MaDCoWS: A scalable distributed shared memory environment for massively parallel multiprocessors

In this paper we present MaDCoWS, a software implementation of a Distributed Shared Memory (DSM) runtime system, specifically designed for massively parallel 2-D grid multiprocessors. The system takes advantage of the network topology in order to minimise the paths of the message sequences realising the shared operations. As a result its performance is increased and the system becomes scalable even to very large processor numbers. We present the basic ideas for 2-D optimisations, the implementation structure and results from synthetic and application benchmarks executed on a 1024 processor Parsytec GCel.