaluating point Evaluating point Evaluating point Skipping point Skipping point Skipping point Evaluating point Evaluating point Evaluating point Evaluating point Evaluating point Evaluating point Evaluating point Evaluating point Evaluating point Evaluating point Evaluating point Evaluating point 5 parl exp2 E parl exp2 2048 par2 ste H parl exp2 4 024 par2 ste 096 par2 ste p p p 1 par3 step2 1 1 par3 step2 1 1 par3 step2 1 parl exp2 1024 par2 stepl 2 par3 step2 1 8 par2 stepl 2 par3 step2 1 parl exp2 4096 par2 stepl 2 par3 step2 1 parl exp2 204 parl exp2 parl exp2 4 parl exp2 parl exp2 4 parl exp2 parl exp2 4 parl exp2 parl exp2 4 Writing the database to disk Database correctly written Saving the database in CSV format Assigned value sum difference 024 par2 ste parl exp2 2048 par2 ste 096 par2 ste 024 par2 ste parl exp2 2048 par2 ste 096 par2 ste 024 par2 step parl exp2 2048 par2 ste 096 par2 ste 024 par2 ste parl exp2 2048
24. ger unit mm2 desired small gt system metric name product type integer unit mW desired small gt system metrics rules rule lt greater equal gt lt parameter name par3_step2 gt lt parameter name par2_step1 gt lt greater equal gt lt rule gt lt rules gt lt design_space gt Figure 4 simple_sim_ds xml In this example we perform a full search exploration of the design space shown in Figure 4 by filtering the final results for the pareto set To start with the exploration we invoke m3explorer with its target design space gt lt installdir gt bin m3explorer x simple_sim_ds xml The target design space is now loaded Now m3explorer knows where the simulator is and which are the parameters asso ciated with it Once in the m3explorer shell we perform the following steps Configure the optimizer to clean the directory at the end of the exploration m3 shell set clean directory on exit true Change the current database to a new database called fu11 db it will be filled by optimizer module with the explo ration results 21 Multicube Explorer User Manual Version 0 5 m3 shell db change current full db Load the m3 full doe DoE Full search considers all the possible combination of the parameters m3 shell doe define doe m3 full doe Load m3 pareto doe optimizer it visits only the solutions defined by the DoE m3 shell opt define optimizer m3 pareto do
25. gurations stored into the current db Example usage gt db_report Shell Command name db_filter_pareto Arguments Options Description Filter the current database keeping only the Pareto points The Pareto concept is defined once a special variable with the name objectives is declared in the environ ment as a list of objective metrics Commands in the example filter the database my_exploration keeping only design points such as no other point is better in terms of energy and delay Example usage gt db_change_current my_exploration gt set objectives energy delay gt db_filter_pareto Shell Command name db_plot_objectives Arguments Databases to be plotted Options 5 Description Graphical investigation of solutions quality Plotting functionalities are available only in 2D thus to use db_plot_objectives a special variable with the name ob jectives must be declared as a list of two objective metrics If no parameters are passed the current database only is investigated and objective metrics of stored points are plotted If one database name parameter is passed the graphical investigation is performed over such db In the case two database name parameters are passed graphical investigation is performed on the same plot for the two databases allowing graphical comparison of solution qualities Example usage gt set objectives energy dela
26. imulator gt The path is specified by using Unix conventions The simulator executable is invoked with three arguments my simulator executable xml system configuration sc path name xml system metrics sm path name reference xsd xsd file name where sc path name is the path name of XML file describing the system configuration to be passed to the simulator The sm path name is the path name of the output XML file which should be used by the simulator for producing the sys tem metrics output The argument reference xsd xsd file name is used for specifying the position of the reference M3Explorer simulator interface XSD file in the file system This argument can be used by the simulator for validating the input and output files exchanged with M3Explorer 6 1 2 Parameters Definition The parameters lt parameters gt is used by the use case and simulator provider to specify the names the types and the ranges of the parameters that can be explored by the DSE tool The section contains a list of parameter markers parameter lt parameter name seed description RNG seed type integer min 0 max 10 gt lt parameter name fetch queue size description instruction fetch queue size type integer min 1 max 8 step 2 gt 15 Multicube Explorer User Manual Version 0 5 lt parameters gt For each parameter a unique name must be provided This name will be used for generating configurations at the i
