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 17.02.0pre2
==============================
-- Add new RPC (REQUEST_EVENT_LOG) so that slurmd and slurmstepd can log events
through the slurmctld daemon.
-- Remove sbatch --bb option. That option was never supported.
-- Automically cleanup task/cgroup cpuset and devices cgroups after steps are
done.
-- Limit job purge run time to 1 second at a time.
-- The database index for jobs is now 64 bit. If you happen to be close to
4 billion jobs in your database you will want to update your slurmctld at
the same time as your slurmdbd to prevent roll over of this variable as
it is 32 bit previous versions of Slurm.
-- Optionally lock slurmstepd in memory for performance reasons and to avoid
possible SIGBUS if the daemon is paged out at the time of a Slurm upgrade
(changing plugins). Controlled via new LaunchParameters options of
slurmstepd_memlock and slurmstepd_memlock_all.
-- Add event trigger on burst buffer errors (see strigger man page,
--burst_buffer option).
-- Add job AdminComment field which can only be set by a Slurm administrator.
-- Add salloc, sbatch and srun option of --delay-boot=<time>, which will
temporarily delay booting nodes into the desired state for a job in the
hope of using nodes already in the proper state which will be available at
a later time.
* Changes in Slurm 17.02.0pre1
==============================
-- burst_buffer/cray - Add support for rounding up the size of a buffer reqeust
if the DataWarp configuration "equalize_fragments" is used.
-- Rename "in" to "input" in slurm_step_io_fds data structure defined in
slurm.h. This is needed to avoid breaking Python with by using one of its
keywords in a Slurm data structure.
-- Remove eligible_time from jobcomp/elasticsearch.
-- Fix issue where if no clusters were added but yet a QOS needed to be
deleted make it possible.
-- SlurmDBD - change all timestamps to bigint from int to solve Y2038 problem.
-- Add salloc/sbatch/srun --spread-job to distribute tasks over as many nodes
as possible. This also treats the --ntasks-node-node option as a maximum
value.
-- Add ConstrainKmemSpace to cgroup.conf, defaulting to yes, to allow
cgroup Kmem enforcement to be disabled while still using ConstrainRAMSpace.
-- Add support for sbatch --bbf option.
-- Add burst buffer support for job arrays. Add new SchedulerParameters option
of bb_array_stage_cnt=# to indicate how many pending tasks of a job array
should be made available for burst buffer resource allocation.
-- Fix small memory leak when a job fails to load from state save.
-- Fix invalid read when attempting to delete clusters from db with running
jobs.
-- Fix small memory leak when deleting clusters from db.
-- Add SLURM_ARRAY_TASK_COUNT environment variable. Total number of tasks in a
job array (e.g. "--array=2,4,8" will set SLURM_ARRAY_TASK_COUNT=3).
-- Add new sacctmgr commands: "shutdown" (shutdown the server), "list stats"
(get server statistics) "clear stats" (clear server statistics).
Tim Wickberg
committed
-- Restructure job accounting query to use 'id_job in (1, 2, .. )' format
instead of logically equivalent 'id_job = 1 || id_job = 2 || ..' .
-- Added start_delay field to jobcomp/elasticsearch.
-- In order to support federated jobs, the MaxJobID configuration parameter
default value has been reduced from 2,147,418,112 to 67,043,328 and its
maximum value is now 67,108,863. Upon upgrading, any pre-existing jobs that
have a job ID above the new range will continue to run and new jobs will get
job IDs in the new range.
-- Added infrastructure for setting up federations in database and establishing
connections between federation clusters.
* Changes in Slurm 16.05.5
==========================
-- Fix accounting for jobs requeued after the previous job was finished.
-- slurmstepd modified to pre-load all relevant plugins at startup to avoid
the possibility of modified plugins later resulting in inconsistent API
or data structures and a failure of slurmstepd.
-- Export functions from parse_time.c in libslurm.so.
-- Export unit convert functions from slurm_protocol_api.c in libslurm.so.
-- Fix scancel to allow multiple steps from a job to be cancelled at once.
-- Update and expand upgrade guide (in Quick Start Administrator web page).
-- burst_buffer/cray: Requeue, but do not hold a job which fails the pre_run
operation.
-- Insure reported expected job start time is not in the past for pending jobs.
* Changes in Slurm 16.05.4
==========================
-- Fix potential deadlock if running with message aggregation.
-- Streamline when schedule() is called when running with message aggregation
on batch script completes.
-- Fix incorrect casting when [un]packing derived_ec on slurmdb_job_rec_t.
-- Document that persistent burst buffers can not be created or destroyed using
the salloc or srun --bb options.
-- Add support for setting the SLURM_JOB_ACCOUNT, SLURM_JOB_QOS and
SLURM_JOB_RESERVAION environment variables are set for the salloc command.
Document the same environment variables for the salloc, sbatch and srun
commands in their man pages.
-- Fix issue where sacctmgr load cluster.cfg wouldn't load associations
that had a partition in them.
-- Don't return the extern step from sstat by default.
-- In sstat print 'extern' instead of 4294967295 for the extern step.
-- Make advanced reservations work properly with core specialization.
-- Fix race condition in the account_gather plugin that could result in job
stuck in COMPLETING state.
-- Regression test fixes if SelectTypePlugin not managing memory and no node
memory size set (defaults to 1 MB per node).
