Hierarchical Interconnection Networks for Multicomputer Systems

Multicomputer systems are distributed-memory The Cosmic Cube [25], the Finite Element Machine [20], the NCUBE/ten [13], and the Transputer system [30] are examples of multicomputer systems. The Butterfly Parallel Proces-'Or [513 the NYU U1tracomputer [l and the CEDAR system 191 are examples of multiprocessor systems. The IBM RP3 [23] is an example of a hybrid form. This system encompasses both shared-memorv and distributed-memorv Daradigms (a MIMD systems. Communication in these systems occurs through explicit message passing. Therefore, the underlying ProceSSor interconnection network plays an important and direct role in determining their performance. Several types of interconnection networks have been proposed in the literature. Unfortunately, no network is " universally " better. Ideally, therefore, systems should use more than one such network. Furthermore, systems < & Y-that have large numbers of processors should be able to exploit mixture of the two dan be locality in communication in order to obtain improved performance. This paper proposes the use of hierarchical interconnec-The Of this paper is to and efficient tion networks to meet both these requirements. static interconnection networks for multicomputer systems. at time). A performance analysis of a class of hierarchical intercon-nection networks is presented. This analysis includes both static analysis (i.e., queueing delays are neglected) and queueing ana1-ysis. In both cases, the hierarchical networks are shown to have (within our model) by several simulation experiments. The impact of two performance enhancement schemes-replication of links and improved routing algorithms-on hierarchical inter-connection network performance is also presented. Section I1 presents details on the hierarchical network structure that is proposed here. Some examples of these hierarchical networks are analyzed in Section 111. This analysis in-analysis. Locality in COmmUniCatiOn iS treated in the analysis. Section IV discusses some performance enhancements to hierarchical interconnection networks. Section v concludes the paper by summarizing the results. better cost-benefit ratios. The queueing analysis is also validated both static (with no 'Ontention) and queueing 11. HIERARCHICAL INTERCONNECTION NETWORKS Hierarchical interconnection networks (HIN's) are intuitively appealing when a large number of processors are to be connected, for the reasons described in Section 11-A. Sev-Section 11-B presents details on some of these networks. Sec-paper.