Implementation and Analysis of Helper Threads with SSMT

In this paper, several scenarios for helper threading were implemented and analyzed on the SPEC2000 mcf benchmark [1]. After determining the delinquent load of the mcf benchmark, three helper threading scenarios were implemented to prefetch the delinquent load into L2 cache. First a static loop-based helper process was implemented, followed by a static loop-based helper thread. Finally, a static sample-based helper threading was performed. All helper threads were simulated and evaluated on the SSMT simulator [2]. Our results indicate that sample-based triggering of helper threads results in the greatest improvement on the delinquent load, with a 71% increase in L2 hits and a 66% decrease in L2 misses, when compared to the single-threaded benchmark. However, sample-based triggering is very costly to the main thread as a whole. Consequently, loop-based triggering of the helper thread performs better on the benchmark as a whole, with only a mild speedup of 1.14%.