Advisor

Luke, Edward A.

Committee Member

Reese, Donna S.

Committee Member

Dandass, Yoginder S.

Date of Degree

12-1-2005

Document Type

Graduate Thesis - Open Access

Major

Computer Science

Degree Name

Master of Science

College

James Worth Bagley College of Engineering

Department

Department of Computer Science and Engineering

Abstract

For distributed memory architectures, communication cost is a significant source of overhead in parallel scientific applications. Many proposed communication optimizations duplicate the behavior of well-written hand-tuned parallel code. Because of continuous changes in architectural components, these types of low-level optimizations are not always effective. This thesis seeks to develop a high-level optimization of work replication in which computations are replicated to minimize communications. There exist performance trade-offs between computation cost and communication cost because of work replication. Due to these trade-offs, it is required to determine which computations should be replicated to improve overall performance. This research presents the development of a model-based approach with heuristics to automatically determine the computations to replicate. Using a computational and communication model, the execution time is predicted to make replication decisions.

URI

https://hdl.handle.net/11668/17470

Share

COinS