How to run HPL/HPCG/IO500 in WSL (5) | 青训营笔记

3. HPCG Test

0. Get source from github.com/hpcg-benchm….

    git clone https://github.com/hpcg-benchmark/hpcg.git && cd hpcg
   

1. Copy example config.

    cp setup/Make.Linux_MPI ./
   

2. Modify Configuration, using nano Make.Linux_MPI.

    TOPdir       = $(HOME)/hpcg
    ​
    MPdir        = /usr/lib/x86_64-linux-gnu/openmpi
    MPinc        = -I$(MPdir)/include
    MPlib        = $(MPdir)/lib/libmpi.so
    ​
    HPCG_OPTS     = -DHPCG_NO_OPENMP -DHPCG_DETAILED_TIMING
   

3. Build from source. If the build fails, delete the folder and download the source code again.

    make arch=Linux_MPI -j4
   

4. Now you can find the xhpcg and hpcg.dat files in the bin directory

5. Test HPCG.

    cd bin
    mpirun -np 4 xhpcg --nx=16 --rt=1800
    # The test will last for 30 minutes (1800 seconds)
   

   Output in file hpcg<identifier>.txt and HPCG-Benchmark_3.1_<identifier>.txt:

    $ cat HPCG-Benchmark_3.1_<identifier>.txt | grep GFLOP
    GFLOP/s Summary=
    GFLOP/s Summary::Raw DDOT=1.46162
    GFLOP/s Summary::Raw WAXPBY=18.9513
    GFLOP/s Summary::Raw SpMV=10.583
    GFLOP/s Summary::Raw MG=8.18653
    GFLOP/s Summary::Raw Total=7.88411
    GFLOP/s Summary::Total with convergence overhead=7.88411
    GFLOP/s Summary::Total with convergence and optimization phase overhead=7.55894
    Final Summary::HPCG result is VALID with a GFLOP/s rating of=7.55894
   

   Result is 7.55894 GFLOP/s

6. Performance Tuning

   See github.com/hpcg-benchm… and modify the HPCG.dat

    HPCG benchmark input file
    Sandia National Laboratories; University of Tennessee, Knoxville
    120 120 120
    120
   

   And run again.

4. IO500 Test

0. Clone the io500 source.

    git clone https://github.com/IO500/io500.git && cd io500
   

1. To Retrieve the required packages and compile the library version, run:

    sudo apt install pkg-config # to solve autoconf errors
    ./prepare.sh
   

   When it’s OK,output:

    OK: All required software packages are now prepared
    ior  md-workbench  mdtest  pfind
   

2. In order to create a new INI file with all the options, you can execute:

    ./io500 --list > config-all.ini
   

3. (Optional) Modify config-all.ini (My disk space is not enough, you may need not modify).

    [ior-easy]
    blockSize = 100m
    ​
    [mdtest-easy]
    n = 100
    ​
    [ior-hard]
    segmentCount = 100
    ​
    [find-hard]
    pfind-queue-length = 100
   

4. Then you can running the io500 test:

    mpiexec -np 8 ./io500 config-all.ini
   

   Output:

    IO500 version io500-isc23 (standard)
    ERROR INVALID (src/phase_dbg.c:15) stonewall-time < 300s
    ERROR INVALID (src/phase_ior.c:24) Write phase needed 1.370603s instead of stonewall 5s. Stonewall was hit at 0.2s
    ERROR INVALID (src/main.c:437) Runtime of phase (1.379476) is below stonewall time. This shouldn't happen!
    ERROR INVALID (src/main.c:443) Runtime is smaller than expected minimum runtime
    [RESULT]       ior-easy-write        0.570005 GiB/s : time 1.379 seconds [INVALID]
    ERROR INVALID (src/main.c:437) Runtime of phase (1.056533) is below stonewall time. This shouldn't happen!
    ERROR INVALID (src/main.c:443) Runtime is smaller than expected minimum runtime
    [RESULT]    mdtest-easy-write       15.711578 kIOPS : time 1.057 seconds [INVALID]
    [      ]            timestamp        0.000000 kIOPS : time 0.000 seconds
    ERROR INVALID (src/phase_ior.c:24) Write phase needed 0.484803s instead of stonewall 5s. Stonewall was hit at 0.0s
    ERROR INVALID (src/main.c:437) Runtime of phase (0.485622) is below stonewall time. This shouldn't happen!
    ERROR INVALID (src/main.c:443) Runtime is smaller than expected minimum runtime
    [RESULT]       ior-hard-write        0.072243 GiB/s : time 0.486 seconds [INVALID]
    ERROR INVALID (src/main.c:443) Runtime is smaller than expected minimum runtime
    [RESULT]    mdtest-hard-write       56.509355 kIOPS : time 8.147 seconds [INVALID]
    [RESULT]                 find     2339.353769 kIOPS : time 0.173 seconds
    [RESULT]        ior-easy-read       14.772937 GiB/s : time 0.057 seconds
    [RESULT]     mdtest-easy-stat     2115.664061 kIOPS : time 1.002 seconds
    [RESULT]        ior-hard-read       18.307577 GiB/s : time 0.002 seconds
    [RESULT]     mdtest-hard-stat     7693.419492 kIOPS : time 1.059 seconds
    [RESULT]   mdtest-easy-delete      481.067125 kIOPS : time 1.003 seconds
    [RESULT]     mdtest-hard-read     1630.211454 kIOPS : time 1.252 seconds
    [RESULT]   mdtest-hard-delete      145.488834 kIOPS : time 3.870 seconds
    [SCORE ] Bandwidth 1.826809 GiB/s : IOPS 499.212193 kiops : TOTAL 30.198765 [INVALID]
    ​
    The result files are stored in the directory: ./results/<date>-<time>
   

   Why is [INVALID], because the running time is not enough for 300 seconds :).

5. Performance Tuning.

   Modify the config-all.ini

    [ior-easy]
    blockSize = 9920000m
    ​
    [mdtest-easy]
    n = 1000000
    ​
    [ior-hard]
    segmentCount = 10000000
    ​
    [find-hard]
    pfind-queue-length = 10000
   

   And run it again.

5. Reference

0. chat.openai.com
1. www.deepl.com/translator
2. www.open-mpi.org
3. www.netlib.org/benchmark/h…
4. www.hpcg-benchmark.org
5. github.com/hpcg-benchm…
6. github.com/IO500/io500
7. www.cnblogs.com/liu-shaobo/…
8. blog.csdn.net/gyx15496246…