Semester

Fall

Date of Graduation

2003

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Hany H. Ammar

Committee Co-Chair

K. Subramani

Abstract

Dynamic load balancing strategies have been shown to be the most critical part of an efficient implementation of various applications on large distributed computing systems. The need for dynamic load balancing strategies increases when the underlying hardware is a non-dedicated heterogeneous network of workstations (HNOW). This research focuses on the single program multiple data (SPMD) programming model as it has been extensively used in parallel programming for its simplicity and scalability in terms of computational power and memory size.;This dissertation formally defines and addresses the problem of designing a scalable dynamic load-balancing algorithm for pipelined SPMD applications on non-dedicated HNOW. During this process, the HNOW parameters, SPMD application characteristics, and load-balancing performance parameters are identified.;The dissertation presents a taxonomy that categorizes general load balancing algorithms and a methodology that facilitates creating new algorithms that can harness the HNOW computing power and still preserve the scalability of the SPMD application.;The dissertation devises a new algorithm, DLAH (Dynamic Load-balancing Algorithm for HNOW). DLAH is based on a modified diffusion technique, which incorporates the HNOW parameters. Analytical performance bound for the worst-case scenario of the diffusion technique has been derived.;The dissertation develops and utilizes an HNOW simulation model to conduct extensive simulations. These simulations were used to validate DLAH and compare its performance to related dynamic algorithms. The simulations results show that DLAH algorithm is scalable and performs well for both homogeneous and heterogeneous networks. Detailed sensitivity analysis was conducted to study the effects of key parameters on performance.

Share

COinS