Abstract:
|
Managing performance variability is an important issue in high-performance computing (HPC). The performance variability is affected by complicated interactions of numerous factors, such as CPU frequency, the number of I/O threads, and I/O scheduler. In this paper, we develop statistical methods for modeling and analyzing HPC I/O variability. The objective is to identify both cause and magnitude of variability in HPC, and to predict performance variability. A complete analysis procedure is applied to deal with complex datasets. Several approximation methods are used to build predictive surface for the variability of HPC system. We evaluate the performance of the proposed method by leave-one- out test and mean prediction error in new system configurations. We also discuss the methodology for future system configuration by using tools in experimental designs.
|