Adapting Distributed Shared Memory Applications in Diverse Environments

A problem with running distributed shared memory applications in heterogeneous environments is that making optimal use of available resources often requires significant changes to the application. In this paper we present a model, dubbed the view model, that provides an abstraction of shared data and separates the concerns of programming model, consistency, and communication. Separating these concerns makes it possible for applications to easily be adapted to different execution environments, allowing them to take full advantages of resources such as high speed interconnects and hardware-based memory coherence, and to be optimised for specific network topologies. Furthermore, it allows different data consistency protocol implementations to be used without requiring changes to the application code itself. We also present an implementation of the view model and provide experimental results showing how the view architecture can be used to improve the performance of a distributed shared memory application running in a heterogeneous multi-cluster environment.