Skip to content

How to treat CUDA_VISIBLE_DEVICES for multi-node packs of tests? #103

@mdmosby

Description

@mdmosby

If we pack a multi-node submission with test cases that require fewer than one node's resources (particularly GPU resources), how should the visible devices be managed? Can we do this? Is it even desirable?

Work with experts (e.g., Kokkos) to better understand the problem and solution space. Implement a solution, or provide configuration guidance for appropriately batching tests in a multi-node environment if it doesn't make sense to manage. Put in guardrails to prevent inadvertent misuse or having things behave unexpectedly.

Update this issue description when time permits for greater clarity/specificity.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions