The Analysis and Detection Method for Nondeterminacy in Parallel/Distributed Program

Abstract

Debugging parallel programs in more difficult than debugging traditional sequential programs. The difficulty is mostly due to the synchronization and communication between components of a parallel program. The concurrent components of a parallel program communicate with each other via either scared variable or message passing. Without proper synchronization between these communications, nondeterminacy may arise. In shared variable model, nondeterminacy is caused by data race, and in message passing model it is caused by message race. In this paper, we presented a graph based analysis technique suitable for message passing system to detect message race. Since this technique can be applied on trace analysis (post-mortem analysis), it can help the programmer find all potential races and the errors caused more efficiently. The new detection method will enhance parallel debuggers the capability of race detection and nondeterminacy elimination.