hpc2: Queue Configuration


The scheduler on hpc2 is OpenPBS (Portable Batch System). This is an Open Source product. There is currently a single submittal queue (dque) and 2 run queues (small and small1), which run on a first-come first served mode. This will likely be modified. The queues maximuns are driven by the memory requirements. The memory requirement is the larger of ncpus*256MB and the memory requirement in the submittal script. Jobs are started if they satisfy the memory limit. The queue characteristics are:

Group Name    Max proc.   Max Real Time  Max memory(GB)  max jobs/user  
                                                                                  
SMALL            130        150 hrs/job      28                8
SMALL1            34         24 hrs/job       8.5              8

The current queue configuration keeps the machine at a high rate of utilization. Small memory requirement jobs will start even though several large memory jobs are running. Also, several short duration jobs can run during the day even though the system run at 100% CPU utilization, and the long running jobs run at night when fewer users are around. The fair-share mechanism works by keeping track of recent usage and allowing jobs from users who have light recent usage to schedule ahead of jobs from users who have heavy recent usage.