Open topic with navigation
You are here: 7.0 Controlling Resource Access - Reservations, Partitions, and QoS Facilities > QoS Facilities
7.3 Quality of Service (QoS) Facilities
This section describes how to do the following:
Moab's QoS facility allows a site to give special treatment to various classes of jobs, users, groups, and so forth. Each QoS object can be thought of as a container of special privileges ranging from fairness policy exemptions, to special job prioritization, to special functionality access. Each QoS object also has an extensive access list of users, groups, and accounts that can access these privileges.
Sites can configure various QoS's each with its own set of priorities, policy exemptions, and special resource access settings. They can then configure user, group, account, and class access to these QoS's. A given job will have a default QoS and may have access to several additional QoS's. When the job is submitted, the submitter may request a specific QoS or just allow the default QoS to be used. Once a job is submitted, a user may adjust the QoS of the job at any time using the setqos command. The setqos command will only allow the user to modify the QoS of that user's jobs and only change the QoS to a QoS that this user has access to. Moab administrators may change the QOS of any job to any value.
Jobs can be granted access to QoS privileges if the QoS is listed in the system default configuration QDEF (QoS default) or QLIST (QoS access list), or if the QoS is specified in the QDEF or QLIST of a user, group, account, or class associated with that job. Alternatively, a user may access QoS privileges if that user is listed in the QoS's MEMBERULIST attribute.
The mdiag -q command can be used to obtain information about the current QoS configuration including specified credential access.
The privileges enabled via QoS settings may be broken into the following categories:
All privileges are managed via the QOSCFG parameter.
|FSTARGET||Specifies QoS fairshare target.|
|FSWEIGHT||Sets QoS fairshare weight offset affecting a job's fairshare priority component.|
|PRIORITY||Assigns priority to all jobs requesting particular QoS.|
|QTTARGET||Sets QoS queuetime target affecting a job's target priority component and QoS delivered.|
|QTWEIGHT||Sets QoS queuetime weight offset affecting a job's service priority component.|
|XFTARGET||Sets QoS XFactor target affecting a job's target priority component and QoS delivered.|
|XFWEIGHT||Sets QoS XFactor weight offset affecting a job's service priority component.|
# assign priority for all qos geo jobs QOSCFG[geo] PRIORITY=10000
The QoS facility can be used to enable special services and to disable default services. These services are enabled/disabled by setting the QoS QFLAGSattribute.
|DEADLINE||Job may request an absolute or relative completion deadline and Moab will reserve resources to meet that deadline. (An alternative priority based deadline behavior is discussed in the PRIORITY FACTORS section.)|
|DEDICATED||Moab dedicates all resources of an allocated node to the job meaning that the job will not share a node's compute resources with any other job.|
|ENABLEUSERRSV||Allow user or personal reservations to be created and managed.|
|IGNALL||Scheduler ignores all resource usage policies for jobs associated with this QoS.|
Specifies how Moab should track the dynamic aspects of a job's priority. The two valid values are ACCRUE and RESET.
ACCRUE indicates that the job will accrue queuetime based priority from the time it is submitted unless it violates any of the policies not specified in JOBPRIOEXCEPTIONS.
RESET indicates that it will accrue priority from the time it is submitted unless it violates any of the JOBPRIOEXCEPTIONS. However, with RESET, if the job does violate JOBPRIOEXCEPTIONS then its queuetime based priority will be reset to 0.
The following old JOBPRIOACCRUALPOLICY values have been deprecated and should be adjusted to the following values:
Specifies exceptions for calculating a job's dynamic priority (QUEUETIME, XFACTOR, TARGETQUEUETIME). Valid values are a comma delimited list of any of the following: DEFER, DEPENDS, SOFTPOLICY, HARDPOLICY, IDLEPOLICY, USERHOLD, BATCHHOLD, and SYSTEMHOLD (ALL or NONE can also be specified on their own).
Normally, when a job violates a policy, is placed on hold, or has an unsatisfied dependency, it will not accrue priority. Exceptions can be configured to allow a job to accrue priority inspite of any of these violations. With DEPENDS a job will increase in priority even if there exists an unsatisfied dependency. With SOFTPOLICY, HARDPOLICY, or IDLEPOLICY a job can accrue priority despite violating a specific limit. With DEFER, USERHOLD, BATCHHOLD, or SYSTEMHOLD a job can accrue priority despite being on hold.
|NOBF||Job is not considered for backfill.|
|NORESERVATION||Job should never reserve resources regardless of priority.|
|NTR||Job is prioritized to run next and backfill is disabled to prevent other jobs from jumping in front of ones with the NTR flag.
