Differences between revisions 31 and 32
Revision 31 as of 2018-11-19 12:08:38
Size: 10719
Comment: QOS factor
Revision 32 as of 2019-07-15 08:29:56
Size: 11482
Comment: Preemption of jobs
Deletions are marked like this. Additions are marked like this.
Line 85: Line 85:
Preemption of jobs by high-priority jobs
----------------------------------------

Slurm supports job preemption_, the act of stopping one or more "low-priority" jobs to let a "high-priority" job run.
Job preemption_ is implemented as a variation of Slurm's Gang Scheduling logic.
When a high-priority job has been allocated resources that have already been allocated to one or more low priority jobs, the low priority job(s) are preempted.
The low priority job(s) can resume once the high priority job completes.
Alternately, the low priority job(s) can be requeued and started using other resources if so configured in newer versions of Slurm.

Preemption is configured in slurm.conf_.

.. _preemption: https://slurm.schedmd.com/preempt.html

Slurm job scheduler

Jump to our Slurm top-level page.

Prerequisites

Before configuring the Multifactor_Priority_Plugin scheduler, you must first configure Slurm_accounting.

Scheduler configuration

The SchedulerType configuration parameter controls how queued jobs are executed. See the Scheduling_Configuration_Guide.

Backfill scheduler

We use the backfill scheduler in slurm.conf:

SchedulerType=sched/backfill
SchedulerParameters=kill_invalid_depend,defer,bf_continue

but there are some backfill parameters that should be considered (see slurm.conf), for example:

...bf_interval=60,bf_max_job_start=20,bf_resolution=600,bf_window=11000

The importance of bf_window is explained as:

  • The default value is 1440 minutes (one day). A value at least as long as the highest allowed time limit is generally advisable to prevent job starvation. In order to limit the amount of data managed by the backfill scheduler, if the value of bf_window is increased, then it is generally advisable to also increase bf_resolution.

So you must configure bf_window according to the longest possible MaxTime in all partitions in slurm.conf:

PartitionName= ... MaxTime=XXX

scontrol top command

The scontrol top job_list command is documented as:

Move the specified job IDs to the top of the queue of jobs belonging to the identical user ID, partition name, account, and QOS.
The job_list argument is a comma separated ordered list of job IDs.
Any job not matching all of those fields will not be effected.
Only jobs submitted to a single partition will be effected.
This operation changes the order of jobs by adjusting job nice values.
The net effect on that user's throughput will be negligible to slightly negative.
This operation is disabled by default for non-privileged (non-operator, admin, SlurmUser, or root) users.
This operation may be enabled for non-privileged users by the system administrator by including the option "enable_user_top" in the SchedulerParameters configuration parameter.

While scontrol top job_list may be useful for the superuser to help with user requests, it is not recommended to configure SchedulerParameters=enable_user_top. The Slurm 17.11 news page (https://slurm.schedmd.com/news.html) highlights this change:

Regular user use of "scontrol top" command is now disabled.
Use the configuration parameter "SchedulerParameters=enable_user_top" to enable that functionality.
The configuration parameter "SchedulerParameters=disable_user_top" will be silently ignored.

There does not seem to be any documentation of why the scontrol top job_list is unwarranted, but we have made the following observations of a bad side effect:

  • A user requests a high priority for a job, and the superuser grants a negative nice value with scontrol update jobid=10208 nice=-10000.
  • The user can now assign a negative nice value to his other jobs with scontrol top jobid=10209,10210, thereby jumping ahead of normal jobs in the queue.

Preemption of jobs by high-priority jobs

Slurm supports job preemption, the act of stopping one or more "low-priority" jobs to let a "high-priority" job run. Job preemption is implemented as a variation of Slurm's Gang Scheduling logic. When a high-priority job has been allocated resources that have already been allocated to one or more low priority jobs, the low priority job(s) are preempted. The low priority job(s) can resume once the high priority job completes. Alternately, the low priority job(s) can be requeued and started using other resources if so configured in newer versions of Slurm.

Preemption is configured in slurm.conf.

Multifactor Priority Plugin scheduler

A sophisticated Multifactor_Priority_Plugin provides a very versatile facility for ordering the queue of jobs waiting to be scheduled. See the PriorityXXX parameters in the slurm.conf file.

