Newer
Older
This file describes changes in recent versions of SLURM. It primarily
documents those changes that are of interest to users and admins.
* Changes in SLURM 2.5.1
========================
-- Correction to hostlist sorting for hostnames that contain two numeric
components and the first numeric component has various sizes (e.g.
"rack9blade1" should come before "rack10blade1")
-- BGQ - Only poll on initialized blocks instead of calling getBlocks on
each block independently.
-- Fix of task/affinity plugin logic for Power7 processors having hyper-
threading disabled (cpu mask has gaps).
-- Fix of job priority ordering with sched/builtin and priority/multifactor.
-- CRAY - Fix for setting up the aprun for a large job (+2000 nodes).
Morris Jette
committed
-- Fix for race condition related to compute node boot resulting in node being
set down with reason of "Node <name> unexpectedly rebooted"
* Changes in SLURM 2.5.0
========================
-- Add DenyOnLimit flag for QOS to deny jobs at submission time if they
request resources that reach a 'Max' limit.
-- Permit SlurmUser or operator to change QOS of non-pending jobs (e.g.
running jobs).
-- BGQ - move initial poll to beginning of realtime interaction, which will
also cause it to run if the realtime server ever goes away.
-- Modify sbcast logic to survive slurmd daemon restart while file a
transmission is in progress.
-- Add retry logic to munge encode/decode calls. This is needed if the munge
deamon is under very heavy load (e.g. with 1000 slurmd daemons per compute
node).
-- Add launch and acct_gather_energy plugins to RPMs.
-- Restore support for srun "--mpi=list" option.
-- CRAY - Introduce step accounting for a Cray.
-- Modify srun to abandon I/O 60 seconds after the last task ends. Otherwise
an aborted slurmstepd can cause the srun process to hang indefinitely.
-- ENERGY - RAPL - alter code to close open files (and only open them once
where needed)
-- If the PrologSlurmctld fails, then requeue the job an indefinite number
of times instead of only one time.
-- Added Prolog and Epilog Guide (web page). Based upon work by Jason Sollom,
Cray Inc. and used by permission.
-- Restore gang scheduling functionality. Preemptor was not being scheduled.
Fix for bugzilla #3.
-- Add "cpu_load" to node information. Populate CPULOAD in node information
reported to Moab cluster manager.
-- Preempt jobs only when insufficient idle resources exist to start job,
regardless of the node weight.
-- Added priority/multifactor2 plugin based upon ticket distribution system.
Work by Janne Blomqvist, Aalto University.
-- Add SLURM_NODELIST to environment variables available to Prolog and Epilog.
-- Permit reservations to allow or deny access by account and/or user.
-- Add ReconfigFlags value of KeepPartState. See "man slurm.conf" for details.
-- Modify the task/cgroup plugin adding a task_pre_launch_priv function and
move slurmstepd outside of the step's cgroup. Work by Matthieu Hautreux.
-- Intel MIC processor support added using gres/mic plugin. BIG thanks to
Olli-Pekka Lehto, CSC-IT Center for Science Ltd.
-- Accounting - Change empty jobacctinfo structs to not actually be used
instead of putting 0's into the database we put NO_VALS and have sacct
figure out jobacct_gather wasn't used.
-- Cray - Prevent calling basil_confirm more than once per job using a flag.
-- Fix bug with topology/tree and job with min-max node count. Now try to
get max node count rather than minimizing leaf switches used.
-- Add AccountingStorageEnforce=safe option to provide method to avoid jobs
launching that wouldn't be able to run to completion because of a
GrpCPUMins limit.
-- Add support for RFC 5424 timestamps in logfiles. Disable with configuration
option of "--disable-rfc5424time". By Janne Blomqvist, Aalto University.
-- CRAY - Replace srun.pl with launch/aprun plugin to use srun to wrap the
aprun process instead of a perl script.
-- srun - Rename --runjob-opts to --launcher-opts to be used on systems other
than BGQ.
-- Added new DebugFlags - Energy for AcctGatherEnergy plugins.
-- start deprecation of sacct --dump --fdump
-- BGQ - added --verbose=OFF when srun --quiet is used
-- Added acct_gather_energy/rapl plugin to record power consumption by job.
Work by Yiannis Georgiou, Martin Perry, et. al., Bull
* Changes in SLURM 2.5.0.pre3
=============================
-- Add Google search to all web pages.
-- Add sinfo -T option to print reservation information. Work by Bill Brophy,
Bull.
-- Force slurmd exit after 2 minute wait, even if threads are hung.
