A distributed shared memory (DSM) system provides an implementation of the shared memory abstraction on a message-passing architecture which has no physically shared mem­ory. This paper describes recent developments in the design of DSM systems in which the shared memory model is supported primarily through software modifications to existing vir­tual memory facilities in the operating system. These DSM systems are characterized by a larger unit of sharing, typically at the page level, and have been designed for loosely-coupled workstation networks as well as message-passing multiprocessor architectures. Our focus is on performance enhancements that have been developed for these DSM systems to reduce the memory latency incurred for coherence, data location, and data access in a system with physically disjoint memories. We describe enhancements designed to reduce (1) latency stemming from the design of coherence algorithms, (2) latency attributed to communication overhead generated by the network software and hardware, and (3) latency due to unnecessary message-passing due to large page sizes and false sharing.