Analysis and Optimization with Programmer-Controlled Memory Consistency Modeling of OpenMP Scientific Codes

This research represents a significant step towards enabling and evaluating the application of classical optimizations to exploit uniprocessor performance within explicitly parallel programs, with a flexible view of memory consistency. The research focuses on answering two important unanswered questions:

1. What kinds of performance gains can be achieved with the Location Consistency (LC) memory model in comparison to the sequential consistency (SC)-derived models for shared memory parallel programs, amidst the new developments of compiler analysis and optimization for SC-derived models?
2. As a compromise between the two divergent approaches, can both the SC-derived models and the LC model be supported within the same program, by developing a programmer-controlled memory consistency strategy supported by compiler technology?

• specification and implemenation of a programming model that assumes an end-to-end view of the memory system based on the LC model, and a study of its programmability,
• compilation analyses to uphold programmer-controlled memory consistency so programmers can choose between the more strict sequential consistency (SC) model and the LC model in different parts of the same application,
• development and refinement of cache consistency protocols based on the LC model in a software caching context,
• thorough experimental evaluation of compiler optimization and program performance under LC, programmer-controlled memory consistency, and more contemporary SC-derived models and cache protocols.

Publications

Contributors

Faculty: Dr. Lori Pollock, in collaboration with Dr. Guang Gao
Graduated Ph.D. Student: Dixie Hisley
Graduated Undergraduates: Mike Tegtmeyer, Ves Stoyanov, Matt Bridges

Funding

National Science Foundation
NSF REU
Army Research Laboratory