On using statistical analysis of SMF data to build usable simulation models of large-scale computer systems

摘要

This paper describes a technique of building workload models from SMF data. In a stable production environment, two classes of jobs were modeled, one employing a synthetic jobstream to model the "back-ground noise", and one employing individual models of "important" jobs. Univariate statistical analysis of SMF data provided the job profiles for the synthetic job-stream, and provided relationships between CPU time and I/O counts of each file for the "important" jobs. The resultant workload was simulated with the System Analysis Machine (SAM) language.