Simulating Performance Sensitivity of Supercomputer Job Parameters

摘要

We report on the use of a supercomputer simulation to study the performance sensitivity to systematic changes in the job parameters of run time, number of CPUs, and interarrival time. We also examine the effect of changes in share allocation and service ratio for job prioritization under a Fair Share queuing algorithm to see the effect on facility figures of merit. We used log data from the ASCI supercomputer Blue Mountain and the ASCI simulator BIRMinator to perform this study. The key finding is that the performance of the supercomputer is quite sensitive to all the job parameters with the interarrival rate of the jobs being most sensitive at the highest rates and increasing run times the least sensitive job parameter with respect to utilization. These findings have important implications for managers of supercomputers who must balance the conflicting needs of high utilization and rapid turnaround. We also find that this facility is running near its maximum practical utilization. Finally, we show the importance of the use of simulation in understanding the performance sensitivity of a supercomputer.