An Empirical Study on how Program

Cache miss rates are quoted for a speciic program, cache connguration, and input set; the eeect of program layout on the miss rate has largely been ignored. This paper examines the miss variation, that is, the variation of the miss rate for instruction and data caches resulting from randomly chosen layouts. New layouts were generated by changing the order of the modules on the command line when linking, changing compiler version and optimization level, and inserting dummy modules when linking. This analysis is performed for several cache sizes, lines sizes, set-associativities, and input sets for ve programs in the SPEC92 benchmark suite. Miss rates were observed that varied from 60% to 180% of the mean miss rate. We did not observe any consistently good layouts across diierent parameters; in contrast, several layouts were consistently bad. Overall, cache line size and input set has little eeect on the miss variation , while increasing the cache size (i.e. decreasing the miss rate), decreasing the set-associativity, or increasing the optimization level increases the miss variation. For a direct-mapped cache, the results in this paper call into question the validity of using a single layout (i) to determine the miss rate of a given program, (ii) to determine how a given compiler optimization aaects the miss rate, and (iii) to make architecture design decisions based on the miss rate.