DVSA and SHOE: support to shared data structures on distributed memory architectures

Abstract

With reference to numerical iterative algorithms, this paper exemplifies a methodology to design the runtime support of applications sharing a set of data structures on a distributed memory architecture. According to the methodology, the support is decomposed into two layers: an application independent one, supplying the basic functionalities to access a shared structure, and an application dependent layer that implements the caching and prefetching strategies most appropriate for the considered application. Starting from this assumption, we introduce DVSA, a package that implements the application independent layer and SHOB, one of the packages that can be developed on top of DVSA. SHOB defines a weak consistency memory model where the user controls the amount of inconsistency due to caching and prefetching. The model is well suitable to implement iterative numerical algorithms. Experimental results of the methodology are presented in the case of a uniform multi-grid method to solve partial differential equations.