Smg: Shared memory for grids

This paper discusses some of the salient issues involved in implementing the illusion of a shared-memory program-ming model across a group of distributed memory proces-sors from a cluster through to an entire Grid. This illusion can be provided by a distributed shared memory (DSM) runtime system. Mechanisms that have the potential to increase the perfor-mance by minimizing high-latency intra site messages & data transfers are highlighted. Relaxed consistency mod-els are investigated, as well as the use of a grid informa-tion system to ascertain topology information. The latter allows for hierarchy-aware management of shared data and synchronization variables. The process of incremental hy-bridization is also explored, where more efcient message-passing mechanisms can incrementally replace DSM ac-tions when circumstances dictate that performance im-provements can be obtained. In this paper we describe the overall design/architecture of a prototype system, SMG, which integrates DSM and message passing paradigms and may be the target of an OpenMP compiler. Our initial ndings based on some triv-ial examples indicate some of the potential benets that can be obtained for grid Applications