Mapping Unstructured Applications into Nested Parallelism

Nested parallel programming models, where the task graph associated to a computation is series-parallel are easy to program and show good analysis properties. These can be exploited for efficient scheduling, accurate cost estimation or automatic mapping to different architectures. Restricting synchronization structures to nested series-parallelism may bring performance losses due to a less parallel solution, as compared to more generic ones based in unstructured models (e.g. message passing). A new algorithmic technique is presented which allows automatic transformation of the task graph of any unstructured application to a series-parallel form (nested-parallelism). The tool is applied to random and irregular application task graphs to investigate the potential performance degradation when conveying them into series-parallel form. Results show that a wide range of irregular applications can be expressed using a structured coordination model with a small loss of parallelism.