Application Characteristics of Many-tasking Execution Models

Performance gain for computer systems through Moore’s Law is jeopardized by the limitations of clock rate growth due to power considerations and the limitations in instruction-level parallelism improvement from processor core computer architecture experienced over the last decade. High performance computer architectures are addressing this challenge through multicore processors that combine many processing units on a single chip. For this class of machines, conventional programming and execution practices, while still effective for some application algorithms, are suffering in scalability for many others. Multithreaded runtime systems offer convenient many-tasking execution model implementations that show increased efficiency and scalability through exploiting medium-grained thread parallelism. This paper provides an overview of the existing runtime systems and libraries used for many-tasking computation. They are empirically examined in terms of overheads and the potential impact resulting on application performance. Intra-node performance aspects are investigated using synthetic benchmarks.