Design, Implementation and Evaluation of A Compiler-Directed Distributed Shared Virtual Memory System

This paper presents the design, implementation and performance evaluation of Locust, a distributed shared virtual memory system that uses compile-time data dependency information to improve performance. The thesis of the Locust project is that the distributed shared memory program performance can closely approximate message passing program performance if compiler-generated data dependency information is eeectively exploited. Locust uses this information to address the issues of false sharing, coherence maintenance, data prefetching, and aanity scheduling. Locust also features a function-shipping approach to support operations on mutual exclusion objects and thereby sig-niicantly reducing network traac. The rst Locust prototype is implemented on a network of 90-MHz Pentium PCI machines running FreeBSD 2.0.5 connected by a 100 Mbps Fast Ethernet. For our test parallel applications Locust performs within 6-13% of message-passing programs, and its simple load balancing scheme always achieves 75% to 85% of the performance of a perfect load balancing scheme.