...
The scheduler's starving code is being deprecated. The motivation behind this is that the starving code is a binary on/off. You are starving or not. One second you have low priority, and the next you have super high priority. It is much better for jobs to slowly gain priority as they wait. This is achieved with eligible_time and the job_sort_formula.
Interface Changes:
Obsolete: help_starving_jobs
- The sched_config option help_starving_jobs is being obsoleted.
- If the option is turned on, a log message will be printed on how to achieve starving support with the to use the job_sort_formula.
- This option will be removed from the default sched_config file
- This option will no longer affect job's priority.
...
- The sched_config option max_starve is being obsoleted
- If this option is turned on, a log message will be printed on how to achieve starving support with the formulato use the job_sort_formual.
- This option will be removed from the default sched_config file
- This option will no longer affect job's priority
...
qmgr> set server eligible_time_enable = True
Example 1:
...
To enable backfilling, edit sched_config:
strict_ordering: True ALL
Crafting a formula will be different for every site, but the thing to know about adding eligible_time to the formula is to correctly normalize it to the other factors. The eligible_time factor will grow large. If a job wait for one day, it is already 86400.
Example 1: If there is no formula already, it is simple:
qmgr> set sched job_sort_formula = eligible_time
Example 2: give shorter jobs priority over longer jobs
qmgr> set sched job_sort_formula = eligible_time / walltime
Example 3: give larger jobs priority. The eligible_time factor won't overtake the ncpus factor for 11.5 days. If this is too short, increase the scaling factor on ncpus.
qmgr> set sched
...
job_sort_formula = 1000000*ncpus + eligible_time
...
Project Documentation Main Page
...