Newer
Older
This file describes changes in recent versions of Slurm. It primarily
documents those changes that are of interest to users and administrators.
* Changes in Slurm 17.11.8
==========================
-- Fix incomplete RESPONSE_[RESOURCE|JOB_PACK]_ALLOCATION building path.
-- Do not allocate nodes that were marked down due to the node not responding
by ResumeTimeout.
-- task/cray plugin - search for "mems" cgroup information in the file
"cpuset.mems" then fall back to the file "mems".
-- Fix ipmi profile debug uninitialized variable.
-- Improve detection of Lua package on older RHEL distributions.
-- MYSQL: Fix issue not handling all fields when loading an archive dump.
-- Allow a job_submit plugin to change the admin_comment field during
job_submit_plugin_modify().
-- job_submit/lua - fix access into reservation table.
-- MySQL - Prevent deadlock caused by archive logic locking reads.
-- Don't enforce MaxQueryTimeRange when requesting specific jobs.
-- Modify --test-only logic to properly support jobs submitted to more than
one partition.
Dominik Bartkiewicz
committed
-- Prevent slurmctld from abort when attempting to set non-existing
qos as def_qos_id.
-- Add new job dependency type of "afterburstbuffer". The pending job will be
delayed until the first job completes execution and it's burst buffer
stage-out is completed.
-- Reorder proctrack/task plugin load in the slurmstepd to match that of slurmd
and avoid race condition calling task before proctrack can introduce.
-- Prevent reboot of a busy KNL node when requesting inactive features.
-- Revert to previous behavior when requesting memory per cpu/node introduced
in 17.11.7.
-- Fix to reinitialize previously adjusted job members to their original value
when validating the job memory in multi-partition requests.
-- Fix _step_signal() from always returning SLURM_SUCCESS.
-- Combine active and available node feature change logs on one line rather
than one line per node for performance reasons.
-- Prevent occasionally leaking freezer cgroups.
-- Fix potential segfault when closing the mpi/pmi2 plugin.
-- Fix issues with --exclusive=[user|mcs] to work correctly
with preemption or when job requests a specific list of hosts.
-- mpi/pmix: Fixed the collectives canceling.
-- SlurmDBD: improve error message handling on archive load failure.
-- Fix incorrect locking when deleting reservations.
-- Fix incorrect locking when setting up the power save module.
-- Fix setting format output length for squeue when showing array jobs.
-- Fix printing out of --hint options in sbatch, salloc --help.
-- Prevent possible divide by zero in _validate_time_limit().
* Changes in Slurm 17.11.7
==========================
-- Fix for possible slurmctld daemon abort with NULL pointer.
-- Fix different issues when requesting memory per cpu/node.
-- PMIx - override default paths at configure time if --with-pmix is used.
-- Have sprio display jobs before eligible time when
PriorityFlags=ACCRUE_ALWAYS is set.
-- Make sure locks are always in place when calling _post_qos_list().
-- Notify srun and ctld when unkillable stepd exits.
-- Fix slurmstepd deadlock in stepd cleanup caused by race condition in
the jobacct_gather fini() interfaces introduced in 17.11.6.
-- Fix slurmstepd deadlock in PMIx startup.
-- task/cgroup - fix invalid free() if the hwloc library does not return a
string as expected.
-- Fix insecure handling of job requested gid field. CVE-2018-10995.
-- Add --without x11 option to rpmbuild in slurm.spec.
* Changes in Slurm 17.11.6
==========================
-- CRAY - Add slurmsmwd to the contribs/cray dir.
-- sview - fix crash when closing any search dialog.
-- Fix initialization of variable in stepd when using native x11.
-- Fix reading slurm_io_init_msg to handle partial messages.
-- Fix scontrol create res segfault when wrong user/account parameters given.
-- Fix documentation for sacct on parameter -X (--allocations)
-- Change TRES Weights debug messages to debug3.
-- FreeBSD - assorted fixes to restore build.
-- Fix for not tracking environment variables from unrelated different jobs.
-- PMIX - Added the direct connect authentication.
When upgrading this may cause issues with jobs using pmix starting on mixed
slurmstepd versions where some are less than 17.11.6.
-- Prevent the backup slurmctld from losing the active/available node
-- Add documentation for fix IDLE*+POWER due to capmc stuck in Cray systems.
-- Fix missing mutex unlock when prolog is failing on a node, leading to a
hung slurmd.
-- Fix locking around Cray CCM prolog/epilog.
-- Fix issue incorrectly setting a job time_start to 0 while requeueing.
-- smail - remove stray '-s' from mail subject line.
Ben Matthews
committed
-- srun - prevent segfault if ClusterName setting is unset but
SLURM_WORKING_CLUSTER environment variable is defined.
-- In configurator.html web pages change default configuration from
task/none to task/affinity plugin and from select/linear plugin to
select/cons_res plus CR_Core.
-- Allow jobs to run beyond a FLEX reservation end time.
-- Fix problem with wrongly set as Reservation job state_reason.
-- Prevent bit_ffs() from returnig value out of bitmap range.
-- Improve performance of 'squeue -u' when PrivateData=jobs is enabled.
-- Make UnavailableNodes value in job reason be correct for each job.
-- Fix 'squeue -o %s' on Cray systems.
-- Fix incorrect error thrown when cancelling part of a job array.
-- Fix error code and scheduling problem for --exclusive=[user|mcs].
