SIG/HPC Meeting 2024-02-22

Attendees

NVIDIA GPU driver Testing - Chris
- https://github.com/mghpcsim/gpu-testing/tree/master
- documented process for configuring instance, installing drivers (open source or proprietary), setting up container runtimes, nvidia container toolkit
  - Benchmarks using forked toolkit from Lambda labs with Rocky customizations
- initial control benchmark (pytorch):
  - closed drivers slightly (4s) faster
- Plan: run benchmarks on progressively newer instances and collect results
  - Publish results on Wiki
Intel driver - Met with them, went well
- Can build this driver into signed kernel modules, add to testing Chris is doing
- This will live in SIG/Kernel because it's a kernel module
- driver toolkit pieces probably will end up in HPC SIG
Kernel Cnode (for MoS)
- Sherif synced with Jeremy
  - Lots of progress has been made, almost all patches backported
  - there are couple problematic patches--they're based on SLES kernels, but a bit different enough to be problematic
    - Pablo will help once the problem is set

Testing - Warewulf, others
- Sherif and Brian Phan synced on warewulf testing
- Not just installibility, upgrade path, etc
- What can we use? Multiple things, probably
  - OpenQA? TMT? Zuul? Whatever OpenHPC uses?
  - Testing team would also love to get more people involved and participating in building tests
- Example tests:
  - Provision cluster
  - nodes communicate
  - etc
- Want: have full end to end testing of all components
- What tests do we want?
  - Functional
    - Create cluster
    - Create user
    - Submit job as user
    - Future:
      - Slurm accounting/dbd, others
Package tracking - PoI tracker
- Neil is looking how we can integrate this
Wiki Updates - Neil and Sherif will work on this at FOSDEM/CentOS Connect
- This didn't really happen specifically, but discussions about ensuring Wikis are up to date did happen