Performance analysis and network bandwidth parameters for Multicluster MIMD systems

Thesis/Dissertation · Wed Jan 01 04:00:00 EST 1986

OSTI ID:5185136

Jeon, C H

This thesis considers the achievement of high speed scientific computation on a class of multiprocessor systems that can be constructed using today's off-the-shelf microprocessors. The realization of high performance is considered from two points of view, the interconnection network design, in hardwater aspect, and an efficient implementation of the algorithms, in software aspect. A general framework of MIMD multiprocessor systems, called Multicluster, is presented. A Multicluster system consists of a set of processor clusters interconnected through a global interconnection network, where each processor cluster, in turn, is compared of a number of processors, I/O processors, attached processors, and a local memory. The choice of the global interconnection network is considered as a main hardware design problem. Two benchmark algorithms, a Near-Neighbor computation and the Fast Fourier Transform, are analyzed to investigate the bandwidth requirements on the interconnection network for an image processing application. Based on the results of a series of simulations, performed using a simple model of a Multicluster system, the network bandwidth requirements are discussed in terms of two critical latency parameters.

• Research Organization: Cornell Univ., Ithaca, NY (USA)
• OSTI ID: 5185136
• Country of Publication: United States
• Language: English

Related Subjects

99 GENERAL AND MISCELLANEOUS
990200* -- Mathematics & Computers
ALGORITHMS
ARRAY PROCESSORS
IMAGE PROCESSING
MATHEMATICAL LOGIC
NETWORK ANALYSIS
PERFORMANCE
PROCESSING