Version: AMD_ZEN_HPL_2024-10-08

Description:

HPL is a software package that solves a (random) dense linear system in double precision (64 bits) arithmetic on distributed-memory computers.

This version is derived from The High-Performance Computing Linpack Benchmark (HPL - 2.3 - December 2, 2018) and has been optimized to run on AMD CPUs.

Dependencies:

  • OpenMPI 4/5:  This release was built with OpenMPI 5.0.5 and should run without issue if OpenMPI 5 or 4 is in the environment.
  • This release was built on Rocky Linux 8.6, and tested on CentOS 8.3, Rocky Linux 9.0, and Ubuntu Linux 22.04.

Recommended Settings:

  • Boost: ON
  • Transparent Hugepages: always
  • SMT: OFF
  • NPS: 4
  • Determinism: Power

How to Run:

  • Ensure the above dependencies have been satisfied, and mpirun from OpenMPI is on your $PATH
  • Optionally, run the ./reset-system.sh script to set various OS parameters to recommended values.
  • Optionally, create an HPL.dat file, as described in the documentation.
    AMD Zen HPL will auto-tune to select reasonable default parameters, so an HPL.dat file is not required. By default, AMD_Zen_HPL will attempt to use roughly 90% of system memory.
  • For single-node runs, run ./run.sh. For multi-node runs, adapt the mpirun line from run.sh to suit the local site.
  • Recommended arguments to mpirun are:
    • --bind-to core --map-by socket:PE=$NUM_CORES_PER_SOCKET - to specify 1 MPI rank per socket, with $NUM_CORES_PER_SOCKET (ideally set to number of cores per socket) CPUs assigned to the rank.