github SchedMD/slurm slurm-26-05-3-1
v26.05.3

3 hours ago

Changes in 26.05.3

  • async srun steps now honor --kill-on-bad-exit and kill peers.
  • Fix sinfo --states to accept REBOOT_REQUESTED and REBOOT_ISSUED filter names documented in the man page.
  • Harden LicenseParameters=RemoteFuzzyMatch matching for license names that unexpectedly contain an '@' character.
  • scontrol - Fix local licenses not appearing for scontrol show licenses=<license_name> when LicenseParameters=RemoteFuzzyMatch is set.
  • auth/jwt - Prevent fatal() on non-RSA keys in mixed JWKS files.
  • slurmctld - Avoid breaking backfill cycle on busy reservation delete.
  • Fix srun intermittently hanging at job or step completion due to a connection manager shutdown race.
  • slurmrestd - Fix HTTP 500 error from GET /slurm/{version}/nodes when the update_time query parameter is passed and only partition data is unchanged.
  • Fix slurmctld segfault when QOS-based SUSPEND preemption occurs on a partition with OverSubscribe=YES.
  • slurmd - Fix corrupted job step cancellation message request when a job exceeds its memory limit with JobAcctGatherParams=OverMemoryKill, resulting in afflicted job steps not being killed.
  • data_parser/v0.0.45 - Avoid hitting an xassert when there is an error while parsing arrays in topology.yaml. The source string in the error will now be the path to the field experiencing the error instead of the function name where the error happened.
  • Fix srun --overlap step being allocated fewer GPUs than the job when the job's allocated cores span fewer sockets than its GPUs.
  • topology/block - Validate BlockSizes and fatal when they are invalid.
  • Give an error at slurmdbd startup when MariaDB sets innodb_snapshot_isolation to ON. Also give a specific fatal message when a write fails with the ER_CHECKREAD error.
  • slurmctld - Fix memory leak when a job update requesting multiple QOS is rejected.
  • slurmctld - Fix memory leak when a job update requesting multiple reservations is rejected.
  • Fix client command crashes when using stepmgr with allocations that are powering up powered down nodes.
  • Fix handling of UIDs and GIDs above 2^31 in gid_from_string(), the data_parser plugins and squeue output.
  • auth/jwt - Fix the token identity being discarded when the requested username matches it, instead of when it differs.
  • sched/backfill - Prevent undoing a job priority update that occurs while backfill is running.
  • sched/backfill - Prevent undoing a job TimeLimit update if the update occurs while backfill is running. Additionally, enforce updates to a job's TimeMin and Deadline during the current backfill cycle.
  • Enable delete/recreate of dynamic cloud/external nodes to schedule jobs.
  • Enable external or mixed external/regular hetjobs
  • Enable external and mixed-type hetjobs to be submitted via sbatch.
  • Enable hetjobs with external components to be submitted via REST interface.
  • Enable sbatch to submit mixed-type regular/external hetjobs.

Don't miss a new slurm release

NewReleases is sending notifications on new releases.