Machine settings - Task 045

Fit to programme

This task has been identified by the working groups as part of the agenda behind WP 1.2.

The task number is 045.

Summary

The performance of a simulation or machine learning run hinges on correct settings in the code: how MPI is instructed to pin jobs, how many ranks are deployed on a NUMA architecture, how GPUs are tied to ranks, and so forth.

Matters are further complicated by the fact that specific libraries are often required, or that particular environment variables (such as MPI buffer sizes) can make a substantial difference to runtime.

This task proposes that a project examine the five high-level rubrics identified by SHAREing and develop standard/default recommendations: which settings should be checked first, independent of the particular cluster, and which issues warrant a deeper dive into how a user has configured a cluster job.

The outcomes should extend the existing SHAREing assessment guidebook, building on material already added to that document.

Methodology

Add to SHAREing guidebook

Outcomes

Add to SHAREing guidebook