Adaptive Granularity Transparent Integration of Fine Grain and Coarse Grain Communications

The granularity of sharing is one of the key components that a ect the performance in distributed shared memory DSM systems Providing only one or two xed size granu larities to the user may not result in an e cient use of resources Providing an arbitrarily variable granularity increases hardware and or software overheads Moreover its e cient implementation requires the user to provide some information on the granularity sacri cing the programmability of the shared memory paradigm In this paper we present a new communication scheme called Adaptive Granularity AG Adaptive Granularity is a communication scheme that e ectively and transparently integrates bulk transfer into the shared memory paradigm Adaptive Granularity provides a limited number of granularities but e cient enough to achieve gains from bulk transfer without sacri cing any programmability of the shared memory paradigm and requiring any additional hardware An Adaptive Granularity protocol consists of two protocols one for ne grain data and the other one for bulk data For small size data the standard hardware DSM protocol is used and the granularity is xed to a cache line For large array data the protocol for bulk data is used and the granularity varies depending on the sharing behavior at runtime Simulation results show that AG improves performance up to over the hardware implementation of DSM e g DASH Compared with the equivalent architecture that supports ne grain memory replication at the xed granularity of a cache line e g Typhoon AG reduces execution time up to