gpuenv-utils v0.1: Utilities for cooperative access to GPUs
Hi all,
I've released gpuenv-utils v0.1 [1] to the unstable distribution of our
APT archive [2]. From the README:
> This set of utilities aims to monitor and to facilitate cooperative
> access to GPUs. The primary intent is to let various concurrently
> running test workers, build workers, and porter box environments have
> exclusive control over one or more GPUs for the duration of their
> work.
The first user of this will be an updated debci (coming shortly), where
workers for Debian, Ubuntu and anything else APT-based will be able to
share GPU access in a serialized fashion. It should also be useful to
worker operators (like myself) wishing to run autopkgtests locally
without having to shutdown debci.
Both functionality and protocol are deliberately very simple. The goal
right now is to gather experience, and to iterate quickly towards
porterboxes, so that most GPUs in our network can be accessed by
contributors.
This is currently deployed on ci-dev, where I'm finalizing the debci
update to make use of this.
Best,
Christian
PS: I dread naming things, and "gpuenv-utils" is a pretty bad name. If
anyone has a better idea, please share.
[1]: https://salsa.debian.org/rocm-team/gpuenv-utils
[2]: https://apt.rocm.debian.net/
Reply to: