Quality of Service (QOS)
One can specify a Quality of Service (QOS) for each job submitted to Slurm. The QOSs are defined in the Slurm database using the sacctmgr command. Jobs request a QOS using the "--qos=" option to the sbatch, salloc, and srun commands.
Contents
Effects on Jobs
The QOS associated with a job will affect the job in three key ways: scheduling priority, preemption, and resource limits.
Job Scheduling Priority
Job scheduling priority is made up of a number of factors as described in the priority/multifactor plugin. One of the factors is the QOS priority. Each QOS is defined in the Slurm database and includes an associated priority. Jobs that request and are permitted a QOS will incorporate the priority associated with that QOS in the job's multi-factor priority calculation.
To enable the QOS priority component of the multi-factor priority calculation, the "PriorityWeightQOS" configuration parameter must be defined in the slurm.conf file and assigned an integer value greater than zero.
A job's QOS only affects is scheduling priority when the multi-factor plugin is loaded.
Job Preemption
Slurm offers two ways for a queued job to preempt a running job, free-up the running job's resources and allocate them to the queued job. See the Preemption description for details.
The preemption method is determined by the "PreemptType" configuration parameter defined in slurm.conf. When the "PreemptType" is set to "preempt/qos", a queued job's QOS will be used to determine whether it can preempt a running job. It is important to note that the QOS used to determine if a job is eligible for preemption is the QOS associated with the job and not a Partition QOS.
The QOS can be assigned (using sacctmgr) a list of other QOSs that it can preempt. When there is a queued job with a QOS that is allowed to preempt a running job of another QOS, the Slurm scheduler will preempt the running job.
The QOS option PreemptExemptTime specifies the minimum run time before the job is considered for preemption. The QOS option takes precedence over the global option of the same name. A Partition QOS with PreemptExemptTime takes precedence over a job QOS with PreemptExemptTime, unless the job QOS has the OverPartQOS flag enabled.
Resource Limits
Each QOS is assigned a set of limits which will be applied to the job. The limits mirror the limits imposed by the user/account/cluster/partition association defined in the Slurm database and described in the Resource Limits page. When limits for a QOS have been defined, they will take precedence over the association's limits.
Partition QOS
A QOS can be attached to a partition. This means the partition will have all
the same limits as the QOS. This does not associate jobs with the QOS, nor does
it give the job any priority or preemption characteristics of the assigned QOS.
Jobs may separately request the same QOS or a different QOS to gain those
characteristics. However, the Partition QOS limits will override the job's QOS.
If the opposite is desired you may configure the job's QOS with
Flags=OverPartQOS
which will reverse the order of precedence.
This functionality may be used to implement a true "floating"
partition, in which a partition may access a limited amount of resources with no
restrictions on which nodes it uses to get the resources. This is accomplished
by assigning all nodes to the partition, then configuring a Partition QOS with
GrpTRES
set to the desired resource limits.
NOTE: Most QOS attributes are set using the sacctmgr command. However, setting a QOS as a partition QOS is accomplished in slurm.conf through the QOS= option in the configuration of the associated partition. The QOS should be created using sacctmgr before it is assigned as a partition QOS.
Relative QOS
Starting in Slurm 23.11, a QOS may be configured to contain relative resource
limits instead of absolute limits by setting Flags=Relative
.
When this flag is set, all resource limits are treated as percentages of the
total resources available. Values higher than 100 are interpreted as 100%.
Memory limits should be set with no units. Although the default units (MB) will
be displayed, the limits will be enforced as a percentage (1MB = 1%).
NOTE: When Flags=Relative is added to a QOS, slurmctld must be restarted or reconfigured for the flag to take effect.
Generally, the limits on a relative QOS will be calculated relative to the
resources in the whole cluster. For example, cpu=50
would be
interpreted as 50% of all CPUs in the cluster.
However, when a relative QOS is also assigned as a partition QOS, some unique conditions will apply:
- Limits will be calculated relative to the partition's resources;
for example,
cpu=50
would be interpreted as 50% of all CPUs in the associated partition. - Only one partition may have this QOS as its partition QOS.
- Jobs will not be allowed to use it as a normal QOS.
NOTE: To avoid unexpected job submission errors, it is recommended not to add a relative partition QOS to any association-based entities.
Other QOS Options
- Flags Used by the slurmctld to override or enforce certain
characteristics. To clear a previously set value use the modify command with a
new value of -1.
Valid options are:- DenyOnLimit If set, jobs using this QOS will be rejected at submission time if they do not conform to the QOS 'Max' limits as stand-alone jobs. Jobs that go over these limits when other jobs are considered, but conform to the limits when considered individually will not be rejected. Instead they will pend until resources are available (as by default without DenyOnLimit). Group limits (e.g. GrpTRES) will also be treated like 'Max' limits (e.g. MaxTRESPerNode) and jobs will be denied if they would violate the limit as stand-alone jobs. This currently only applies to QOS and Association limits.
- EnforceUsageThreshold If set, and the QOS also has a UsageThreshold, any jobs submitted with this QOS that fall below the UsageThreshold will be held until their Fairshare Usage goes above the Threshold.
