Julia multi-AMDGPU test solver to run on OLCF test system

2-minute read
julia•amdgpu•coding

We successfully launched a test simulation of a 2D linear diffusion on AMD MI250x using ROCm-aware MPI on OLCF Crusher system.

ROCm-aware MPI (AMDGPU.jl & ImplicitGlobalGrid.jl)

Ludovic Räss

3D viscous Stokes solution

During the Julia hackathon in April 2022, we finalised AMDGPU support (AMDGPU.jl ROCm programming) in MPI.jl and ImplicitGlobalGrid.jl for both ROCm-aware and standard AMDGPU-MPI capabilities.

We are thrilled to announce that the 2D diffusion solver successfully executed on 8 AMD MI250x GPUs using ROCm-aware MPI on OLCF's Crusher test system, a small replica of Frontier, the first exascale supercomputer:

Start diffusion process                            
Global grid: 32762x16382x1 (nprocs: 8, dims: 4x2x1)
Process 0 selecting device 1
Process 1 selecting device 2
Process 2 selecting device 3
Process 3 selecting device 4
Process 5 selecting device 6
Process 6 selecting device 7
Process 7 selecting device 8
Process 4 selecting device 5
Starting the time loop 🚀...done
Executed 1000 steps in = 2.679e+00 sec (@ T_eff = 595.00 GB/s)

The code and setup can be found in https://github.com/luraess/ROCm-MPI. On Crusher, we used Julia 1.8.0-rc1 and ROCm 5.1.0.

Thanks to William F. Godoy (OLCF) for having performed the tests.