Performance evaluation of distributed crossbar switch hypermesh

Loucif, Samia (1999) Performance evaluation of distributed crossbar switch hypermesh. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of scanned version of the original print thesis] PDF (scanned version of the original print thesis)
Download (4MB)
Printed Thesis Information: https://eleanor.lib.gla.ac.uk/record=b1810436

Abstract

The interconnection network is one of the most crucial components in any multicomputer as it greatly influences the overall system performance. Several recent studies have suggested that hypergraph networks, such as the Distributed Crossbar Switch Hypermesh (DCSH), exhibit superior topological and performance characteristics over many traditional graph networks, e.g. k-ary n-cubes. Previous work on the DCSH has focused on issues related to implementation and performance comparisons with existing networks. These comparisons have so far been confined to deterministic routing and unicast (one-to-one) communication. Using analytical models validated through simulation experiments, this thesis extends that analysis to include adaptive routing and broadcast communication. The study concentrates on wormhole switching, which has been widely adopted in practical multicomputers, thanks to its low buffering requirement and the reduced dependence of latency on distance under low traffic. Adaptive routing has recently been proposed as a means of improving network performance, but while the comparative evaluation of adaptive and deterministic routing has been widely reported in the literature, the focus has been on graph networks. The first part of this thesis deals with adaptive routing, developing an analytical model to measure latency in the DCSH, and which is used throughout the rest of the work for performance comparisons. Also, an investigation of different routing algorithms in this network is presented. Conventional k-ary n-cubes have been the underlying topology of contemporary multicomputers, but it is only recently that adaptive routing has been incorporated into such systems. The thesis studies the relative performance merits of the DCSH and k-ary n-cubes under adaptive routing strategy. The analysis takes into consideration real-world factors, such as router complexity and bandwidth constraints imposed by implementation technology. However, in any network, the routing of unicast messages is not the only factor in traffic control. In many situations (for example, parallel iterative algorithms, memory update and invalidation procedures in shared memory systems, global notification of network errors), there is a significant requirement for broadcast traffic. The DCSH, by virtue of its use of hypergraph links, can implement broadcast operations particularly efficiently. The second part of the thesis examines how the DCSH and k-ary n-cube performance is affected by the presence of a broadcast traffic component. In general, these studies demonstrate that because of their relatively high diameter, k-ary n-cubes perform poorly when message lengths are short. This is consistent with earlier more simplistic analyses which led to the proposal for the express-cube, an enhancement of the basic k-ary n-cube structure, which provides additional express channels, allowing messages to bypass groups of nodes along their paths. The final part of the thesis investigates whether this "partial bypassing" can compete with the "total bypassing" capability provided inherently by the DCSH topology.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Additional Information: Adviser: Dr. Lewis Mackenzie.
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Colleges/Schools: College of Science and Engineering
Date of Award: 1999
Depositing User: Enlighten Team
Unique ID: glathesis:1999-71800
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 17 May 2019 09:31
Last Modified: 25 Oct 2022 08:43
Thesis DOI: 10.5525/gla.thesis.71800
URI: https://theses.gla.ac.uk/id/eprint/71800

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year