Developer Guide

Developer Guide for Intel® oneAPI Math Kernel Library Linux*

ID 766690
Date 11/07/2023
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

Running the Intel® Distribution for LINPACK* Benchmark and the Intel® Optimized HPL-AI* Benchmark

To run the Intel® Distribution for LINPACK Benchmark on multiple nodes or on one node with multiple MPI processes, you need to use MPI and either modify HPL.dat or use Ease-of-use Command-line Parameters. The following example describes how to run the dynamically-linked prebuilt Intel® Distribution for LINPACK Benchmark binary using the script provided. To run other binaries, adjust the steps accordingly; specifically, change line 58 of runme_intel64_dynamic to point to the appropriate binary.

  1. Load the necessary environment variables by accessing setup scripts.

  2. In HPL.dat, set the problem size N to 10000. Because this setting is for a test run, the problem size should be small.

  3. For better performance, enable non-uniform memory access (NUMA) on your system and configure to run an MPI process for each NUMA socket as explained below.

    • Refer to your BIOS settings to enable NUMA on your system.

    • Set the following variables at the top of the runme_intel64_dynamic script according to your cluster configuration:

      MPI_PROC_NUM
      The total number of MPI processes.
      MPI_PER_NODE
      The number of MPI processes per each cluster node.
      NUMA_PER_MPI
      The number of NUMA nodes per each MPI process.
      USE_HPL_AI
      Uncomment this to enable the Intel® Optimized HPL-AI* Benchmark.
      USE_HPL_GPU
      Uncomment this to enable GPUs.
      HPL_NUMSTACK
      The number of stacks on each GPU.
      HPL_NUMDEV
      The number of GPUs.
    • In the HPL.dat file, set the parameters Ps and Qs so that Ps * Qs equals the number of MPI processes. For example, for two processes, set Ps to 1 and Qs to 2. Alternatively, leave the HPL.dat file as is and launch with the -p and -q command-line parameters.

  4. Execute the runme_intel64_dynamic script:

    ./runme_intel64_dynamic

  5. Rerun the test, increasing the size of the problem until the matrix size uses about 80% of the available memory. To do this, either modify Ns in line 6 of HPL.dat or use the -n command-line parameter:

    • For 16 GB: 40000 Ns

    • For 32 GB: 56000 Ns

    • For 64 GB: 83000 Ns

Product and Performance Information

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.

Notice revision #20201201