We are pleased to announce the availability of Slurm version 22.05.6.
This includes a fix to core selection for steps which could result in
random task launch failures, alongside a number of other moderate
severity issues.
- Marshall
--
Marshall Garey
Release Management, Support, and Development
SchedMD LLC - Commercial Slurm Development and Support
* Changes in Slurm 22.05.6
==========================
-- Fix a partition's DisableRootJobs=no from preventing root jobs from working.
-- Fix the number of allocated cpus for an auto-adjustment case in which the
job requests --ntasks-per-node and --mem (per-node) but the limit is
MaxMemPerCPU.
-- Fix POWER_DOWN_FORCE request leaving node in completing state.
-- Do not count magnetic reservation queue records towards backfill limits.
-- Clarify error message when --send-libs=yes or BcastParameters=send_libs
fails to identify shared library files, and avoid creating an empty
"<filename>_libs" directory on the target filesystem.
-- Fix missing CoreSpec on dynamic nodes upon slurmctld restart.
-- Fix node state reporting when using specialized cores.
-- Fix number of CPUs allocated if --cpus-per-gpu used.
-- Add flag ignore_prefer_validation to not validate --prefer on a job.
-- Fix salloc/sbatch SLURM_TASKS_PER_NODE output environment variable when the
number of tasks is not requested.
-- Permit using wildcard magic cookies with X11 forwarding.
-- cgroup/v2 - Add check for swap when running OOM check after task
termination.
-- Fix deadlock caused by race condition when disabling power save with a
reconfigure.
-- Fix memory leak in the dbd when container is sent to the database.
-- openapi/dbv0.0.38 - correct dbv0.0.38_tres_info.
-- Fix node SuspendTime, SuspendTimeout, ResumeTimeout being updated after
altering partition node lists with scontrol.
-- jobcomp/elasticsearch - fix data_t memory leak after serialization.
-- Fix issue where '*' wasn't accepted in gpu/cpu bind.
-- Fix SLURM_GPUS_ON_NODE for shared GPU gres (MPS, shards).
-- Add SLURM_SHARDS_ON_NODE environment variable for shards.
-- Fix srun error with overcommit.
-- Fix bug in core selection for the default cyclic distribution of tasks
across sockets, that resulted in random task launch failures.
-- Fix core selection for steps requesting multiple tasks per core when
allocation contains more cores than required for step.
-- gpu/nvml - Fix MIG minor number generation when GPU minor number
(/dev/nvidia[minor_number]) and index (as seen in nvidia-smi) do not match.
-- Fix accrue time underflow errors after slurmctld reconfig or restart.
-- Surpress errant errors from prolog_complete about being unable to locate
"node:(null)".
-- Fix issue where shards were selected from multiple gpus and failed to
allocate.
-- Fix step cpu count calculation when using --ntasks-per-gpu=.
-- Fix overflow problems when validating array index parameters in slurmctld
and prevent a potential condition causing slurmctld to crash.
-- Remove dependency on json-c in slurmctld when running with power saving.
Only the new "SLURM_RESUME_FILE" support relies on this, and it will be
disabled if json-c support is unavailable instead.
