A stochastic disk i/o simulation technique

Processor speeds continue to increase while disk I/O speeds lag far behind. This mismatch causes an I/O bottleneck reducing the performance of systems running I/O intensive applications. There is a need to model these types of systems to determine and quantify the performance impact of alternative designs. In this paper, we describe a technique to construct accurate stochastic simulation models from acquired trace data. The resulting simulation models accept input trace data and return estimates of disk request service times. In addition, we describe a trace collection technique capable of collecting accurate trace data from Novell Netware servers. These traces are used to construct a simulation model of a Novell Netware I/O subsystem. This simulator is evaluated and found to be an accurate predictor of disk service times.