On the modeling of workload dependent memory faults

Abstract

A modeling approach to investigating the interdependencies between memory faults and system performance is presented. Describing the program behavior by an independent reference model, the author develops a fault occurrence model that depends on workload characteristics such as task sojourn times, the number of page references, and the number of page faults. Determining the probability that a fault is detected at a page test, the author quantifies the workload required for fault handling. Using a queuing network in a stationary analysis, he evaluates the average performance decrease caused by memory faults. The interdependencies between performance and reliability quantities are described by a set of nonlinear equations. An iterative method for evaluating the model is given. The results of some experiments demonstrate that the performance decrease caused by memory error depends on system workload and operating system characteristics.<>