|PREEMPTCONFIG||User jobs may specify options to alter how preemption impacts the job such as minpreempttime.|
|PREEMPTEE||Job may be preempted by higher priority PREEMPTOR jobs.|
|PREEMPTFSV||Job may be preempted by higher priority PREEMPTOR jobs if it exceeds its fairshare target when started.|
|PREEMPTOR||Job may preempt lower priority PREEMPTEE jobs.|
|PREEMPTSPV||Job may be preempted by higher priority PREEMPTOR jobs if it currently violates a soft usage policy limit.|
|PROVISION||If the job cannot locate available resources with the needed OS or software, the scheduler may provision a number of nodes to meet the needed OS or software requirements.|
|RESERVEALWAYS||Job should create resource reservation regardless of job priority.|
|RUNNOW||Boosts a job's system priority and makes the job a preemptor.
|TRIGGER||The job is able to directly specify triggers.|
|USERESERVED[:<RSVID>]||Job may only use resources within accessible reservations. If <RSVID> is specified, job may only use resources within the specified reservation.|
Example: For lowprio QoS job, disable backfill and make job preemptible
Example: Bind all jobs to chemistry reservation
Other QoS Attributes
In addition to the flags, there are attributes that alter service access.
|RSVCREATIONCOST||The base cost to users when creating reservations.|
|SYSPRIO||Sets the system priority on jobs associated with this QoS.|
Example: All jobs submitted under a QoS sample receive a system priority of 1
If desired, jobs associated with a particular QoS can be locked into a reservation or reservation group using the REQRID attribute. For example, to force jobs using QoS jasper to only use the resources within the failsafestanding reservation, use the following:
QOSCFG[jasper] REQRID=failsafe ...
Minimum network service levels can be specified using the MINBANDWIDTH and MAXLATENCY attributes. If specified, Moab does not allow a job to start until these criteria can be satisfied and allocates or reserves specific resources or combinations of resources to fulfill these constraints. Further, if the Network Management System supports it, Moab dynamically creates network fabric partitions with guaranteed service levels.
QOSCFG[highspeed] MINBANDWIDTH=1GB MAXLATENCY=10us ...
All credentials, including QoS, allow specification of job usage limits as described in the Basic Fairness Policies overview. In such cases, jobs are constrained by the most limiting of all applicable policies. With QoS's, an override limit may also be specified and with this limit, jobs are constrained by the override, regardless of other limits specified.The following parameters can override the throttling policies from other credentials:
OMAXJOB, OMAXNODE, OMAXPE, OMAXPROC, OMAXPS, OMAXJPROC, OMAXJPS, OMAXJWC, and OMAXJNODE.
# staff QoS should have a limit of 48 jobs, ignoring the user limit USERCFG[DEFAULT] MAXJOB=10 QOSCFG[staff] OMAXJOB=48
Jobs can be granted access to services such as preemption and reservation creation, and they can be granted access to resource reservations. However, with QoS thresholds, this access can be made conditional on the current queuetime and XFactor metrics of an idle job. The following table lists the available QoS service thresholds:
|PREEMPTQTTHRESHOLD||A job with this QoS becomes a preemptor if the specified queuetime threshold is reached.|
|PREEMPTXFTHRESHOLD||A job with this QoS becomes a preemptor if the specified XFactor threshold is reached.|
|RSVQTTHRESHOLD||A job with this QoS can create a job reservation to guarantee resource access if the specified queuetime threshold is reached.|
|RSVXFTHRESHOLD||A job with this QoS can create a job reservation to guarantee resource access if the specified XFactor threshold is reached.|
|ACLQTTHRESHOLD||A job with this QoS can access reservations with a corresponding QoS ACL only if the specified queuetime threshold is reached.|
|ACLXFTHRESHOLD||A job with this QoS can access reservations with a corresponding QoS ACL only if the specified XFactor threshold is reached.|
|TRIGGERQTTHRESHOLD||If a job with this QoS fails to run before this threshold is reached, any failure triggers associated with this QoS will fire.|
Job preemption facilities can be controlled on a per-QoS basis using the PREEMPTEE and PREEMPTOR flags. Jobs that are preemptible can optionally be constrained to only be preempted in a particular manner by specifying the QoS PREEMPTPOLICY attribute as in the following example:
QOSCFG[special] QFLAGS=PREEMPTEE PREEMPTPOLICY=CHECKPOINT
For preemption to be effective, a job must be marked as a preemptee and must be enabled for the requested preemption type. For example, if the PREEMPTPOLICY is set to suspend, a potential target job must be both a preemptee and marked with the job flag SUSPENDABLE. (See suspension for more information.) If the target job is not suspendable, it will be either requeued or canceled. Likewise, if the PREEMPTPOLICY is set to requeue, the job will be requeued if it is marked restartable. Otherwise, it will be canceled.
