Performance Analysis of the Simultaneous Optical Multiporcessor Exchange Bus Architecture
Abstract
The computing world has entered an era in which executing high-performance applications has become commonplace in the scientific, engineering, and business workplace. The trend in these applications is toward the execution of larger or more detailed models in less time, with parallel processing seen as a useful tool for performing this work. The focus of this dissertation is the performance evaluation of a parallel-processing architecture, known as the Simultaneous Optical Multiprocessor Exchange Bus (SOME-Bus), that exploits the parallelism in these applications via efficient data communication between processors. The SOME-Bus is a low-latency, high-bandwidth, broadcast-based fiber-optic interconnection network which can efficiently interconnect over one hundred processing nodes. Performance potential of the SOME-Bus architecture is examined by comparing its quantifiable network characteristics to those of existing static and dynamic network topologies. The need for a broadcast-based parallel-processing design is examined by reviewing research into the workload, communication patterns, and performance of current parallel systems. Theoretical and simulation models are developed for both message-passing and distributed-shared-memory parallel processing paradigms. The message-passing model consists of a closed queueing network. An efficient solution method is developed using Norton's Theorem for Queueing Networks. Two distributed-shared-memory models are constructed, one based on a closed, multi-class queueing network, the other on a Markov-chain. Mean Value Analysis is used to evaluate the first model; the second model is evaluated by examining state probabilities. Performance results from the models and simulations are compared to results from crossbar and torus system simulations. Models and simulators were implemented using MATLAB and C languages.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1998
- Accession Number
- ADA359922
Entities
People
- Edward K. Doskocz
Organizations
- Air Force Institute of Technology