- NoDecay If set, this QOS will not have its GrpTRESMins, GrpWall and UsageRaw decayed by the slurm.conf PriorityDecayHalfLife or PriorityUsageResetPeriod settings. This allows a QOS to provide aggregate limits that, once consumed, will not be replenished automatically. Such a QOS will act as a time-limited quota of resources for an association that has access to it. Account/user usage will still be decayed for associations using the QOS. The QOS GrpTRESMins and GrpWall limits can be increased or the QOS RawUsage value reset to 0 (zero) to again allow jobs submitted with this QOS to run (if pending with QOSGrp{TRES}MinutesLimit or QOSGrpWallLimit reasons, where {TRES} is some type of trackable resource).
- NoReserve If this flag is set and backfill scheduling is used, jobs using this QOS will not reserve resources in the backfill schedule's map of resources allocated through time. This flag is intended for use with a QOS that may be preempted by jobs associated with all other QOS (e.g use with a "standby" QOS). If this flag is used with a QOS which can not be preempted by all other QOS, it could result in starvation of larger jobs.
- OverPartQOS If set, jobs using this QOS will be able to override any limits used by the requested partition's QOS limits.
- PartitionMaxNodes If set, jobs using this QOS will be able to override the requested partition's MaxNodes limit.
- PartitionMinNodes If set, jobs using this QOS will be able to override the requested partition's MinNodes limit.
- PartitionTimeLimit If set, jobs using this QOS will be able to override the requested partition's TimeLimit.
- Relative If set, the QOS limits will be treated as percentages of
the cluster or partition instead of absolute limits (numbers should be less than
100). The controller should be restarted or reconfigured after adding the
Relative flag to the QOS.
If this is used as a partition QOS:- Limits will be calculated relative to the partition's resources.
- Only one partition may have this QOS as its partition QOS.
- Jobs will not be allowed to use it as a normal QOS.
- RequiresReservation If set, jobs using this QOS must designate a reservation when submitting a job. This option can be useful in restricting usage of a QOS that may have greater preemptive capability or additional resources to be allowed only within a reservation.
- UsageFactorSafe If set, and AccountingStorageEnforce includes Safe, jobs will only be able to run if the job can run to completion with the UsageFactor applied.
- GraceTime Preemption grace time to be extended to a job which has been selected for preemption.
UsageFactor A float that is factored into a job's TRES usage (e.g. RawUsage, TRESMins, TRESRunMins). For example, if the usagefactor was 2, for every TRESBillingUnit second a job ran it would count for 2. If the usagefactor was .5, every second would only count for half of the time. A setting of 0 would add no timed usage from the job.
- UsageThreshold A float representing the lowest fairshare of an association allowable to run a job. If an association falls below this threshold and has pending jobs or submits new jobs those jobs will be held until the usage goes back above the threshold. Use sshare to see current shares on the system.
The usage factor only applies to the job's QOS and not the partition QOS.
If the UsageFactorSafe flag is set and AccountingStorageEnforce includes Safe, jobs will only be able to run if the job can run to completion with the UsageFactor applied.
If the UsageFactorSafe flag is not set and AccountingStorageEnforce includes Safe, a job will be able to be scheduled without the UsageFactor applied and will be able to run without being killed due to limits.
If the UsageFactorSafe flag is not set and AccountingStorageEnforce does not include Safe, a job will be able to be scheduled without the UsageFactor applied and could be killed due to limits.
See AccountingStorageEnforce in slurm.conf man page.
Default is 1. To clear a previously set value use the modify command with a new value of -1.
Configuration
To summarize the above, the QOSs and their associated limits are defined in the Slurm database using the sacctmgr utility. The QOS will only influence job scheduling priority when the multi-factor priority plugin is loaded and a non-zero "PriorityWeightQOS" has been defined in the slurm.conf file. The QOS will only determine job preemption when the "PreemptType" is defined as "preempt/qos" in the slurm.conf file. Limits defined for a QOS (and described above) will override the limits of the user/account/cluster/partition association.
QOS examples
QOS manipulation examples. All QOS operations are done using the sacctmgr command. The default output of 'sacctmgr show qos' is very long given the large number of limits and options available so it is best to use the format option which filters the display.
By default when a cluster is added to the database a default qos named normal is created.
$ sacctmgr show qos format=name,priority Name Priority ---------- ---------- normal 0
Add a new QOS
$ sacctmgr add qos zebra Adding QOS(s) zebra Settings Description = QOS Name $ sacctmgr show qos format=name,priority Name Priority ---------- ---------- normal 0 zebra 0
Set QOS priority
$ sacctmgr modify qos zebra set priority=10 Modified qos... zebra $ sacctmgr show qos format=name,priority Name Priority ---------- ---------- normal 0 zebra 10
Set some other limits
$ sacctmgr modify qos zebra set GrpTRES=cpu=24 Modified qos... zebra $ sacctmgr show qos format=name,priority,GrpTRES Name Priority GrpTRES ---------- ---------- ------------- normal 0 zebra 10 cpu=24
Add a QOS to a user account
$ sacctmgr modify user crock set qos=zebra $ sacctmgr show assoc format=cluster,user,qos Cluster User QOS ---------- ---------- -------------------- canis_major normal canis_major root normal canis_major normal canis_major crock zebra
Users can belong to multiple QOSs
$ sacctmgr modify user crock set qos+=alligator $ sacctmgr show assoc format=cluster,user,qos Cluster User QOS ---------- ---------- -------------------- canis_major normal canis_major root normal canis_major normal canis_major crock alligator,zebra
Finally, delete a QOS
$ sacctmgr delete qos alligator Deleting QOS(s)... alligator
Last modified 22 April 2023