Debugging and analysis of large-scale parallel programs. Doctoral thesis
One of the most serious problems in the development cycle of large-scale parallel programs is the lack of tools for debugging and performance analysis. Parallel programs are more difficult to analyze than their sequential counterparts for several reasons. First, race conditions in parallel programs can cause non-deterministic behavior, which reduces the effectiveness of traditional cyclic debugging techniques. Second, invasive, interactive analysis can distort a parallel program's execution beyond recognition. Finally, comprehensive analysis of a parallel program's execution requires collection, management, and presentation of an enormous amount of information. This dissertation addresses the problem of debugging and analysis of large-scale parallel programs executing on shared-memory multiprocessors. It proposes a methodology for top-down analysis of parallel program executions that replaces previous ad-hoc approaches. To support this methodology, a formal model for shared-memory communication among processes in a parallel program is developed. It is shown how synchronization traces based on this abstract model can be used to create indistinguishable executions that form the basis for debugging. This result is used to develop a practical technique for tracing parallel program executions on shared-memory parallel processors so that their executions can be repeated deterministically on demand.
- Research Organization:
- Rochester Univ., NY (USA). Dept. of Computer Science
- OSTI ID:
- 6920106
- Report Number(s):
- AD-A-221678/6/XAB; TR-312
- Resource Relation:
- Other Information: Thesis
- Country of Publication:
- United States
- Language:
- English
Similar Records
Parallel execution of Lisp programs. Doctoral thesis
A mechanism for efficient debugging of parallel programs