Multifactor configuration

The Fairshare is configured with PriorityX parameters in the Configuration section of the Multifactor_Priority_Plugin page, also documented in the slurm.conf page:

  • PriorityType
  • PriorityDecayHalfLife
  • PriorityCalcPeriod
  • PriorityUsageResetPeriod
  • PriorityFavorSmall
  • PriorityMaxAge
  • PriorityWeightAge
  • PriorityWeightFairshare
  • PriorityWeightJobSize
  • PriorityWeightPartition
  • PriorityWeightQOS
  • PriorityWeightTRES

An example slurm.conf fairshare configuration may be:

PriorityType=priority/multifactor
PriorityDecayHalfLife=7-0
PriorityFavorSmall=NO
PriorityMaxAge=10-0
PriorityWeightAge=100000
PriorityWeightFairshare=1000000
PriorityWeightJobSize=100000
PriorityWeightPartition=100000
PriorityWeightQOS=100000

PropagateResourceLimitsExcept=MEMLOCK
PriorityFlags=ACCRUE_ALWAYS,FAIR_TREE
AccountingStorageEnforce=associations,limits,qos,safe

PriorityWeightXXX values are all 32-bit integers. The final Job Priority is a 32-bit integer.

IMPORTANT: Set PriorityWeight values high to generate wide range of job priorities.

Fairshare

Several Fairshare Algorithms may be used:

The Fairshare algorithm is explained in the Multifactor_Priority_Plugin page. Note the meaning of the special value fairshare=parent:

  • If all users in an account are configured with:

    FairShare=parent

    the result is that all the jobs drawing from that account will get the same fairshare priority, based on the accounts total usage. No additional fairness is added based on users individual usage.

Quality of Service (QOS)

One can specify a Quality of Service (QOS) for each job submitted to Slurm. A description and example are in the QOS page.

Example:

sacctmgr show qos format=name,priority

To enforce user jobs to have a QOS you must (at least) have:

AccountingStorageEnforce=qos

see the slurm.conf and Resource_Limits documents. The AccountingStorageEnforce options include:

  • associations - This will prevent users from running jobs if their association is not in the database. This option will prevent users from accessing invalid accounts.
  • limits - This will enforce limits set to associations. By setting this option, the 'associations' option is also set.
  • qos - This will require all jobs to specify (either overtly or by default) a valid qos (Quality of Service). QOS values are defined for each association in the database. By setting this option, the 'associations' option is also set.
  • safe - limits and associations will automatically be set.

The Quality of Service (QOS) Factor is defined in the Multifactor_Priority_Plugin page as:

Each QOS can be assigned an integer priority. The larger the number, the greater the job priority will be for jobs that request this QOS. This priority value is then normalized to the highest priority of all the QOS's to become the QOS factor.

A non-zero weight must be defined in slurm.conf, for example:

PriorityWeightQOS=100000

Resource Limits

To enable any limit enforcement you must at least have:

AccountingStorageEnforce=limits

in your slurm.conf, otherwise, even if you have limits set, they will not be enforced. Other options for AccountingStorageEnforce and the explanation for each are found on the Resource_Limits document.

Limiting (throttling) jobs in the queue

It is desirable to prevent individual users from flooding the queue with jobs, in case there are idle nodes available, because it may block future jobs by other users. Note:

With Slurm it appears that the only way to achieve user job throttling is the following:

  • Using the GrpTRESRunMins parameter defined in the Resource_Limits document. See also the TRES definition.

  • The GrpTRESRunMins limits can be applied to associations (accounts or users) as well as QOS. Set the limit by:

    sacctmgr modify association where name=XXX set GrpTRESRunMin=cpu=1000000   # For an account/user asociation
    sacctmgr modify qos where name=some_QOS set GrpTRESRunMin=cpu=1000000      # For a QOS

Partition factor priority

If some partition XXX (for example big memory nodes) should have a higher priority, this is explained in Multifactor_Priority_Plugin by:

(PriorityWeightPartition) * (partition_factor) +

The Partition factor is controlled in slurm.conf, for example:

PartitionName=XXX ... PriorityJobFactor=10
PriorityWeightPartition=1000

Scheduling commands

View scheduling information for the Multifactor_Priority_Plugin by the commands:

  • sprio - view the factors that comprise a job's scheduling priority:

    sprio     # List job priorities
    sprio -l  # List job priorities including username etc.
    sprio -w  # List weight factors used by the multifactor scheduler
  • sshare - Tool for listing the shares of associations to a cluster:

    sshare
    sshare -l    # Long listing with additional information
    sshare -a    # Listing with also user information
  • sdiag - Scheduling diagnostic tool for Slurm

Niflheim: Slurm_scheduler (last edited 2019-07-15 08:29:56 by OleHolmNielsen)