Cache Behaviour of the SPEC 95

Previous cache studies are becoming rather outdated, since they deal with older benchmarks (such as those in SPEC92) and smaller cache conngurations than are now feasible. We remedy this situation by performing a battery of cache tests on several interesting programs in the SPEC95 suite. The applications of chief interest to us are those that are represent current trends in computing, such as ijpeg, perl and database programs. We try out a variety of key cache conngurations and do some analyses of second-level caches. The results generally corroborate older studies, with some subtle and interesting diierences. Some conclusions are: (1) a block size of 32 bytes remains optimal for rst-level caches; (2) LRU buys little more than random replacement; (3) going beyond 2-associativity results in little improvement; (4) the 2:1 \cache rule-of-thumb" is highly application-dependent; (5) second-level caches do not need to be very large, and there is little point in pushing their associativity beyond 2-way or making their replacement policy LRU; and (6) second-level caches may warrant blocks of 64 bytes or more, depending upon their size. We also found some idiosyncrasies of diierent classes of programs; for instance, the optimal block size for our database application vortex was lower than that for our other benchmarks, indicating highly non-sequential accesses.