Transparent Middleware Support for Shared State on Heterogeneous Clusters

University Of Rochester, Rochester NY

Investigators

Abstract

The explosive growth of both large and small scale distributed systems is leading to a host of new applications, most of which employ some notion of distributed shared state: information required at more than one location. Building on the experience of the PIs, this research will develop convenient, efficient middleware for applications running on heterogeneous platforms. The goal is to dramatically reduce the programmer effort required to develop distributed applications, and in particular to make it as easy to run applications on a local-area cluster as it is to to run them on a shared-memory multiprocessor. The proposed research will therefore employ a shared-memory programming model. The PIs will extend their previous work in software distributed shared memory (S-DSM) to accommodate heterogeneous languages and machine types, and to transparently provide shared state for independently-deployed processes. They will introduce application-specific notions of when a cached copy of data is ``recent enough'' to use, thereby facilitating optimized communication in more globally distributed environments. The applications driving the research include both scientific simulations and codes developed by local colleagues in the areas of datamining and intelligent multi-sense environments. The research will facilitate the effective distribution of these applications across the resources they require.