Bug #10004
closed
[Crunch] Terminate on node failure in setup steps
Description
If an allocation includes multiple nodes, it's possible for a node to fail during one of the setup tasks executed by crunch-job (such as docker load). When a node fails, srun can hang.
The current code checks squeue to see if the step has gone away, and kills srun. However, if one node fails but there are other nodes still in the allocation, the job will still be in squeue. As a result, the stuck srun process isn't terminated.
Updated by Peter Amstutz over 9 years ago
- Subject changed from [Crunch] Handle partial node failure to [Crunch] Handle partial node failure in setup steps
- Description updated (diff)
Updated by Peter Amstutz over 9 years ago
- Subject changed from [Crunch] Handle partial node failure in setup steps to [Crunch] Terminate on node failure in setup steps
- Assigned To set to Peter Amstutz
- Target version set to 2016-09-14 sprint
Updated by Tom Clegg over 9 years ago
10004-check-sinfo @ 19ad5c5
This should have a comment explaining why it's necessary, and why it only makes sense during srun_sync (i.e., at that point, the failure of even one node makes any further effort futile).
The $last_sinfo_check variable is superfluous. Might as well just compare and update $sinfo_checked directly.
Other than that, LGTM. Thanks.
Updated by Peter Amstutz over 9 years ago
- Status changed from New to Resolved
- % Done changed from 0 to 100
Applied in changeset arvados|commit:b54478ea1b7c8aaeaf565d591f32769bcdc09b8f.