Improving performance visualization of OpenMP task-based applications

OpenMP is becoming a more powerful environment for exploiting task-based parallelism. Recent specification versions add support for new task clauses, while the OMPT interface provides a standard API for performance monitoring. In this paper, we present a workflow to improve the performance visualization of OpenMP task-based applications. We rely on open-source solutions such as the Tikki OMPT tracing tool and the StarVZ performance analysis framework to create enriched space-time views. We demonstrate this workflow with three applications: Strassen matrix multiply, SparseLU factorization, and a dense Cholesky factorization. For two of them, our strategy enables a better understating of the performance impact of the OpenMP task depend, task wait, and priority constructions.