DescriptionHigh performance MPI applications employ nondeterministic asynchronous communication to achieve greater performance. However, this nondeterminism can significantly hamper debugging. Various software tools have been developed to control nondeterminism in HPC applications, but a high-level application-agnostic taxonomy for this nondeterminism is absent and limits these tools' effectiveness in practice. We propose to address this need by extracting common nondeterministic communication motifs from representative applications.
We present a first step toward capturing nondeterminism motifs by way of a workflow for detecting and summarizing sender nondeterminism in HPC applications.