The minimum time a job must run before being considered eligible for preemption can also be configured on a per-QoS basis using the PREEMPTMINTIME parameter, which is analogous to the JOBPREEMPTMINACTIVETIME. Conversely, PREEMPTMAXTIME sets a threshold for which a job is no longer eligible for preemption; see JOBPREEMPTMAXACTIVETIME for analogous details.
The PREEMPTEES attribute allows you to specify which QoSes that a job in a specific QoS is allowed to preempt. The PREEMPTEES list is a comma-delimited list of QOSIDs. When a PREEMPTEES attribute is specified, a job using that QoS can only preempt jobs using QoSes listed in the PREEMPTEES list. In turn, those QoSess must be flagged as PREEMTEEs as in the following example:
QOSCFG[a] QFLAGS=PREEMPTOR PREEMPTEES=b,c QOSCFG[b] QFLAGS=PREEMPTEE QOSCFG[c] QFLAGS=PREEMPTEE
In the example, jobs in the 'a' QoS can only preempt jobs in the "b" and "c" QoSes.
You can define the privileges allowed within a QoS by using the QOSCFGparameter; however, in most cases access to the QoS is enabled via credential specific *CFG parameters, specifically the USERCFG, GROUPCFG, ACCOUNTCFG, and CLASSCFG parameters, which allow defining QoS access lists and QoS defaults. Specify credential specific QoS access by using the QLIST and/or QDEF attributes of the associated credential parameter.
To enable QoS access, the QLIST and/or QDEF attributes of the appropriate user, group, account, or class/queue should be specified as in the following example:
# user john's jobs can access QOS geo, chem, or staff with geo as default USERCFG[john] QDEF=geo QLIST=geo,chem,staff # group system jobs can access the development qos GROUPCFG[systems] QDEF=development # class batch jobs can access the normal qos CLASSCFG[batch] QDEF=normal
By default, jobs may request a QoS if access to that QoS is allowed by any of the job's credentials. (In the previous example, a job from user john submitted to the class batch could request QoS's geo, chem, staff, or normal).
If desired, QoS access can be masked or logically ANDed if the QoS access list is specified with a terminating ampersand (&) as in the following example:
# user john's jobs can access QOS geo, chem, or staff with geo as default USERCFG[john] QDEF=geo QLIST=geo,chem,staff # group system jobs can access the development qos GROUPCFG[systems] QDEF=development # class batch jobs can access the normal qos CLASSCFG[batch] QDEF=normal # class debug jobs can only access the development or lowpri QoS's regardless of other credentials CLASSCFG[debug] QLIST=development,lowpri&
QoS access may also be specified from within the QoS object using the QoS MEMBERULIST attribute as in the following example:
|By default, if a job requests a QoS that it cannot access, Moab places a hold on that job. The QOSREJECTPOLICY can be used to modify this behavior.|
Internal charging is being deprecated. It is anticipated that internal charging will no longer be supported by Moab by mid-year 2012.
The following policies are affected:
Charging policies and rates may be configured on a per-QoS basis. This capability provides a powerful way of allowing end-users to select the level of service they desire and the price they are willing to pay for that service.
Several policies allow control over exactly how a job gets charged; these include CHARGERATEPOLICY and CHARGEMETRICPOLICY. The first parameter, CHARGERATEPOLICY, allows a site to specify how jobs are charged for the resources they use.
A job's total cost can be adjusted on a per QoS basis using the DEDRESCOST attribute:
|DEDRESCOST||Cost multiplier for dedicated resources associated with using the specified QoS.|
QOSCFG[fast] DEDRESCOST=4.0 QOSCFG[slow] DEDRESCOST=0.5 ...
|These charging parameters only impact internal charging job costs written to accounting records. They do not currently affect charges sent to an external allocation manager.|
By default, jobs inherit a default QoS based on the user, group, class, and account associated with the job. If a job has access to multiple QoS levels, the submittor can explicitly request a particular QoS using the QoS resource manager extension as in the following example:
> msub -l nodes=1,walltime=100,qos=special3 job.cmd
By default, Moab allows all users access to special attributes such as node access policy. By enabling the QoS facility SPECATTRS, the access to these policies can be restricted. For example, to enable the facility, in the moab.cfg file, specify QOSCFG[DEFAULT] SPECATTRS=. Then, to allow access to the special attributes, indicate which special attributes a specific QoS may access.
QOSCFG[DEFAULT] SPECATTRS= QOSCFG[high] SPECATTRS=NACCESSPOLICY
Copyright © 2012 Adaptive Computing Enterprises, Inc.®