A distributed shared memory (DSM) is an implementation of the shared memory abstraction on a multicomputer architecture which has no physically shared memory. Shared memory is important (as a programming model) not only because of the vast number of existing applications which use it, but also because it is a more appropriate paradigm for certain algorithms. The DSM concept was demonstrated to be viable by Li, in IVY. Recently, there has been a surge of new projects which implement DSM in a variety of software and hardware environments.

This paper gives an integrated overview of distributed shared memory. We discuss theoretical lower bounds on the performance of DSM systems, design choices such as structure and granularity, access, coherence semantics, scalability, and heterogeneity, and open problems in DSM. In addition, we describe algorithms used to implement and improve efficiency: reducing thrashing, eliminating false sharing, matching the coher­ence protocol to the type of sharing, and relaxing the semantics of the memory coherence provided. A spectrum of current DSM systems are used as illustrative examples.