Two Layers Distributed Shared Memory

Abstract

This paper presents a methodology to design a distributed shared memory by decomposing it into two layers. An application independent layer supplies the basic functionalities to access shared structures and optimizes these functionalities according to the underlying architecture. On top of this layer, that can be seen as an application independent run time support, an application dependent layer defines the most suitable consistency model for the considered class of applications and it implements the most appropriate caching and prefetching strategies for the consistency model. To exemplify this methodology, we introduce DVSA, a package that implements the application independent layer and SHOB, an example of the second layer. SHOB defines a release consistency model for iterative numerical algorithms and it implements the corresponding caching and prefetching strategies. We present some experimental results of the methodology and discuss the performance of a uniform multigrid method developed through SHOB on a massively parallel architecture, the Meiko CS2, and on a cluster of workstations.