Figure 4

Tuning the load balancing algorithm. The load balancing algorithm uses a parameter t to determine, where possible, the ideal run time for a single parallel job. We determined that the optimal t value, minimizing the wall clock time for the task, was 15 minutes. At t  < = 5 minutes, only one repeat was performed per parallel job, and the job submission overhead outweighed the benefits of running in parallel. At t  > = 20 minutes, too many repeats were performed by each parallel job, and the benefits of running in parallel were reduced. Our tests were performed by running 1500 repeats of an optimization task, where each optimization had a 5 minute run time, on a Condor pool containing 2000 available executing nodes.