Cluster system which is low-cost commodity PCs interconnected with high-speed network is very attractive to construct large parallel machine. Small-scale SMP(Symmetric Multiprocessor) is widely used as a building block of cluster system. SDSM(Software Distributed Shared Memory) system provides shared memory abstraction with no special hardware support. In this thesis, we implement and evaluate the performance of SDSM system for SMP clusters. Our SDSM system provides HLRC(Home-based Lazy Release Consistency) memory consistency model. Our protocol utilize share memory in hardware within same SMP node, so that page fetch and message passing through network can be reduced. We optimize protocol reusing write-notice table to reduce memory overhead. It is implemented on 8 node of 2-way Pentium-III SMP interconnected with 100Mbps Fast Ethernet. The communication layer is TCP/IP. The experiment with eight applications shows that our SMP protocol achieves maximum 33% speedup improvement and 13%∼52% reduction of page fetch compared with uniprocessor protocol. Our work shows that utilizing SMP node achieves good performance on low-cost commodity environment such as 2-way SMPs interconnected with Fast Ethernet.
