Experiences with Helper Threads for Speculative Parallelization in Software?

Abstract. Software schemes to speculative parallelization require no changes to the hardware of existing shared-memory systems, but can suer from signicant overheads. This paper evaluates the use of helper threads to assist the execution of the compute threads by performing some of the most expensive speculative operations on their behalf. Per-formance gains can be achieved if alleviating such overheads compensates for the loss of processing resources available for the main computation. Moving speculative operations from the main threads to helper threads, however, is not trivial and introduces several new race conditions, which are addressed in this paper. Experimental results so far show that for an SMP system helper threads fail to materialize any signicant perfor-mance gains, but that for a system with multithreading a more promising outcome can be expected.