• DocumentCode
    3235050
  • Title

    Support for Dependency Driven Executions among OpenMP Tasks

  • Author

    Ghosh, Prosenjit ; Yan, Y. ; Chapman, Barbara

  • Author_Institution
    Dept. of Comput. Sci., Univ. of Houston, Houston, TX, USA
  • fYear
    2012
  • fDate
    19-23 Sept. 2012
  • Firstpage
    48
  • Lastpage
    54
  • Abstract
    OpenMP 3.x introduced the concept of creation of tasks with the aim of handling unstructured parallelism by providing support for efficient load-balancing and dynamic scheduling of asynchronous units of work, namely tasks. This work is focused towards extending task parallelism further to exploit task dependence relationships. The goal is to extend the OpenMP tasking model to provide more flexible synchronizations based on the data access relationship among such tasks. Our approach implements extensions to the current OpenMP task directive, which aim at providing task-level granularity for synchronization of tasks sharing the same parent. The new extensions focus primarily on adding additional functionality to the OpenUH compiler runtime referenced at the time of scheduling of tasks to honor only specified explicit dependencies. The extensions provide simplicity of use and help in achieving more parallelism owing to a more flexible approach in synchronizing tasks wherein a task waits only as long as its specified explicit dependencies are honored. The extensions when tested on a basic linear algebraic algorithm - LU Decomposition, presented a 9% and 20% speedup compared to the tasking implementations of Intel and GNU compilers respectively.
  • Keywords
    linear algebra; parallel processing; processor scheduling; program compilers; resource allocation; synchronisation; GNU compilers; Intel compilers; LU decomposition; OpenMP 3.x; OpenMP task directive; OpenMP tasking model; OpenUH compiler runtime; data access relationship; dependency driven executions; dynamic asynchronous work unit scheduling; flexible synchronizations; linear algebraic algorithm; load balancing; task creation; task dependence relationships; task synchronization approach; task-level granularity; unstructured parallelism handling; Matrix decomposition; Parallel processing; Radiation detectors; Runtime; Scalability; Scheduling; Synchronization;
  • fLanguage
    English
  • Publisher
    ieee
  • Conference_Titel
    Data-Flow Execution Models for Extreme Scale Computing (DFM), 2012
  • Conference_Location
    Minneapolis, MN
  • Type

    conf

  • DOI
    10.1109/DFM.2012.16
  • Filename
    6612860