An Approach to Verification of MPI Applications Defined in a High-Level Model

Abstract

Kaira is a prototyping tool for developing MPI programs (Bohm et al., Petri Nets 2014). The user of Kaira suggests a parallelization of a given sequential C++ code by creating a visual Petri-net based model, the tool then automatically generates the corresponding stand-alone MPI application. One of the tool modules enables to verify that the suggested parallelization has not introduced unexpected behaviors caused by unintended orders of communications among processes. The verification uses a state-space exploration procedure. Here we report on a new substantial enhancement of this procedure, leading to a significant reduction of the explored state space. The enhancement is based on the well-known methods of stubborn sets and ample sets, and is tailored to the use in the above mentioned visual model. Our concrete method is particularly simple but the experiments show that it works very well for practical examples from the area of scientific computing in distributed environment. The experiments also confirm that the Kaira model together with this method outperforms other tools for verifying MPI programs (w.r.t. instance-size/time).