SCIENCE AND TECHNOLOGY

COLLEGE OF ELECTRICAL ENGINEERING

ABSTRACT:
In this thesis the structural design of the massively parallel processing system is investigated. The main contribution is proposing a new structure, i.e., multistage clustering structure (MSCS), and constructing an analytical model for evaluating the MSCS with different configuration. MSCS is expandable in vertical and horizontal manner, and is so flexible that can cover many of the previous parallel machine designs.

The queuing theory and Jackson queuing network is applied for constructing the analytical model, that gives a closed-form solution for the system performance metrics, such as processor’s waiting time, memory utilization, interconnection network utilization and processing power. The analytical model has been developed for shared variable and message passing organization, separately. There are two novel points in development of these analytical models:

1. application of open queuing network rules for analyzing a closed queuing network,

2. calculation of the input rate of each service center as a function of the input rate for previous center.

The above points reduce the volume of calculation and simulation time considerably. A program is developed in C++ to simulate the analytical model. This program is general enough to enable the study of a system configuration of any size.

To demonstrate the design capability of the model, a system has been designed with 2 TIPS ( 2 million MIPS) processing power. A system designer can use similar method for designing MPPs.

So a powerful tool is developed for structural designing of the new massively parallel processing systems, or evaluating the existing systems.

FOR MORE INFORMATION CONTACT:

H.S.Shahhoseini Ph.D.

Iran University of Science & Technology
Electrical Engineering Department
Narmak, Tehran, 16844, Iran.
Direct FAX: (0098-21) 888 68 68

h_s_shahhoseini@hotmail.com