|SGE_PRIORITY(5)||Grid Engine File Formats||SGE_PRIORITY(5)|
sge_priority - Grid Engine job priorities
Grid Engine provides a means for controlling job dispatch and run-time priorities. The dispatch priority indicates the relative importance of pending jobs and determines the order in which Grid Engine initially considers dispatching jobs to queue instances, if the relevant resources are available. (The actual running order may be influenced may be influenced by other factors, such as reservations.) The run-time priority determines the CPU allocation that the operating system assigns to jobs.
JOBS DISPATCH PRIORITY¶
A job's dispatch priority (display in the "reduced" output format of qstat(1)), is affected by a number of factors:
- the identity of the submitting user;
- the project under which the job is submitted (or alternatively, the default project of the submitting user);
- any resources requested by the job;
- the job's submit time;
- the job's initiation deadline time (if specified);
- the -p priority specified for the job (also known as the POSIX priority "pprio").
The effect of each of these is governed by the overall policy setup, which is split into three top-level contributions. Each of these is configured through the sched_conf(5) parameters weight_priority, weight_ticket and weight_urgency. These three parameters control to what degree POSIX priority, ticket policy, and urgency policy are in effect. To facilitate interpretation, the raw priorities ("tckts"/"urg"/"ppri") are normalized ("ntckts"/"nurg"/"npprior") before they are used to calculate job priorities ("prio"). Normalization maps each raw urgency/ticket/priority value into a range between 0 and 1.
npprior = normalized(ppri)
nurg = normalized(urg)
ntckts = normalized(tckts)
prio = weight_priority * npprio +
weight_urgency * nurg +
weight_ticket * ntckts
The higher a job's priority value, the earlier it gets dispatched.
The urgency policy defines an urgency value for each job. The
urg = rrcontr + wtcontr + dlcontr
consists of the resource requirement contribution (rrcontr), the waiting time contribution (wtcontr) and the deadline contribution (dlcontr).
The resource requirement contribution adds up all resource
requirements of a job into a single numeric value.
rrcontr = Sum over all(hrr)
with an hrr for each hard resource request. Depending on the resource type, two different methods are used to determine the value to be used for hrr here. For numeric type resource requests, the hrr represents how much of a resource a job requests (on a per-slot basis for PE jobs) and how "important" this resource is considered in comparison to other resources. This is expressed by the formula:
hrr = rurg * assumed_slot_allocation * request
where the resource's urgency value (rurg) is as specified under urgency in complex(5), the job's assumed_slot_allocation represents the number of slots supposedly assigned to the job, and the per-slot request is that which was specified using the -l qsub(1) option. For string-type requests the formula is simply
hrr = rurg
and directly assigns the resource urgency value as specified under urgency in complex(5).
The waiting time contribution represents a weighted waiting time
of the jobs
wtcontr = waiting_time * weight_waiting_time
with the waiting time in seconds and the weight_waiting_time value as specified in sched_conf(5).
The deadline contribution has an increasing effect as jobs
approach their deadline initiation time (see the -dl option in
qsub(1)). It is defined as the quotient of the weight_deadline
value from sched_conf(5) and the (steadily decreasing) free time in
seconds until deadline initiation time
dlcontr = weight_deadline / free_time
or is set to 0 for non-deadline jobs. After the deadline passes, the value is static and equal to weight_deadline. The qstat(1) -urg option displays urgency information for jobs.
The ticket policy unites functional, override and share tree
policies in the ticket value (tckts), defined as the sum of the
specific ticket values (ftckt/otckt/stckt) for each
sub-policy (functional, override, share):
tckts = ftckt + otckt + stckt
The ticket policies provide a broad range of means for influencing both job dispatch and runtime priorities on a per job, per user, per project, and per department basis. The qstat(1) -ext option displays ticket information for jobs.
JOB RUN-TIME PRIORITY¶
The run-time priority can be dynamically adjusted in order to meet the goals set with the ticket policy when execution hosts are over-subscribed. Dynamic run-time priority adjustment can be turned on globally using reprioritize in sge_conf(5) and reprioritize_interval in sched_conf(5). If no dynamic run-time priority adjustment is done at a host level, the priority specification in queue_conf(5) is in effect.
Note that urgency and POSIX priorities do not affect runtime priority.
See sge_intro(1) for a full statement of rights and permissions.