Platform-Independent Cache Optimization by Pinpointing Low-Locality Reuse

摘要

For many applications, cache misses are the primary performance bottleneck. Even though much research has been performed on automatically optimizing cache behavior at the hardware and the compiler level, many program executions remain dominated by cache misses. Therefore, we propose to let the programmer optimize, who has a better high-level program overview, needed to resolve many cache problems. In order to assist the programmer, a visualization of memory accesses with poor locality is developed. The aim is to indicate causes of cache misses independent of actual cache parameters such as associativity or size. In that way, the programmer is steered towards platform-independent locality optimizations. The visualization was applied to three programs from the SPEC2000 benchmarks. After optimizing the source code based on the visualization, an average speedup of 3.06 was obtained on different platforms with Athlon, Itanium and Alpha processors; indicating the feasibility of platform-independent cache optimizations.