A workload characterization by clustering technique

Abstract

An improvement of a method for the characterization of computer workload by means of arrival patterns is presented. A numerical fitting technique provides a suitable representation of the arrival rate function of jobs over one day period. In order to classify such arrival patterns, we suggest the application of the MacQueen algorithm with coarsening and refining parameters [2], which does not need the number of clusters to be fixed a priori, as in the original approach [4]. Solutions for some numerical problems related to the approximation of the arrival rate function are also provided. The proposed technique has been implemented and experimental results concerning the workload characterization of an educational system are given.