Date of Award

Fall 1993

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computing Sciences - (Ph.D.)

Department

Computer and Information Science

First Advisor

E. Gelenbe

Second Advisor

Peter A. Ng

Third Advisor

Constantine N. Manikopoulos

Fourth Advisor

David Nassimi

Fifth Advisor

Bruce David Parker

Abstract

Performance modeling of distributed and parallel systems is of considerable importance to the high performance computing community. To achieve high performance, proper task or process assignment and data or file allocation among processing sites is essential. This dissertation describes an elegant approach to model distributed and parallel systems, which combines the optimal static solutions for data allocation with dynamic policies for task assignment. A performance-efficient system model is developed using analytical tools and techniques.

The system model is accomplished in three steps. First, the basic client-server model which allows only data transfer is evaluated. A prediction and evaluation method is developed to examine the system behavior and estimate performance measures. The method is based on known product form queueing networks. The next step extends the model so that each site of the system behaves as both client and server. A data-allocation strategy is designed at this stage which optimally assigns the data to the processing sites. The strategy is based on flow deviation technique in queueing models. The third stage considers process-migration policies. A novel on-line adaptive load-balancing algorithm is proposed which dynamically migrates processes and transfers data among different sites to minimize the job execution cost. The gradient-descent rule is used to optimize the cost function, which expresses the cost of process execution at different processing sites.

The accuracy of the prediction method and the effectiveness of the analytical techniques is established by the simulations. The modeling procedure described here is general and applicable to any message-passing distributed and parallel system. The proposed techniques and tools can be easily utilized in other related areas such as networking and operating systems. This work contributes significantly towards the design of distributed and parallel systems where performance is critical.

Share

COinS