Estimating the Impact of External Interference on Application Performance

Abstract

The wall-clock execution time of applications on HPC clusters is commonly subject to run-to-run variation, often caused by external interference from concurrently running jobs. Because of the irregularity of this interference from the perspective of the affected job, performance analysts do not consider it an intrinsic part of application execution, which is why they wish to factor it out when measuring execution time. However, if chances are high enough that at least one interference event strikes while the job is running, merely repeating runs several times and picking the fastest run does not guarantee a measurement free of external influence. In this paper, we present a novel approach to estimate the impact of sporadic and high-impact interference on bulk-synchronous MPI applications. An evaluation with several realistic benchmarks shows that the impact of interference can be estimated already based on a single run.