HPCG Cuda Binary with multiple GPUs? about hpcg

[luszczek]

I would first suggest to use NVIDIA's HPC-Benchmark container from NVIDIA's NGC collection at https://catalog.ngc.nvidia.com/orgs/nvidia/containers/hpc-benchmarks because they have --config option with presets for multi-GPU nodes. The binary you reference is no longer maintained by NVIDIA.

[tcarneirop]

Hello Piotr,

thank you.

I see. However, I could not run this container they provide and I would like just to run the binary for some comparison purposes.

Do you know the parameters to pass to the binary so the code get executed on multiple gpus?

Thank you and best regards,

Tiago Carneiro

[luszczek]

The binary code from HPCG software page expects one GPU per MPI rank. So you should have as many MPI ranks as you have GPUs on your node. The MPI message will not use the network since the MPI ranks will all be sharing the same main memory space.

[tcarneirop]

Dear Piotr,

Thank you! Would it be possible to ask to add this information to the documentation?

[luszczek]

Done. Great suggestion, thanks!