The Time-Space Model for Instruction Reference Behavior

Abstract

In this paper, an abstract model is established for describing the behavior of instruction references. From this model, we can obtain accurate predictions of cache miss ratios without the expense of trace-driven simulation. The time-space model explores the relationship between the sojourn times and working spaces of program executions. The average sojourn time per reference (ASPR) of each block of program activity is used as an indicator of the likelihood that such blocks stay in a cache. A time-space list is created, ordered by ASPR, to keep track of how much time is spent in each block of the working space. This list incorporates the information as to phase separations, loop sizes, loop counts, and so on. With this time-space relation of a trace and a few constants regarding the cache memory configuration, the miss ratio curve of a program can be predicted, based on the assumption that the higher ASPR blocks always stay in the cache. We present an evaluation of this technique using a wide variety of programs. The miss ratio curves predicted by the timespace model for fully associative caches are compared with the curves produced by trace-driven simulation, with very good agreement in most cases.