A performance study of cache coherence protocols and write coaches for parallel-multithreaded shared-memory multiprocessors

Abstract

According to published research results, no directory-based cache coherence protocol provides best performance for all application programs in conventional multiprocessor systems that use sequential consistency models. However, recently it has been claimed that competitive-update protocols are superior to other protocols under a relaxed consistency model. Moreover, incorporating write caches improves the system performance of clean and competitive-update protocols. In this paper, we examine the different effects that occur when processing elements are replaced by parallel-multithreaded processors. According to our simulation results, the clean protocol provided the best performance for five out of six SPLASH programs. After augmentation with write caches, the clean protocol outperformed others for all applications. Though competitive-update protocols have been improved, their performance is not better than that of write-invalidate protocols for most programs.