Smt-SPRINTS: Software precompilation with intelligent streaming for resource-constrained SMTs

Abstract

We present SPRINTS, a source-level speculative precomputation framework for scientific applications running on SMTs with two execution contexts. Our framework targets memory-bound applications and reduces memory latency by prefetching long streams of delinquent data accesses. A unique aspect of SPRINTS is that it requires neither hardware nor compiler support. It is based on partial cache simulation and a compression algorithm which can accurately summarize very long streams of cache misses. SPRINTS extracts patterns from the streams, which are in turn used to generate source-level, highly optimized precomputation code. SPRINTS achieves significant performance improvements over plain thread-level parallelization and indiscriminate precomputation based on code cloning. We demonstrate these improvements using two realistic scientific applications. © Springer-Verlag Berlin Heidelberg 2005.