Designing processor-cluster based systems: interplay between cluster organizations and broadcasting algorithms
Debashis Basak, D.K. Panda · 2002
Past research on designing processor-cluster based parallel systems has focused mainly on studing the packaging technologies affecting the inter-cluster network. To make such a design approach more attractive, there is a strong need to understand the details about the topology inside the cluster its memory organization, and the impact of this organization on system performance. In this paper we analyze the communication costs for accessing inter-cluster and intra-cluster memories under different cluster organizations. The merits of these organizations are evaluated based on the performance of a commonly used U-mesh broadcast algorithm. Our results indicate that tightly coupled cluster organizations with shared access to memory offer faster intra-cluster communication. This leads to such organizations to outperform loosely coupled cluster organizations. We also demonstrate that such faster intracluster access in clustered systems can be exploited to design better collective communication algorithms. We propose a new broadcasting algorithm on clustered meshes clus-mesh which outperforms existing u-mesh on clustered systems by up to 20%.