|
|
|
* Description:
|
|
|
|
Job distributes tasks over multiple nodes even if they fit one node.
|
|
|
|
This might lead to increased communication latency and redundant network traffic when reading from filesystems.
|
|
|
|
|
|
|
|
* Criterion:
|
|
|
|
Number of allocated cores fit one node.
|
|
|
|
Additional criterion: High I/O or Lustre Bandwidth indicates high communication overhead (possibly unreliable metric on shared nodes).
|
|
|
|
Possibly, check SLURM arguments.
|
|
|
|
|
|
|
|
* Possible false positives:
|
|
|
|
MPI distributed program uses multiple accelerators and scheduler allocates accelerators on different nodes.
|
|
|
|
Program actually benefits from multiples nodes since it uses a lot of memory bandwidth.
|
|
|
|
|
|
|
|
* Possible cures/workarounds: Use -N to specify number of nodes.
|
|
|
|
Use --contiguous to use contiguous resources. |
