Intelligent program optimization and parallelization for parallel computers
Abstract
Parallel compilers and programming environments need a high degree of intelligence to cope with the complex task of program optimization for parallel computers. In this thesis, we discuss essential requirements for intelligent parallel programming environments and compilers and suggest some possible ways of meeting these requirements. A new program optimization model, the feature-directed program optimization model, is introduced. Under this model, the program optimization process is driven by architectural features and program dependence graphs. A framework for realizing this model into intelligent parallel compilers is presented. Issues discussed include: knowledge manipulation and organization for program-restructuring heuristics and machine features, heuristic-based state-space search algorithms, performance prediction as a base for intelligent decision making, new program optimization techniques for distributed-memory parallel computers (message consolidation and array reshaping), and AI techniques to increase the degree of intelligence and improve the efficiency of the compilers. Finally, a prototype intelligent parallel programming environment and lessons learned from this experiment are discussed.
Degree
Ph.D.
Advisors
Gannon, Purdue University.
Subject Area
Computer science|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.