|
After users submit their jobs to the
LoadLeveler batch system, the scheduler determines the order in which jobs are
run. Numerous factors determine which jobs will run next, including:
- Priority of the job class
- Time when the job was submitted
- The job's request for number of nodes plus amount of time
- The project (proposal id) under which the job was submitted
- The allocation threshold for the project under which the job was submitted
This article briefly describes how two allocation thresholds now affect the
order in which jobs run, and the best queue to select.
Changes in the job scheduler
With the increasing number of jobs on
blackforest, SCD changed the job scheduler on 4 December 2000 to manage jobs from
projects and divisions that exceed their allocations. The process by which queued
jobs are selected for execution on blackforest and babyblue has also been changed.
These changes affect all users, both community computing users and Climate Simulation
Laboratory (CSL) users.
For projects and groups with a monthly allocation (including NCAR divisions
and CSL proposals):
- When a project has exceeded one allocation threshold (either its one-month
cutoff limit or its three-month cutoff limit), it is flagged and any jobs submitted
under that project will run after jobs with non-flagged projects.
- Projects that exceed both allocation thresholds (one-month cutoff limit
and three-month cutoff limit), are flagged in a different way. These jobs will
run after jobs with non-flagged projects and after jobs submitted under projects
that have exceeded only one threshold.
Jobs from flagged projects (projects exceeding either one or both thresholds)
stay in the queue to which they were submitted and are charged the same number
of GAUs as other jobs in that queue, but they run only after all jobs from non-flagged
projects have run in that queue.
The system sends email to the user to notify them when a job they submit is
going to be delayed because their project has been flagged.
GAUs charged are based on the queue the job was submitted to; there is no
longer any charging benefit for jobs that started in a higher-priority queue and
completed in the standby queue. You must submit jobs to the standby queue to get
the benefit of no charges on blackforest.
Note: You can employ two strategies to avoid charges for jobs from
flagged projects.
- After your project has been flagged, you can submit jobs to the standby queue
to avoid charges. The turnaround can be slower, but no charges will accrue.
- Before your project exceeds its one-month or three-month allocation, select
lower-priority work to be run in the standby queue and save GAUs for higher-priority
work that will need to be completed before the end of the time period.
Note: Standby jobs still generate MSS charges in the normal fashion.
Projects with a lifetime allocation (including university projects and some
joint projects) will not be able to run jobs once they exceed their lifetime allocation,
as in the past. University projects may request additional resources by contacting
Ginger Caldwell at cal@ucar.edu or 303-497-1229.
Finally, the round-robin fairness scheme used by the job scheduler has been
modified on blackforest and babyblue to take into account the number of CPUs used
by jobs.
A more thorough description of job scheduling on blackforest and babyblue is
provided on the LoadLeveler job scheduler
page.
In 2001 these scheduler changes will be applied to the other SCD compute servers.
For more information
|
