Managing Nondeterminism in PVM Programs

Clusters of workstations gain more and more success and acceptance. One reason ar emerging standard libraries for interprocess communication, like PVM and M PI. But despite all success it is still hard to exploit the available computing power of these systems. Beside the fact that design and coding of parallel programs is more com plex than in the sequential case, testing and debugging is more complicated since nond eterminism caused by interprocess communication plays a major role. Nondeterminism implies two unpleasant properties: successive execution s of the same program with the same input values often do not exhibit identical b ehavior and watching the program influences the execution. The former effect is called unpredictable behavior, the latter one is called probe effect. Consequences of these properties are poor test coverage, only partially useful regression tests, and the impossib ility of applying methods based on cyclic debugging. Of course, there exist several approaches for debugging parallel programs. But the problems which are entailed by nondeterminism are only partially addr essed. The only tool focusing on nondeterminism is mdb [1]. However, none of f rs an appropriate methodology for debugging nondeterministic programs in a systemat ic w y. Facing this fact we investigated a new idea addressing the problems of non determinism in a different way. In this paper we present our first approach to th is problem, which followws the concept of controlled execution driven by patterns. T he first implementation of this approach is currently being realized for PVM within the sco pe of the TOOL–SET environment [2].