Skip to content

(3.11.x) Job submission failure with Amazon Linux 2023

hgreebe edited this page Nov 18, 2024 · 1 revision

The issue

We have discovered an issue in the official ParallelCluster AMI for Amazon Linux 2023 that consistently leads to job submission failure on p4 compute nodes.

If your cluster is affected by this issue, you will experience job submission failures caused by compute nodes failing to bootstrap. The bootstrap error is:

================================================================================
Error executing action `configure` on resource 'fabric_manager[Configure fabric manager]'
================================================================================

Mixlib::ShellOut::CommandTimeout
--------------------------------
service[nvidia-fabricmanager] (aws-parallelcluster-platform::nvidia_config line 35) had an error: Mixlib::ShellOut::CommandTimeout: Command timed out after 900s:
Command exceeded allowed execution time, process terminated

After consistent bootstrap errors, the cluster is eventually set to protected mode, where the partitions are deactivated. See here how to recover from protected mode.

We are investigating the root cause preventing the nvidia-fabricmanager service to start. The issue is impacting NVIDIA drivers 550.90.07 on Amazon Linux 2023. This version of NVIDIA drivers is included in 3.11.0 and 3.11.1 ParallelCluster AMIs.

Affected versions (OSes, schedulers)

  • Official AMIs of ParallelCluster 3.11.0 and 3.11.1 for Amazon Linux 2023
  • We observed impact on p4 instances, but we cannot exclude impact on other GPU instance types.
    • We tested g4dn, g5, g6, and p5 instances and found this issue does not manifest on them.

Mitigation

To mitigate this issue you can either:

  1. Build a custom AMI for Amazon Linux 2023 with NVIDIA driver 535.183.01, following the instructions in Upgrade the NVIDIA GPU driver on a Slurm cluster managed with AWS ParallelCluster.
  2. downgrade to ParallelCluster 3.10.1
  3. Consider using different GPU instance types. We tested g4dn, g5, g6, and p5 instances and found this issue does not manifest on them.
Clone this wiki locally