A100 `amperenodes` communication to User Community

TODO

For this issue

• ETA of release
• Review of hpc-announce email

Timeline

This is tentative and subject to change.

• 2023-07-27: Physical installation complete
• 2023-08-04: Slurm/node configuration complete (first pass)
• 2023-08-09: Single-GPU testing complete
• 2023-09-25: Necessary fixes complete
• 2023-09-25: Release

For A100s generally

• Plan for local NVMe drives
  • Mike proposed RAID 0 striping of the two drives (performance?)
  • Mount path would be /local
• Remaining tasks for A100s
  • Node definitions in slurm.conf
  • QoS definitions in slurm.conf
  • Consistent shell variable for /local
  • Validating A100s
  • Testing A100s ()
  • Performance comparison of A100s to P100s
  • Add amperenodes to live OOD: #461 (closed)
• CUDA: rc/cluster-software#103
  • At least CUDA/toolkit >= 11.8
  • Ideally >= 12.0
  • cuDNN compiled against CUDA/toolkit
  • tensorrt compiled against CUDA/toolkit [optional]

Release coordination

• slurm.conf - remove restritions on access for amperenodes* partitions - Done https://gitlab.rc.uab.edu/rc/rc-slurm/-/merge_requests/38
• Add amperenodes* partitions to OOD Prod - #461 (closed)
• Communications Prepared
  • Shell MOTD
  • OOD MOTD
  • Docs Announcement
  • Docs Pages
  • HPC Announce
• Remove reservation in scontrol
• Release HPC Announce
• Notify and close relevant ServiceNow tickets

See the wiki page for current state of information to communicate.