Sparse Generalized HyperGrids for performance scalability

Abstract

Generalized HyperGrid multiprocessors (GHG), the generalization of k-ary n-cubes with N _d nodes in each dimension d, are often criticized for their inability to scale efficiently, as the number of processors grows. In this paper, we define the concept of performance scalability as the property of a multiprocessor network to sustain a constant throughput per processor, as the number of processors increases. To satisfy this property, we introduce the concept of Sparse Point-to-Point GHGs, in which, only a subset of the network's nodes is allowed to generate messages. We prove, both using theoretical analysis and simulation results, that Sparse Point-to-Point GHGs are performance scalable, even for high and irregular traffic. Finally, we show that GHGs are more easily expandable than k-ary n-cubes and we describe how this property extends to Sparse GHGs.