An Intuitive Approach to Performance Prediction with Application to Workload Management in VM SP/HPO

Abstract

Workload management can be greatly facilitated by using predictive models to anticipate the impact on system resources when workloads arc assigned to computer systems. In the IBM operating system VM SP/HPO (Virtual Machine System Product with High Performance Option), contention for the first 16 megabytes of storage is often a major performance problem, and its resolution often requires workload management. Such storage contention is indicated by three variables: page steals from the low demand paging area, (LOSTEALRAT), free storage extends (FREXTND), and prime free utilization (PRIMEUTL). We know of no existing predictive model for these variables. Herein, we outline our efforts to use a statistical approach to construct such a model, and we present an apparently new technique which has been developed in the process - non-parametric interpolative forecasting for monotone functions (NIMF). Although NIMF is discussed in the context of VM SP/HPO performance analysis, it is a generally applicable statistical technique. Indeed, we have used NIMF to analyze data from MVS/XA systems. Our attempts to use least squares regression to predict LOSTEALRAT have met with little success, primarily because of an inability to find an algebraic expression relating LOSTEALRAT to the workload variables (number of logged-on users and total virtual machine IO rate). However, the regression studies do indicate that LOSTEALRAT is monotonically increasing in the workload variables. When such a monotone relationship exists, an intuitive approach often used by performance analysts is to bound a predicted value by using the monotone relationship to select a lower and upper bound from the measurement data. This idea motivated the development of NIMF. A NIMF model only reauires specifying a monotone relationship between the independent and dependent variables. Using data from several VM SP/HPO 4.2 systems, NIMF and regression arc compared as to their accuracy of predicting contention for the first 16 megabytes of storage. These studies indicate that NIMF can be as accurate as a good regression model. However, formulating NIMF models is much easier than formulating regression models since NIMF only requires specifying a monotone relationship between independent and dependent variables while regression requires an algebraic relationship. We conclude by using NIMF to determine how workload can be managed to control LOSTEALRAT in a large VM SP/HPO system.