Schedulers
From UF HPC Wiki
Schedulers
Other Information
- The scheduler selects jobs based on priority. The idle job with the highest priority is the one the scheduler is working hardest to start.
- The caveat to this is that the scheduler can "backfill". This simply means that if the scheduler can find jobs to run on idle resources while it waits for all of the resources needed for the next highest priority job, it will run them. Otherwise, those resources would go unused. It will run these "backfill" jobs only if they will finish prior to the time at which all of the resources for the highest priority job will become available - that is to say, only if they will not interfere with the start time of the job set to run next.
- Priority is determined by several components which include a base component, fair-share component, time in the queue, length of the job.
- By far, the base priority and fair-share priority are the largest contributors to the overall priority. That is by design. The time in the queue and the length (walltime) of the job usually are minor components. However, note that for backfill, the walltime can be very important because it is usually short jobs that fit the criteria for effective backfill.
It is important to remember that the HPC facility is not a "free" resource. It exists because a number of research faculty have contributed funds to create and support the facility. We allow all faculty and faculty-sponsored graduates students who request an account to have one. This is because the departments, colleges, and the university have also contributed or provided matching funds. However, those researchers (and their students) who have directly contributed to the facility have a higher base priority and are entitled to a certain share of the cluster as determined by their investment. Note that the fair-share priority component can work for or against you. If your group is over its fair-share, the fair-share component of your job's priority will be negative and will reduce the overall priority of the job. If your group is below their fair-share, the fair-share component will be positive and will increase the overall priority.
Finally note that there are some hard limits in place to keep any one group or user (investor or otherwise) from dominating or using dramatically more than their share of the cluster for an extended period of time.
