An effective cache coherence scheme using data locality

Abstract

The locality of data references in distributed shared memory systems affects the performance significantly. Data allocation methods reflecting the locality of data references can improve the performance of DSM systems. This paper proposes an effective cache coherence scheme by utilizing the near processor locality, and evaluates its performance. The proposed scheme reduces memory overhead by using a presence table. If a processor is within a predefined distance, the identification of a processor is set to the presence table. Otherwise, its identification is stored to a directory pool. This scheme can solve the disadvantages of the limited directory scheme and the dynamic point allocation directory scheme. Simulation are conducted using three application programs which have various shared data. The results show that the optimal data allocation method with the proposed scheme reduces the directory pool access up to 4.8 times.