-- Change node_req field in struct job_resources from 8 to 32 bits so we can
run more than 256 jobs per node.
-- sched/backfill: Improve accuracy of expected job start with respect to
reservations.
-- sinfo partition field size will be set the the length of the longest
partition name by default.
-- Make it so the parse_time will return a valid 0 if given epoch time and
set errno == ESLURM_INVALID_TIME_VALUE on error instead.
-- Correct srun --no-alloc logic when node count exceeds node list or task
task count is not a multiple of the node count. Work by Hongjia Cao, NUDT.
-- Completed integration with IBM Parallel Environment including POE and IBM's
NRT switch library.
* Changes in SLURM 2.5.0.pre2
=============================
-- When running with multiple slurmd daemons per node, enable specifying a
range of ports on a single line of the node configuration in slurm.conf.
-- Add reservation flag of Part_Nodes to allocate all nodes in a partition to
a reservation and automatically change the reservation when nodes are
added to or removed from the reservation. Based upon work by
Bill Brophy, Bull.
-- Add support for advanced reservation for specific cores rather than whole
nodes. Current limiations: homogeneous cluster, nodes idle when reservation
created, and no more than one reservation per node. Code is still under
development. Work by Alejandro Lucero Palau, et. al, BSC.
-- Add DebugFlag of Switch to log switch plugin details.
-- Correct job node_cnt value in job completion plugin when job fails due to
down node. Previously was too low by one.
-- Add new srun option --cpu-freq to enable user control over the job's CPU
frequency and thus it's power consumption. NOTE: cpu frequency is not
currently preserved for jobs being suspended and later resumed. Work by
Don Albert, Bull.
-- Add node configuration information about "boards" and optimize task
placement on minimum number of boards. Work by Rod Schultz, Bull.
* Changes in SLURM 2.5.0.pre1
=============================
-- Add new output to "scontrol show configuration" of LicensesUsed. Output is
"name:used/total"
-- Changed jobacct_gather plugin infrastructure to be cleaner and easier to
maintain.
-- Change license option count separator from "*" to ":" for consistency with
the gres option (e.g. "--licenses=foo:2 --gres=gpu:2"). The "*" will still
be accepted, but is no longer documented.
-- Permit more than 100 jobs to be scheduled per node (new limit is 250
-- Restructure of srun code to allow outside programs to utilize existing
logic.
* Changes in SLURM 2.4.6
========================
-- Correct WillRun authentication logic when issued for non-job owner.
-- BGQ - Fix to check block for action 'D' if it also has nodes in error.
* Changes in SLURM 2.4.5
========================
-- Cray - On job kill requeust, send SIGCONT, SIGTERM, wait KillWait and send
SIGKILL. Previously just sent SIGKILL to tasks.
-- BGQ - Fix issue when running srun outside of an allocation and only
specifying the number of tasks and not the number of nodes.
-- BGQ - validate correct ntasks_per_node
-- BGQ - when srun -Q is given make runjob be quiet
Morris Jette
committed
-- Modify use of OOM (out of memory protection) for Linux 2.6.36 kernel
or later. NOTE: If you were setting the environment variable
SLURMSTEPD_OOM_ADJ=-17, it should be set to -1000 for Linux 2.6.36 kernel
or later.
-- BGQ - Fix job step timeout actually happen when done from within an
allocation.
-- Reset node MAINT state flag when a reservation's nodes or flags change.
-- Accounting - Fix issue where QOS usage was being zeroed out on a
slurmctld restart.
-- BGQ - Add 64 tasks per node as a valid option for srun when used with
overcommit.
-- BLUEGENE - With Dynamic layout mode - Fix issue where if a larger block
was already in error and isn't deallocating and underlying hardware goes
bad one could get overlapping blocks in error making the code assert when
a new job request comes in.
-- BGQ - handle pending actions on a block better when trying to deallocate it.
-- Accounting - Fixed issue where if nodenames have changed on a system and
you query against that with -N and -E you will get all jobs during that
time instead of only the ones running on -N.
-- Accounting - If a job start message fails to the SlurmDBD reset the db_inx
so it gets sent again. This isn't a major problem since the start will
happen when the job ends, but this does make things cleaner.
-- If an salloc is waiting for an allocation to happen and is canceled by the
user mark the state canceled instead of completed.
-- Fix issue in accounting if a user puts a '\' in their job name.
-- Accounting - Fix for if asking for users or accounts that were deleted
with associations get the deleted associations as well.
-- BGQ - Handle shared blocks that need to be removed and have jobs running
on them. This should only happen in extreme conditions.
