Evaluating OpenMP Tasking at Scale for the Computation of Graph Hyperbolicity

We describe using OpenMP to compute δ-hyperbolicity, a quantity of interest in social and information network analysis, at a scale that uses up to 1000 threads. By considering both OpenMP workshare and tasking models to parallelize the computations, we find that multiple task levels permits finer grained tasks at runtime and results in better performance at scale than worksharing constructs. We also characterize effects of task inflation, load balancing, and scheduling overhead in this application, using both GNU and Intel compilers. Finally, we show how OpenMP 3.1 tasking clauses can be used to mitigate overheads at scale.