DaSH is the first benchmark suite for the runtime evaluation of (low-level) parallel programming models, which combine the dataflow and shared memory paradigms. Currently, OmpSs, atomic dataflow, and Intel TBB Flow Graph are supported. The suite features 11 benchmarks, each covering a different application class, such as branch and bound, dense linear algebra, dynamic programming, and so on. For each benchmark, a sequential implementation (based on C++), two shared-memory implementations (based on OpenMP and TBB), and three implementations, based on the hybrid models, are provided.

The main contribution, besides the benchmarks themselves, is the experimental evaluation, which shows that implementations based on the hybrid models, on average, perform 27 percent better on small datasets (and 24 percent better on large datasets) than the corresponding OpenMP implementations. This is mainly caused by a reduced number of barrier synchronizations.

The paper also discusses the comparison of the hybrid models in terms of programmability by using code metrics such as the number of lines of code, the number of barrier synchronizations, and so on. The benchmarking shows that hybrid programming models allow for more convenient programming by combining the advantages of both programming paradigms: dataflow (enabling convenient composition of tasks) and shared memory (enabling convenient programming of tasks).