Abstract
Increasing cache hit-ratios has proved to be instrumental in improving performance of cache-based computers. This is particularly true for computers which have a high cache-miss/cache-hit memory reference delay ratio. Although software policies are often used for main vs. secondary memory "caching", the speed required for an implementation of a CPU vs. main memory cache policy has prompted only investigation of policies which can be implemented directly in hardware. Based on compile-time analysis, it is possible to predict program behavior, thereby increasing the hit-ratio beyond the capability of pure run-time (hardware) techniques. In this report, compiler-driven techniques for this kind of cache policy are described. The SCP Model (software cache policy model) provides an optimal cache prefetch and placement/replacement policy when given an arbitrary memory reference string. In addition to suggesting a simplified cache hardware model, the SCP Model can be applied to various cache organizations such as direct mapping, set associative, and full associative. Analytic results demonstrate significant improvements in cache performance. The current work discusses an optimal cache policy which applies where the string of references is known at compile time. However, this constraint can be relaxed to encompass reference strings which are known only statistically, i.e., reference strings in which data aliases make the target of some references ambiguous. Companion reports, currently in preparation, detail the extension of the SCP Model to incorporate aliases, code incorporating loops, and conditional branches.
Date of this Version
6-1-1987