(An) adaptive granularity handling scheme for page-based DSM

Abstract

The distributed shared memory (DSM) across a network of workstations (NoW) gains wide interests due to good scalability and cost/performance empowered by easy programming paradigm. Among various implementation approaches, page-based DSM utilizes an existing paged virtual memory scheme for providing shared memory view. Page-based DSM systems, however, suffer from coherence overhead mainly due to false sharing since they use a large page as a coherence unit. The optimal page size is required to satisfy communication efficiency as well as computational locality which are dynamically affected by application characteristics. Therefore, a fixed-size page cannot satisfy various applications even if it is small as a cache line size. At the same time, for some applications an invalidation-based protocol is preferred while for others an update-based protocol is more desirable. Even within a single application, different portion of shared data may be better handled by one protocol or the other. That means, a single protocol cannot satisfy various applications. Many previous researches use multiple coherence protocols according to a programmer``s annotation. Unfortunately, it is very difficult for every programmer to understand dynamic behavior of applications. Consequently, a desirable DSM system should be able to dynamically maintain the page size and the coherence protocol to satisfy the sharing behavior of a given application without disturbing programmers. In this thesis, a mechanism which can dynamically detect data sharing patterns and apply multiple coherence protocols is proposed. Data sharing patterns can be largely divided into a migratory fashion and a producer-consumer fashion. The proposed detection mechanism can distinguish the page of each pattern from the pages showing both characteristics regardless data is associated with a synchronization primitive or not. Three software-only coherence protocols, BCP (Buddy Coherence Protocol), ERP (Exclusive Read...