Efficient and realistic simulation of disk cache performance

Abstract

This paper describes an improved method for evaluating disk cache performance using trace driven simulation. This method differentiates between reads and writes in the trace data which results in higher miss ratios than when all traced events are treated alike. It allows for simulating various update policies and update intervals, physical blocks that are part of different files at different times, and the optimum replacement policy GOPT. These methods are applied to traces from a VAX 11/780 running 4.3 BSD UNIX to illustrate the importance of including writes in the traces and also to analyze the effects of choosing various update intervals.