02_cache/02_mm-vs-mv

Hands-on: Matrix-matrix product: Well-tuned DGEMM vs. others

• Choose either C/C++ (c.boost_eco) or Fortran (fortran.boost_eco) samples. Both of them are fine, as well.
  • boost_eco: run a job using the boost eco mode (freq=2200, eco_state=2)
  • normal: run a job using the normal mode (normal=2000, eco_state=0)

C/C++

How to compile and how to execute

1. Run a script of creating working space

• edit task.sh (modify --gname option).
• See objst in create_project.sh.
  • It indicates a list of the settings of compile options.
• Running create_project.sh, the executable file (run.x) is generated in Obj_*** directory.
  • This directory is also automatically generated.
  • When you desire to use the own compiler, set MAKE_DIR variable in create_project.sh from config to config.own.
• Example:

$ cd c.boost_eco 
$ bash create_project.sh 
$ ls 
Obj_fj-ssl2  Obj_fj-ssl2so

2. Run program

• A job script to execute the program is located in Obj_***/results.
• You can run the program either:

## To run as a batch job
$ cd Obj_fj-ssl2/results
$ pjsub task.sh  
## Or, to run in an interactive job
$ cd Obj_fj-ssl2/results
$ bash task.sh  

• Each of the samples in the Exercises will be completed within 6 minutes.
  • For safety, we set the elapsed time of the job script as 10 minutes.
  • One job script contains the execution with different kinds of matrix dimension. We treat only square matrices.
    • On the other hand, in c.fapp we fix a single specific matrix dimension.

Exercises A

• E1: Examine difference between a direct use of DGEMM and other implementations of matrix-matrix product in main.c.
• E2: Compare the performance (GFlop/s) of DGEMM to that in the others when changing matrix dimension, in Obj_fj-ssl2/.
  • Also, compare those to the peak value of GFlop/s of the single core in the boost eco mode.

Exercises B (advanced)

• E3: Try to use the dynamic library of Fujitsu SSL2, instead of the static one. Check config/Makefile.fj-ssl2so.
• E4: In c.normal, task.sh file is set to use the normal mode (--rsc-list "freq=2000, eco_state=0"). Run jobs in the directory c.normal and compare the performance with those in directory c.boost_eco.
• E5: In c.fapp, perform the basic CPUPA analysis using fapp. Using the CPUPA reports, compare dgemm to dgemv from the viewpoints of use of cache.
  • On the basic CPUPA, fapp-measurements are performed five times (i.e., -Hevent=pa1, -Hevent=pa2, -Hevent=pa3, -Hevent=pa4, and -Hevent=pa5).
  • Run the job with task-Bpa.sh. After the job completion, perform analyses using perf-Bpa.sh. Using the resultant CSV files and cpu_pa_report.xlsm, you can obtain the CPUPA report.
  • Indeed, you can find that there are various kinds of difference between dgemm and dgemv, other than cache.
• E6. Check the value of floating-point operation peak ratio in the CPUPA reports in c.fapp. Compare the values of the ratios between the normal mode and the boost eco mode.

Fortran

How to compile and how to execute

1. Run a script of creating working space

• edit task.sh (modify --gname option).
• See objst in create_project.sh.
  • It indicates a list of the settings of compile options.
• Running create_project.sh, the executable file (run.x) is generated in Obj_*** directory.
  • This directory is also automatically generated.
  • When you desire to use the own compiler, set MAKE_DIR variable in create_project.sh from config to config.own.
• Example:

$ cd fortran.boost_eco
$ bash create_project.sh 
$ ls 
Obj_fj-ssl2  Obj_fj-ssl2.fast  Obj_fj-ssl2so

2. Run program

• A job script to execute the program is located in Obj_***/results.
• You can run the program either:

## To run as a batch job
$ cd Obj_fj-ssl2/results
$ pjsub task.sh  
## Or, to run in an interactive job
$ cd Obj_fj-ssl2/results
$ bash task.sh  

• Each of the samples in the Exercises will be completed within 5 minutes.
  • For safety, we set the elapsed time of the job script as 10 minutes.
  • One job script contains the execution with different kinds of matrix dimension. We treat only square matrices.
    • On the other hand, in fortran.fapp we fix a single specific matrix dimension.

Exercises A

• E1: Examine difference between a direct use of DGEMM and other implementations of matrix-matrix product in main.F90.
• E2: Compare the performance (GFlop/s) of DGEMM to that in the others when changing matrix dimension, in Obj_fj-ssl2/. Also, compare those to the peak value of GFlop/s of the single core in the boost eco mode.

Exercises B (advanced)

• E3: Try the case of fortran/Obj_fj-ssl2.fast, in which -Kfast option is used. Are there any differences from the case of Obj_fj-ssl2?
• E4: Try to use the dynamic library of Fujitsu SSL2, instead of the static one. Check confing/Makefile.fj-ssl2so.
• E5: In fortran.normal, task.sh file is set to use the normal mode (--rsc-list "freq=2000, eco_state=0"). Run jobs in the directory fortran.normal and compare the performance with those in directory fortran.boost_eco.
• E6: In fortran.fapp, perform the basic CPUPA analysis using fapp. Using the CPUPA reports, compare dgemm to dgemv from the viewpoints of use of cache.
  • On the basic CPUPA, fapp-measurements are performed five times (i.e., -Hevent=pa1, -Hevent=pa2, -Hevent=pa3, -Hevent=pa4, and -Hevent=pa5).
  • Run the job with task-Bpa.sh. After the job completion, perform analyses using perf-Bpa.sh. Using the resultant CSV files and cpu_pa_report.xlsm, you can obtain the CPUPA report.
  • Indeed, you can find that there are various kinds of difference between dgemm and dgemv, other than cache.
• E7: Check the value of floating-point operation peak ratio in the CPUPA reports in fortran.fapp. Compare the values of the ratios between the normal mode and the boost eco mode.
• E8: Consider the case of using Fortran matmul function as a matrix-matrix-product routine.
  • See fortran/config/Makefile.fj-ssl2.mamul. We have two cases:
    • matmul with -Nalloc_assign -O3 option. The job execution time might become longer. (~10 minutes)
    • matmul with -Kfast, but not specifying -Nalloc_assign.