-- Fix inconsistency for hostlists that have more than 1 range.
-- BGQ - Add mutex around recovery for the Real Time server to avoid hitting
DB2 so hard.
-- BGQ - If an allocation exists on a block that has a 'D' action on it fail
job on future step creation attempts.
* Changes in SLURM 2.4.4
========================
-- BGQ - minor fix to make build work in emulated mode.
-- BGQ - Fix if large block goes into error and the next highest priority jobs
are planning on using the block. Previously it would fail those jobs
erroneously.
-- BGQ - Fix issue when a cnode going to an error (not SoftwareError) state
with a job running or trying to run on it.
-- Execute slurm_spank_job_epilog when there is no system Epilog configured.
-- Fix for srun --test-only to work correctly with timelimits
-- BGQ - If a job goes away while still trying to free it up in the
database, and the job is running on a small block make sure we free up
the correct node count.
-- BGQ - Logic added to make sure a job has finished on a block before it is
purged from the system if its front-end node goes down.
-- Modify strigger so that a filter option of "--user=0" is supported.
-- Correct --mem-per-cpu logic for core or socket allocations with multiple
threads per core.
-- Fix for older < glibc 2.4 systems to use euidaccess() instead of eaccess().
-- BLUEGENE - Do not alter a pending job's node count when changing it's
partition.
-- BGQ - Add functionality to make it so we track the actions on a block.
This is needed for when a free request is added to a block but there are
jobs finishing up so we don't start new jobs on the block since they will
fail on start.
-- BGQ - Fixed InactiveLimit to work correctly to avoid scenarios where a
user's pending allocation was started with srun and then for some reason
the slurmctld was brought down and while it was down the srun was removed.
-- Fixed InactiveLimit math to work correctly
-- BGQ - Add logic to make it so blocks can't use a midplane with a nodeboard
in error for passthrough.
-- BGQ - Make it so if a nodeboard goes in error any block using that midplane
for passthrough gets removed on a dynamic system.
-- BGQ - Fix for printing realtime server debug correctly.
-- BGQ - Cleaner handling of cnode failures when reported through the runjob
interface instead of through the normal method.
-- smap - spread node information across multiple lines for larger systems.
-- Cray - Defer salloc until after PrologSlurmctld completes.
-- Correction to slurmdbd communications failure handling logic, incorrect
error codes returned in some cases.
* Changes in SLURM 2.4.3
========================
-- Accounting - Fix so complete 32 bit numbers can be put in for a priority.
-- cgroups - fix if initial directory is non-existent SLURM creates it
correctly. Before the errno wasn't being checked correctly
-- BGQ - fixed srun when only requesting a task count and not a node count
to operate the same way salloc or sbatch did and assign a task per cpu
by default instead of task per node.
-- Fix salloc --gid to work correctly. Reported by Brian Gilmer
-- BGQ - fix smap to set the correct default MloaderImage
-- Close the batch job's environment file when it contains no data to avoid
leaking file descriptors.
-- Fix sbcast's credential to last till the end of a job instead of the
previous 20 minute time limit. The previous behavior would fail for
large files 20 minutes into the transfer.
-- Return ESLURM_NODES_BUSY rather than ESLURM_NODE_NOT_AVAIL error on job
submit when required nodes are up, but completing a job or in exclusive
job allocation.
-- Add HWLOC_FLAGS so linking to libslurm works correctly
-- BGQ - If using backfill and a shared block is running at least one job
and a job comes through backfill and can fit on the block without ending
jobs don't set an end_time for the running jobs since they don't need to
end to start the job.
-- Initialize bind_verbose when using task/cgroup.
-- BGQ - Fix for handling backfill much better when sharing blocks.
-- BGQ - Fix for making small blocks on first pass if not sharing blocks.
-- BLUEGENE - Remove force of default conn_type instead of leaving NAV
when none are requested. The Block allocator sets it up temporarily so
this isn't needed.
-- BLUEGENE - Fix deadlock issue when dealing with bad hardware if using
static blocks.
-- Fix to mysql plugin during rollup to only query suspended table when jobs
reported some suspended time.
-- Fix compile with glibc 2.16 (Kacper Kowalik)
-- BGQ - fix for deadlock where a block has error on it and all jobs
running on it are preemptable by scheduling job.
-- proctrack/cgroup: Exclude internal threads from "scontrol list pids".
Patch from Matthieu Hautreux, CEA.
-- Memory leak fixed for select/linear when preempting jobs.
-- Fix if updating begin time of a job to update the eligible time in
accounting as well.
