Reduction of false sharing by using process affinity in page-based distributed shared memory multiprocessor systems

In page-based distributed shared memory systems, a large page size makes efficient use of interconnection network, but increases the chance of false sharing, while a small page size reduces the level of false sharing but results in an inefficient use of the network. This paper proposes a technique that uses process affinity to achieve data pages clustering so as to optimize the temporal data locality on DSM systems, and therefore reduces the chance of false sharing and improves the data locality. To quantify the degree of process affinity for a piece of data, a measure called process affinity index is used that indicates the closeness between this piece of data and the process. Simulation results show that process affinity technique improves the execution performance as page size increases due to the effective reduction of fair sharing. In the best case an order of magnitude performance improvement is achieved