Abstract
Computer performance measurement and evaluation (CPME) studies are conducted for the purpose of sizing and selecting a new system (selection studies); during the design phase of either a new system or a hardware/software modification to an existing system to assess the impact of the new system/modification (performance projection studies); or to assess and improve the level of performance of an existing system (performance monitoring studies). Nearly all performance measures used are related to the workload being processed by the system. There is the need for a workload which emulates the actual workload, yet executes in less time and does not compromise the adequacy of the measurements. Such a workload is called a drive or test workload. A statistical methodology is proposed to aid in the construction of a test workload. The major elements of this methodology are (a) selecting the workload subset by constructing an overall workload profile and then choosing a period which exhibits characteristics pertinent to the evaluation study, (b) choosing a set of descriptor variables which is detailed enough to represent the demand placed upon the major system resources, but is not so detailed as to complicate later stages of the analysis, (c) collecting data reflecting the values of the descriptor variables for the worksteps in the selected subset, (d) scaling the resource demand matrix so that each descriptor has mean 0 and variance 1, (e) applying principal components analysis to the scaled resource demand matrix and retaining only those components needed to explain the major part of the variability in the data, (f) clustering the transformed resource demand vectors in the principal components space using a non-hierarchical clustering algorithm with a weighted Euclidean distance measure, (g) designing synthetic jobs for each of the isolated clusters using regression analysis to obtain predictor equations for the parameter settings, (h) forming a synthetic job mix by combining a sufficient number of copies of the various synthetic jobs with appropriate parameter settings and the desired arrival time of each, and (i)...
Graybeal, Wayne Thomas (1979). A statistical methodology for constructing synthetic test workloads. Texas A&M University. Texas A&M University. Libraries. Available electronically from
https : / /hdl .handle .net /1969 .1 /DISSERTATIONS -183646.