Binary Information Press
Task model is the most significant feature in the OpenMP, which provides a way to handle irregular and unstructured parallelism. This paper presents a compiler system, CellOTP, which firstly supports OpenMP task parallelism on cell BE heterogeneous multi-core architecture. CellOTP runtime library adopts a novel hybrid tasks queue structure which fully utilizes master-slave features of cell BE as well as increases the scalability of task parallelism. Dynamic load balancing is supported by asynchronous work stealing, which reduces the contention of SPE threads.