27. lorer User Manual Version 0 5 Shell Command name opt tune Arguments Options 2 Description Launch the exploration process defined in the optimizer also called optimization or tuning operation During optimization all modules optimizer driver doe inter acts thus they should be all loaded into the environment when opt tune is called With commands in the following example the aprs optimization process is launched over the mpeg usecase available in the example subdirectory of the installation di rectory The DoE adopted as initial experimental design is a random one Results of optimization process are placed into the current database root database when the shell is just opened Example usage gt set xml design space file m3 mpeg use case xml gt drv define driver m3 xml driver gt doe define doe m3 random doe gt opt define optimizer m3 aprs gt opt tune 8 3 Database Commands Simulation results are stored into database db in memory Databases can be loaded stored from to the hard disk in such a way that is easy to handle simulation data obtained in different working sessions Operations that can be performed on databases are various and allow to extrapolate and visualize some fundamental high levelinformation needed to investigate solution quality of multiobjective optimization Following are m3explorer commands for database handling Shell Command name
28. lt rules gt This corresponds to the following expression 12_cache_block_size gt l1 dcache block size AND 12_cache_size gt 2048 18 Multicube Explorer User Manual Version 0 5 If then else rule An if E then A predicate is introduced and it is evaluated as TRUE if E is FALSE AifEis TRUE An example for this rule is the following lt rule name branch prediction design space reduction gt if lt not equal gt lt parameter name bpred gt lt constant value bimod gt lt not equal gt lt then gt lt equal gt lt parameter name bpred_bmod_size gt lt constant value 0 gt lt equal gt lt then gt lt if gt lt rule gt This associated predicate expression is if bpred bmod then bpred_bmod_size 0 This rule forces to generate configurations where if bpred bmod then bpred_bmod_size 0 These rules can effectively re duce the overall design space An if E then A else B predicate is introduced and it is evaluated as BifEis FALSE AifEis TRUE 6 2 M3Explorer Simulator Interface The M3Explorer Simulator interface is composed by 2 files one in output from M3Explorer to the simulator the other one in the opposite direction 6 2 1 Simulator input file The simulator input file should contain a preamble and a sequence of lt parameter gt sections where for each parameter the name and the value is specified lt xml version 1 0 encoding UTF 8 gt lt
29. nput of the simulator Feasible parameter names are identified by the following regular expression A Za z A Za zo 9 The parameters types can be divided into two categories Scalar types Variable vector types Scalar parameter types The scalar parameter type can be integer and boolean The integer type specifies a simple sequential integer progression associated that specific parameter The min and max attributes which are mandatory specify the boundaries of the progression The step attribute can be used to produce non unitary progressions The Boolean type is an integer progression with min 0 and max 1 exp2 The values associated with an exp2 parameter type should be computed by M3Explorer by using a power of two progression For example lt parameter name ili1 cache block size bytes description type exp2 min 8 max 64 gt should be interpreted by M3Explorer as a parameter with range values 8 16 32 64 string Inthe case of string parameters a list of possible string values shouldbe used instead of the min max attributes lt parameter name bpred description branch predictor type type string gt lt item value nottaken gt lt item value taken gt lt item value perfect gt lt item value bimod gt lt item value 2lev gt lt item value comb gt lt parameter gt Variable vector types The following types are introduced for producing variable vector type
30. o identify the planning of an information gathering experimentation cam paign where a set of variable parameters can be tuned The reason for DoEs is that very often the designer is interested in the effects of some parameter s tuning on the system response Design of experiments is a discipline that has very broad appli cation across natural and social sciences and encompasses a set of techniques whose main goal is the screening and analysis ofthe system behavior with a small number of simulations Each DoE plan differs in terms ofthe layout of the selected design points in the design space The available DoEs in the M3Explorer framework are Full Search Random Two Levels Full Factorial 5 1 1 Full Search It is the simplest DoE in the discrete world It consider all the possible configuration of the design space m3 shell doe define doe m3 full doe 5 1 2 Random The design space configurations are picked up randomly by following a Probability Density Function PDF In the imple mented plugin uses a uniformly distributed PDF m3 shell doe define doe m3 random doe The variable solutions number can be used to define the number of points of the random DoE e g for a random doe with 15 points m3 shell set solutions number 15 5 1 3 Two levels Full Factorial In statistics a factorial experiment is an experiment whose design consists of two or more parameters each with discrete possible values or levels and wh
