GROMACS benchmarks

The repository provides benchmarks for GROMACS.

Status

Under development.

Maintainers

• James Gebbie-Rayet
• Robert Welch

Overview

Code/library

GROMACS

Architectures

• CPU: x86, Arm64, PPC64
• GPU: NVIDIA, AMD, Intel*

Languages and programming models

• Programming languages: C, C++, Python*
• Parallel models: MPI, OpenMP*
• Accelerator offload models: CUDA, OpenCL, SYCL

Seven 'dwarfs'

• Dense linear algebra
• Sparse linear algebra
• Spectral methods
• N-body methods
• Structured grids
• Unstructured grids
• Monte Carlo

Building the benchmark

The benchmark can be built using Spack or manually. If you are using the ReFrame method to run the benchmark described below, it will automatically perform the build step for you.

Once it has been built the benchmark executable is called name-of-exe.x

Spack build

Gromacs can be installed from a spack repository maintained by the core developers:

For GROMACS with MPI for CPU based architecture

spack install [email protected] +mpi

For an MPI + CUDA build for NVIDA GPU architecture

spack install [email protected] +mpi +cuda

For a multinode, multiGPU compilation targetting specific GPU arch

spack install [email protected] +mpi +cuda cuda_arch=70,80,90 +cufftmp

Often better for single NVIDA GPU tests is to use threadMPI and CUDA

spack install [email protected] ~mpi +cuda

Enable SYCL for use with Intel and AMD GPUs, may need to add hardware specific backends to the commands

spack install [email protected] +sycl

Note: to use Spack, you must have Spack installed on the system you are using and a valid Spack system configuration. Example Spack configurations are available in a separate repository: [https://github.com/ukri-bench/system-configs]

Manual build

• ADD: Describe (or link to) the manual build process for a systems where baseline performance has been measured.
  • Links to multiple sub-pages of instructions can be added if they are too long to fit on this page

Running the benchmark

Example performance data

This section contains example performance data from selected HPC systems.

ADD: Example performance data

License

This benchmark description and associated files are released under the MIT license.