-- Fix build when lz4 is in a non-standard location.
-- Be able to force power_down of cloud node even if in power_save state.
-- Allow cloud nodes to be recognized in Slurm when booted out of band.
-- Fixes race condition in _pack_job_gres() when is called multiple times.
-- Increase duration of "sleep" command used to keep extern step alive.
-- Remove unsafe usage of pthread_cancel in slurmstepd that can lead to
to deadlock in glibc.
-- Fix total TRES Billing on partitions.
-- Don't tear down a BB if a node fails and --no-kill or resize of a job
happens.
-- Remove unsafe usage of pthread_cancel in pmix plugin that can lead to
to deadlock in glibc.
-- Fix fatal in controller when loading completed trigger
-- Ignore reservation overlap at submission time.
-- GRES type model and QOS limits documentation added
-- slurmd - fix ABRT on SIGINT after reconfigure with MemSpecLimit set.
-- PMIx - move two error messages on retry to debug level, and only display
the error after the retry count has been exceeded.
-- Increase number of tries when sending responses to srun.
-- Fix checkpointing requeued/completing jobs in a bad state which caused a
segfault on restart.
-- Fix srun on ppc64 platforms.
-- Prevent slurmd from starting steps if the Prolog returns an error when using
PrologFlags=alloc.
-- priority/multifactor - prevent segfault running sprio if a partition has
just been deleted and PriorityFlags=CALCULATE_RUNNING is turned on.
-- job_submit/lua - add ESLURM_INVALID_TIME_LIMIT return code value.
Tim Wickberg
committed
-- job_submit/lua - print an error if the script calls log.user in
job_modify() instead of returning it to the next submitted job erroneously.
-- select/cons_res - improve handling of --cores-per-socket requests.
* Changes in Slurm 17.11.5
==========================
-- Fix cloud nodes getting stuck in DOWN+POWER_UP+NO_RESPOND state after not
responding by ResumeTimeout.
-- Add job's array_task_cnt and user_name along with partitions
[max|def]_mem_per_[cpu|node], max_cpus_per_node, and max_share with the
SHARED_FORCE definition to the job_submit/lua plugin.
-- srun - fix for SLURM_JOB_NUM_NODES env variable assignment.
-- sacctmgr - fix runaway jobs identification.
-- Fix for setting always the correct status on job update in mysql.
-- Fix issue if running with an association manager cache (slurmdbd was down
when slurmctld was started) you could loose QOS usage information.
-- CRAY - Fix spec file to work correctly.
-- Set scontrol exit code to 1 if attempting to update a node state to DRAIN
or DOWN without specifying a reason.
-- Fix race condition when running with an association manager cache
(slurmdbd was down when slurmctld was started).
-- Print out missing SLURM_PERSIST_INIT slurmdbd message type.
-- Fix two build errors related to use of the O_CLOEXEC flag with older glibc.
-- Add Google Cloud Platform integration scripts into contribs directory.
-- Fix minor potential memory leak in backfill plugin.
-- Add missing node flags (maint/power/etc) to node states.
-- Fix issue where job time limits may end up at 1 minute when using the
NoReserve flag on their QOS.
-- Fix security issue in accounting_storage/mysql plugin by always escaping
strings within the slurmdbd. CVE-2018-7033.
-- Soften messages about best_fit topology to debug2 to avoid alarm.
-- Fix issue in sreport reservation utilization report to handle more
allocated time than 100% (Flex reservations).
-- When a job is requesting a Flex reservation prefer the reservation's nodes
over any other nodes.
* Changes in Slurm 17.11.4
==========================
-- Add fatal_abort() function to be able to get core dumps if we hit an
"impossible" edge case.
-- Link slurmd against all libraries that slurmstepd links to.
Alejandro Sanchez
committed
-- Fix limits enforce order when they're set at partition and other levels.
-- Add slurm_load_single_node() function to the Perl API.
-- slurm.spec - change dependency for --with lua to use pkgconfig.
-- Fix small memory leaks in node_features plugins on reconfigure.
-- slurmdbd - only permit requests to update resources from operators or
administrators.
-- Fix handling of partial writes in io_init_msg_write_to_fd() which can
lead to job step launch failure under higher cluster loads.
-- MYSQL - Fix to handle quotes in a given work_dir of a job.
-- sbcast - fix a race condition that leads to "Unspecified error".
-- Log that support for the ChosLoc configuration parameter will end in Slurm
version 18.08.
-- Fix backfill performance issue where bf_min_prio_reserve was not respected.
-- Print MaxQueryTimeRange in "sacctmgr show config".
-- Correctly check return codes when creating a step to check if needing to
wait to retry or not.
-- Fix issue where a job could be denied by Reason=MaxMemPerLimit when not
requesting any tasks.
-- In perl tools, fix for regexp that caused extra incorrectly shown results.
-- Add some extra locks in fed_mgr to be extra safe.
-- Minor memory leak fixes in the fed_mgr on slurmctld shutdown.
-- Make sreport job reports also report duplicate jobs correctly.
-- Fix issues restoring certain Partition configuration elements, especially
when ReconfigFlags=KeepPartInfo is enabled.
-- Don't add TRES whose value is NO_VAL64 when building string line.
-- Fix removing array jobs from hash in slurmctld.
-- Print out missing user messages from jobsubmit plugin when srun/salloc are
Loading
Loading full blame...