31. onal behavior of the design of experiments and optimization algorithms M3explorer XML DoE Lin al Design i ERN Space Plug in im Definition Core Design Space Representation Optimizer Use Plug in an Case Simulator Figure 2 Architecture of the tool The kernel module is responsible for reading in the design space definition file in XML format and accepting commands from the shell interface or the corresponding script It then exposes the parameters of the design space to all the modules involved in the optimization process DoE Optimization Algorithms by means of a core design space representation The core design space representation provides a set of abstract operations that are mapped on the specific use case under analysis The abstract operations are represented by iterators over the feasible design space among which we can find Full search iterators Random search iterators global and neighborhood Factorial iterators two level two level center point The core design space representation provides also services for validating architectural choices at the optimizer level and evaluating the associated objective functions The objective functions are defined as a subset of the use case system level metrics and can be manipulated by the user by interacting with M3explorer Multicube Explorer User Manual Version 0 5 1 3 Interaction with the simulator The design space exploration is pe
32. orer commands define the optimizer module shows information about current optimizer Start the exploration process quit the current m3explorer session read script from file set a variable to a specific value Shows the variables in the current shell The design space exploration problem within M3Explorer is defined by a driver An XML driver m3 xml driver is provided in M3Explorer Such driver allow the integration of M3Explorer with other performance estimation tools by exploiting well defined XML interfaces that will be described later in section 6 Other two drivers m3 dtlz driver and m3 test driver are distributed in the purpose of methodological test of new design space exploration techniques To load the description of the design space for the following exploration process M3Explorer should be lauched with the x design space file xml flag 10 Multicube Explorer User Manual Version 0 5 E g lt installdir gt bin m3explorer x simple sim ds xml The use of M3Explorer can be both in interaction mode through the shell or in a script mode This second mode can be enabled by writing all the M3Explorer commands into a script file and then launching M3Explorer with the f M3Explorer commands file scr flag lt installdir gt bin m3explorer x simple sim ds xml f simple sim scr scr 11 Multicube Explorer User Manual Version 0 5 5 Available Plugins 5 1 DOoEs The term Design of Experiments DoE is used t
33. ors force M3Explorer to skip to the next configuration If an error marker is present in the output file system metric markers are ignored by M3Explorer 20 Multicube Explorer User Manual Version 0 5 7 Example of exploration with a simple simulator In this section we report a simple example usage of m3explorer The example consists of the exploration of the parameter space of a simple simulator The files associated with this example can be located in lt installdir gt examples simple_sim directory namely they correspond to simple sim py Python script representing the simulator of the target architecture to be explored simple sim ds xml Design space to be explored see Figure 4 simple sim scr scr m3explorer script file which automates the steps of the exploration lt xml version 1 0 encoding UTF 8 gt lt design_space xmlns http www multicube eu version 1 3 gt lt simulator gt lt simulator_executable path usr bin python image examples simple_sim simple_sim py gt lt simulator gt lt parameters gt lt parameter name pari_exp2 type exp2 min 1024 max 4096 gt lt parameter name par2_stepi type integer min 1 max 2 step 1 gt lt parameter name par3_step2 type integer min 1 max 5 step 2 gt lt parameters gt lt system_metrics gt lt system_metric name sum type integer unit cycles desired small gt lt system_metric name difference type inte
34. ose experimental units take on all possible combinations of these levels across all such pa rameters Such an experiment allows studying the effects of each parameter on the response variable as well as the effects of interactions between parameters on the response variable In this plugin we consider a 2 level full factorial DoE where the only levels considered are the minimum and maximum for each parameter m3 shell doe define doe m3 two level ff 5 2 Optimizers The available optimizer plugins in M3Explorer are Pareto DoE 12 Multicube Explorer User Manual Version 0 5 APRS Adaptive windows Pareto Random Search MOSA Multi Objective Simulated Annealing MOPSO Multi Objective Particle Swarm Optimizer NSGA II Non dominated Sorting Genetic Algorithm 5 2 1 Pareto DoE This is not a real optimizer but it is only a simple method used to evaluate the point selected by the DoE m3 shell opt define optimizer m3 pareto doe 5 2 2 APRS This algorithm called Adaptive windows Pareto Random Search is an algorithm that has a dynamic windows size which are reduced with the time spent in the exploration and with the goodness of the point found in the current windows The win dows are centered on the current pareto solutions and the new configurations are randomly selected within the windows m3 shell opt define optimizer m3 aprs 5 2 3 MOSA Simulated annealing is a Monte Carlo approach for minimizing multivariat
