New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Automatic NUMA binding #1874

Open

fxmarty opened this issue May 10, 2024 · 0 comments

Collaborator

fxmarty commented May 10, 2024

Feature request

Multiple NUMA nodes may have a big impact on latency, see https://huggingface.co/docs/optimum/main/en/amd/amdgpu/perf_hardware#numa-nodes.

Nvidia recommends to use correct NUMA bindings, or to disable numa balancing (which is NOT the default on many cloud VMs):
https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#numa-best-practices
https://docs.nvidia.com/grace-performance-tuning-guide.pdf

more context: https://huggingface.slack.com/archives/C043JTYE1MJ/p1715076269250249 & https://huggingface.slack.com/archives/C043JTYE1MJ/p1715163026161219

We may handle ourselves numa bindings for the GPUs.

Motivation

No slowdown due to wrong balancing

Your contribution

I may work on it sometime

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment