LSF queue changes
We have been working on improving our LSF queuing policy. The goal is to maximize the use of the available CPU power and at the same time, minimize the waiting (pending) times on the queues, thus speeding up the overall work-flow for our users. After careful consideration we have implemented the following changes which have been well received from users:
- ** The short_serial and normal_serial queues have been modified such that jobs from these queues are allowed to run on all “owned” compute nodes when there are idle cpus. When higher priority jobs are submitted to “owned” nodes, jobs from the short_serial or normal_serial queues will be killed and re-queued as needed so the higher priority jobs can run. This policy allows jobs from the short_serial and normal_serial queues to run on many more cores, while satisfying the requirement that higher priority jobs don’t contend with resources taken by short_serial and normal_serial jobs. Consequently, you may receive email alerts from LSF that some of your jobs have been automatically re-queued.
- ** The long_serial queue had an additional 278 cores added. A total of 344 cores are available to this queue now. This should improve job throughput for this queue.