A contextual approach for effective recovery of inter-process communication patterns from HPC traces

Studies have shown that understanding of interprocess communication patterns is an enabler to effective analysis of high performance computing (HPC) applications. In previous work, we presented an algorithm for recovering communication patterns from traces of HPC systems. The algorithm worked well on small cases but it suffered from low accuracy when applied to large (and most interesting) traces. We believe that this was due to the fact that we viewed the trace as a mere string of operations of inter-process communication. That is, we did not take into account program control flow information. In this paper, we improve the detection accuracy by using function calls to serve as a context to guide the pattern extraction process. When applied to traces generated from two HPC benchmark applications, we demonstrate that this contextual approach improves precision and recall by an average of 56% and 66% respectively over the non-contextual method.