Multivariate point process models for response times in multiprogrammed systems

Abstract

We consider the formulation of marked multivariate point process models for job response times in multiprogrammed computer systems. Complementing queueing network representation of the structure of the system to be modeled, the particular R-process (Response time process) model we propose permits representation of resource contention, facilitates the incorporation of realistic workload characteristics into system performance predictions, and can reproduce inhomogeneities observed in running systems. Specification of the structure of the R-process model is conditional on workload marks; this effectively separates the difficult problem of formal representation of workload characteristics from the overall problem of response time prediction. To illustrate these ideas, an application to database management systems is considered. Evidence of the predictive capability of the R-process model, based on statistical analysis of response time data from an IMS system, is also given. © 1978 Plenum Publishing Corporation.