-- BGQ - make it so you can signal steps when signaling the job allocation.
-- BGQ - Remove extra overhead if a large block has many cnode failures.
-- Priority/Multifactor - Fix issue with age factor when a job is estimated to
start in the future but is able to run now.
-- CRAY - update to work with ALPS 5.1
-- BGQ - Handle issue of speed and mutexes when polling instead of using the
realtime server.
-- BGQ - Fix minor sorting issue with sview when sorting by midplanes.
-- Accounting - Fix for handling per user max node/cpus limits on a QOS
correctly for current job.
-- Update documentation for -/+= when updating a reservation's
users/accounts/flags
-- Update pam module to work if using aliases on nodes instead of actual
host names.
-- Correction to task layout logic in select/cons_res for job with minimum
and maximum node count.
-- BGQ - Put final poll after realtime comes back into service to avoid
having the realtime server go down over and over again while waiting
for the poll to finish.
-- task/cgroup/memory - ensure that ConstrainSwapSpace=no is correctly
handled. Work by Matthieu Hautreux, CEA.
-- CRAY - Fix for sacct -N option to work correctly
-- CRAY - Update documentation to describe installation from rpm instead
or previous piecemeal method.
-- Fix sacct to work with QOS' that have previously been deleted.
-- Added all available limits to the output of sacctmgr list qos
* Changes in SLURM 2.4.2
========================
-- BLUEGENE - Correct potential deadlock issue when hardware goes bad and
there are jobs running on that hardware.
-- If job is submitted to more than one partition, it's partition pointer can
be set to an invalid value. This can result in the count of CPUs allocated
on a node being bad, resulting in over- or under-allocation of its CPUs.
Patch by Carles Fenoy, BSC.
-- Fix bug in task layout with select/cons_res plugin and --ntasks-per-node
option. Patch by Martin Perry, Bull.
-- BLUEGENE - remove race condition where if a block is removed while waiting
for a job to finish on it the number of unused cpus wasn't updated
correctly.
-- BGQ - make sure we have a valid block when creating or finishing a step
allocation.
-- BLUEGENE - If a large block (> 1 midplane) is in error and underlying
hardware is marked bad remove the larger block and create a block over
just the bad hardware making the other hardware available to run on.
-- BLUEGENE - Handle job completion correctly if an admin removes a block
where other blocks on an overlapping midplane are running jobs.
-- BLUEGENE - correctly remove running jobs when freeing a block.
-- BGQ - correct logic to place multiple (< 1 midplane) steps inside a
multi midplane block allocation.
-- BGQ - Make it possible for a multi midplane allocation to run on more
than 1 midplane but not the entire allocation.
-- BGL - Fix for syncing users on block from Tim Wickberg
-- Fix initialization of protocol_version for some messages to make sure it
is always set when sending or receiving a message.
Alejandro Lucero Palau
committed
-- Reset backfilled job counter only when explicitly cleared using scontrol.
Patch from Alejandro Lucero Palau, BSC.
-- BLUEGENE - Fix for handling blocks when a larger block will not free and
while it is attempting to free underlying hardware is marked in error
making small blocks overlapping with the freeing block. This only
applies to dynamic layout mode.
-- Cray and BlueGene - Do not treat lack of usable front-end nodes when
slurmctld deamon starts as a fatal error. Also preserve correct front-end
node for jobs when there is more than one front-end node and the slurmctld
daemon restarts.
Morris Jette
committed
-- Correct parsing of srun/sbatch input/output/error file names so that only
the name "none" is mapped to /dev/null and not any file name starting
with "none" (e.g. "none.o").
-- BGQ - added version string to the load of the runjob_mux plugin to verify
the current plugin has been loaded when using runjob_mux_refresh_config
-- CGROUPS - Use system mount/umount function calls instead of doing fork
exec of mount/umount from Janne Blomqvist.
-- BLUEGENE - correct start time setup when no jobs are blocking the way
from Mark Nelson
-- Fixed sacct --state=S query to return information about suspended jobs
current or in the past.
-- FRONTEND - Made error warning more apparent if a frontend node isn't
configured correctly.
-- BGQ - update documentation about runjob_mux_refresh_config which works
correctly as of IBM driver V1R1M1 efix 008.
* Changes in SLURM 2.4.1
========================
-- Fix bug for job state change from 2.3 -> 2.4 job state can now be preserved
correctly when transitioning. This also applies for 2.4.0 -> 2.4.1, no
state will be lost. (Thanks to Carles Fenoy)
Loading
Loading full blame...