Predictive modeling of I/O characteristics in high performance computing systems

Thomas Lux, Layne T. Watson, Tyler Chang

Abstract

Each of high performance computing, cloud computing, and computer security have their own interests in modeling and predicting the performance of computers with respect to how they are configured. An effective model might infer internal mechanics, minimize power consumption, or maximize computational throughput of a given system. This paper analyzes a four-dimensional dataset measuring the input/output (I/O) characteristics of a cluster of identical computers using the benchmark IOzone. The I/O performance characteristics are modeled with respect to system configuration using multivariate interpolation and approximation techniques. The analysis reveals that accurate models of I/O characteristics for a computer system may be created from a small fraction of possible configurations, and that some modeling techniques will continue to perform well as the number of system parameters being modeled increases. These results have strong implications for future predictive analyses based on more comprehensive sets of system parameters.

Thomas C. H. Lux, Layne T. Watson, Tyler H. Chang, Jon Bernard, Bo Li, Li Xu, Godmar Back, Ali Raza Butt, Kirk W. Cameron, Yili Hong: Predictive modeling of I/O characteristics in high performance computing systems. SpringSim (HPC) 2018: 8:1-8:12