35. par2 ste 096 par2 ste Filtering the database for pareto points p O io io Mo uo p p p p p product 1 par3 step2 3 1 par3 step2 3 1 par3 step2 3 2 par3 step2 3 2 par3 step2 3 2 par3 step2 3 1 par3 step2 5 1 par3 step2 5 1 par3 step2 5 2 par3 step2 5 2 par3 step2 5 2 par3 step2 5 to objectives full db Current parl exp2 1024 parl exp2 1024 parl exp2 1024 parl exp2 1024 parl exp2 1024 Number of points Information Information Information Information Information Figure 5 M3Explorer output while it is running simple sim scr scr database contents par2 stepl 1 par2 stepl 1 par2 stepl 2 par2 stepl 1 par2 stepl 2 in the DB 5 par3 step2 1 par3 step2 3 par3 step2 3 par3 step2 5 par3 step2 5 Writing the database to disk Database correctly written Saving the database in CSV format Removing xml driver Exiting from Multicube Explorer shell 23 1026 1022 1024 1028 1020 3072 1029 1019 6144 1030 1018 5120 1031 1017 10240 Multicube Explorer User Manual 8 Shell Command List This section reports a brief survey of commands available in the Multicube Explorer optimization tool The commands are organized in three main classes Basic commands adopted to control the m3explorer environment in terms of defined variables and other basic func tionalities Plugins commands which give powerful flexibility to the user on handling the modul
36. rformed by using the simulation abstraction layer exported by the XML driver to the optimizer plug ins In principle the optimizer instantiates a set of architectural configurations by means of the design space iterators and passes the corresponding representation to the XML driver which will execute the simulator see Figure 3 Information about simulator runs will be displayed directly on the M3explorer shell XML System Metrics Use Case Simulator Multicube Explorer XML System Configuration Figure 3 Interaction of Multicube Explorer with the Use Case Simulator M3Explorer creates a specific directory to execute each instance ofthe simulator In this directory a valid system parameters file is created before starting the simulator A system metrics file is expected to be obtained as the output of the simulator execution Multicube Explorer User Manual Version 0 5 2 License Multicube explorer is open source and it is released under the BSD license Authors Vittorio Zaccaria Gianluca Palermo Giovanni Mariani Copyright c 2008 Politecnico di Milano and Universit della Svizzera italiana All rights reserved Redistribution and use in source and binary forms with or without modification are permitted provided that the follow ing conditions are met Redistributions of source code must retain the above copyright notice this list of conditions and the following dis claimer
37. s with specific constraints on the possible combinations of the components on off mask The on off mask is essentially a vector combination of boolean values with a specific dimension We use the on set size attribute to specify the amount of elements which should be on in the resulting vector lt parameter type on off mask name active processors dimension 7 on set size Onumber of threads gt The on set size can be a fixed value or a reference to a variable value In the case of reference to variable values the notation parameter should be used For example the notation 9number of threads indicates that the on set size should be equal to the number of threads parameter of the configuration under evaluation In this example we assume that the number of threads parameter type is an integer progression without explicit steps as a matter of 16 Multicube Explorer User Manual Version 0 5 fact the notation 9 parameter can refer only to integer parameters with a step 1 When the on set size attribute is not specified all the possible combinations of the Boolean vector are considered in the generation of the associated progression The dimension of the on off mask can be variable as well lt parameter type on off mask name Qo8 priorities dimension Onumber of threads gt The previous parameter specification contains as an example the Boolean QoS priorities for each of the active nodes of a target multi pro
38. ssi Mohammad ALaRI Universit della Svizzera italiana Carlos Kavka ESTECO Sara Bocchio STMicroelectonics Hector Posadas University of Cantabria This work is supported by the EC under grant FP7 216693 MULTICUBE http www multicube eu The Multicube Explorer tool and the documentation can be found at the following address http home dei polimi it zaccaria multicube explorer 30
39. y db change current first exploration db read first exploration db db change current second exploration V MM M db read second exploration db gt plot objectives in the second database loaded the current db gt db plot objectives gt plot objectives in the first database loaded gt db plot objectives first exploration gt plot objectives in the from both databases gt db plot objectives first exploration second exploration 28 Multicube Explorer User Manual Version 0 5 Shell Command name db show optimum Arguments Options Description Displays the optimal point stored in the current db The optimum point is consid ered the one with minimum product of objective metrics Usage example refers to use case example provided with the distribution of m3explorer Example usage gt db_read full_mpeg4 db gt set_objectives energy cycles gt db_show_optimum Information Computing optimum point please note that the cost function is the product of all the metrics Optimum point ics 16384 dcs 8192 12cs 131072 icw 1 dcw 8 12cw 2 iwidth 2 cbs 16 pn 8 cost 2 92596e 07 Shell Command name db compute ADRS Arguments Options Description Computes the Average Distance from Reference Set ADRS Such distance is a quantitative measure of solution qualities of a multiobjective ex ploration that is available for
Download Pdf Manuals
Related Search