Dominik Bartkiewicz
committed
-- Add missing partition write locks to _slurm_rpc_dump_nodes/node_single to
prevent a race condition leading to inconsistent sinfo results.
-- Fix task:CPU binding logic for some processors. This bug was introduced
in version 16.05.1 to address KNL bunding problem.
-- Fix two minor memory leaks in slurmctld.
-- Improve partition-specific limit logging from slurmctld daemon.
-- Fix incorrect access check when using MaxNodes setting on the partition.
Dominik Bartkiewicz
committed
-- Fix issue with sacctmgr when specifying a list of clusters to query.
-- Fix issue when calculating future StartTime for a job.
-- Make EnforcePartLimit support logic work with any ordering of partitions
in job submit request.
Tim Wickberg
committed
-- Prevent restoration of wrong CPU governor and frequency when using
multiple task plugins.
-- Prevent slurmd abort if hwloc library fails to populate the "children"
arrays (observed with hwloc version "dev-333-g85ea6e4").
-- burst_buffer/cray: Add "--groupid" to DataWarp "setup" command.
Iakovos Panourgias
committed
-- Fix lustre profiling putting it in the Filesystem dataset instead of the
Network dataset.
-- Fix profiling documentation and code to match be consistent with
Filesystem instead of Lustre.
-- Correct the way watts is calculated in the rapl plugin when using a poll
frequency other than AcctGatherNodeFreq.
-- Don't about step launch if job reaches expected end time while node is
configuring/booting (NOTE: The job end time will be adjusted after node
becomes ready for use).
Tim Wickberg
committed
-- Fix several print routines to respect a custom output delimiter when
printing NO_VAL or INFINITE.
-- Correct documented configurations where --ntasks-per-core and
--ntasks-per-socket are supported.
-- task/affinity plugin buffer allocated too small, can corrupt memory.
* Changes in Slurm 16.05.3
==========================
-- Make it so the extern step uses a reverse tree when cleaning up.
-- If extern step doesn't get added into the proctrack plugin make sure the
sleep is killed.
-- Fix areas the slurmctld can segfault if an extern step is in the system
cleaning up on a restart.
-- Prevent possible incorrect counting of GRES of a given type if a node has
the multiple "types" of a given GRES "name", which could over-subscribe
GRES of a given type.
-- Add web links to Slurm Diamond Collectors (from Harvard University) and
collectd (from EDF).
-- Add job_submit plugin for the "reboot" field.
-- Make some more Slurm constants (INFINITE, NO_VAL64, etc.) available to
job_submit/lua plugins.
-- Send in a -1 for a taskid into spank_task_post_fork for the extern_step.
-- MYSQL - Sightly better logic if a job completion comes in with an end time
of 0.
-- task/cgroup plugin is configured with ConstrainRAMSpace=yes, then set soft
memory limit to allocated memory limit (previously no soft limit was set).
-- Document limitations in burst buffer use by the salloc command (possible
access problems from a login node).
-- Fix proctrack plugin to only add the pid of a process once
(regression in 16.05.2).
-- Fix for sstat to print correct info when requesting jobid.batch as part of
a comma-separated list.
-- CRAY - Fix issue if pid has already been added to another job container.
-- CRAY - Fix add of extern step to AELD.
-- burstbufer/cray: avoid batch submit error condition if waiting for stagein.
-- CRAY - Fix for reporting steps lingering after they are already finished.
-- Testsuite - fix test1.29 / 17.15 for limits with values above 32-bits.
-- CRAY - Simplify when a NHC is called on a step that has unkillable
processes.
-- CRAY - If trying to kill a step and you have NHC_NO_STEPS set run NHC
anyway to attempt to log the backtraces of the potential
unkillable processes.
-- Fix gang scheduling and license release logic if single node job killed on
bad node.
-- Make scontrol show steps show the extern step correctly.
-- Do not scheduled powered down nodes in FAILED state.
-- Do not start slurmctld power_save thread until partition information is read
in order to prevent race condition that can result invalid pointer when
trying to resolve configured SuspendExcParts.
-- Add SLURM_PENDING_STEP id so it won't be confused with SLURM_EXTERN_CONT.
-- Fix for core selection with job --gres-flags=enforce-binding option.
Previous logic would in some cases allocate a job zero cores, resulting in
slurmctld abort.
-- Minimize preempted jobs for configurations with multiple jobs per node.
-- Improve partition AllowGroups caching. Update the table of UIDs permitted to
use a partition based upon it's AllowGroups configuration parameter as new
valid UIDs are found rather than looking up that user's group information
for every job they submit. If the user is now allowed to use the partition,
then do not check that user's group access again for 5 seconds.
-- Add routing queue information to Slurm FAQ web page.
-- Do not select_g_step_finish() a SLURM_PENDING_STEP step, as nothing has
been allocated for the step yet.
-- Fixed race condition in PMIx Fence logic.
-- Prevent slurmctld abort if job is killed or requeued while waiting for
reboot of its allocated compute nodes.
-- Treat invalid user ID in AllowUserBoot option of knl.conf file as error
rather than fatal (log and do not exit).
-- qsub - When doing the default output files for an array in qsub style
make them using the master job ID instead of the normal job ID.
-- Create the extern step while creating the job instead of waiting until the
end of the job to do it.
-- Always report a 0 exit code for the extern step instead of being canceled
Loading
Loading full blame...