Date of Award

1995

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

First Advisor

Ahmed El-Amawy

Abstract

This dissertation develops a formal and systematic methodology for designing optimal, synchronous multiple bus systems (MBSs) realizing given (classes of) parallel algorithms. Our approach utilizes graph and group theoretic concepts to develop the necessary model and procedural tools. By partitioning the vertex set of the graphical representation CFG of the algorithm, we extract a set of interconnection functions that represents the interprocessor communication requirement of the algorithm. We prove that the optimal partitioning problem is NP-Hard. However, we show how to obtain polynomial time solutions by exploiting certain regularities present in many well-behaved parallel algorithms. The extracted set of interconnection functions is represented by an edge colored, directed graph called interconnection function graph (IFG). We show that the problem of constructing an optimal MBS to realize an IFG is NP-Hard. We show important special cases where polynomial time solutions exist. In particular, we prove that polynomial time solutions exist when the IFG is vertex symmetric. This is the case of interest for the vast majority of important interconnection function sets, whether extracted from algorithms or correspond to existing interconnection networks. We show that an IFG is vertex symmetric if and only if it is the Cayley color graph of a finite group $\Gamma$ and its generating set $\Delta.$ Using this property, we present a particular scheme to construct a symmetric $MBS\ M(\Gamma,\Delta)$ with minimum number of buses as well as minimum number of interfaces realizing a vertex symmetric IFG. We demonstrate several advantages of the optimal $MBS\ M(\Gamma,\Delta)$ in terms of its symmetry, number of ports per processor, number of neighbors per processor, and the diameter. We also investigate the fault tolerant capabilities and performance degradation of $M(\Gamma,\Delta)$ in the case of a single bus failure, single driver failure, single receiver failure, and single processor failure. Further, we address the problem of designing an optimal MBS realizing a class of algorithms when the number of buses and/or processors in the target MBS are specified. The optimality criteria are maximizing the speed and minimizing the number of interfaces.

Pages

207

Share

COinS