Adaptive optimization of a time-sharing system's performance

Abstract

The complex task of managing a virtual memory multiprogramming system is considered as one which can be achieved by allowing the operating system to make use of measurement data gathered on-line in the scheduling decisions it has to make. System performance optimization is achieved by continuous monitoring of critical system parameters and workload characteristics and by use of this information in a real-time adaptive feedback control policy. As a specific application of this approach, the maximization of system throughput by the regulation of the degree of multiprogramming in a virtual memory system is examined. The specific form of this performance measure as a function of the number of active processes sharing main memory is used in the design of an adaptive and statistical maximum-seeking algorithm designed to respond to abrupt changes in program locality. The data gathering and smoothing procedures and the optimization policy are then implemented in a simulator of a virtual memory time-sharing system and evaluated in simulation runs with a random and time-varying workload. These experiments are used to tune the various parameters of the algorithm and to demonstrate its ability to maintain the system at an optimal level of performance. Statistical confidence intervals for these simulation runs are given in order to provide a measure of significance to the experiments.