Speculative Optimizations for Parallel Programs on Multicores

Abstract

The advent of multicores presents a promising opportunity for exploiting fine grained parallelism present in programs. Programs parallelized in the above fashion, typically involve threads that communicate via shared memory, and synchronize with each other frequently to ensure that shared memory dependences between different threads are correctly enforced. Such frequent synchronization operations, although required, can greatly affect program performance. In addition to forcing threads to wait for other threads and do no useful work, they also force the compiler to make conservative assumptions in generating code.

We analyzed a set of parallel programs with fine grained barrier synchronizations, and observed that the synchronizations used by these programs enforce interprocessor dependences which arise relatively infrequently. Motivated by this observation, our approach consists of creating two versions of the section of code between consecutive synchronization operations; one version is a highly optimized version created under the optimistic assumption that no interprocessor dependences that are enforced by the synchronization operation will actually arise. The other version is unoptimized code created under the pessimistic assumption that interprocessor dependences will arise. At runtime, we first speculatively execute the optimistic code and if misspeculation occurs, the results of this version will be discarded and the non-speculative version will be executed. Since interprocessor dependences arise infrequently, misspeculation rate remains low. To detect misspeculation efficiently, we modify existing architectural support for data speculation and adapt it for multicores. We utilize this scheme to perform two speculative optimizations to improve parallel program performance. First, by speculatively executing past barrier synchronizations, we reduce time spent idling on barriers, translating into a 12% increase in performance. Second, by promoting shared variables to registers in the presence of synchronization, we reduce a significant amount of redundant loads, translating into an additional performance increase of 2.5%.