• DocumentCode
    3449576
  • Title

    MPIActor: A thread-based MPI program accelerator

  • Author

    Liu, Zhiqiang ; Song, Junqiang ; Peng, Shaoliang

  • Author_Institution
    Coll. of Comput., Nat. Univ. of Defense Technol., Changsha, China
  • fYear
    2010
  • fDate
    16-18 June 2010
  • Firstpage
    1
  • Lastpage
    2
  • Abstract
    Towards gaining the performance improvement benefited from threaded MPI while supporting MPI standard well, in this paper, we propose a thread-based MPI program accelerator (MPIActor). MPIActor is a transparent middleware to assist general MPI libraries. People can choose to adopt or abandon MPIActor freely in compiling time for any MPI program (Currently only support C code). With the join of MPIActor, in each node, the MPI processes will be mapped as several threads of one process, and the intra-node point-to-point communication and collective communication will have been enhanced by take advantage of thread based mechanism. We have implemented the point-to-point communication module of our design and evaluated it on a real platform. Comparing with MVAPICH2, the experimental results of OSU PINGPONG benchmark show a significant performance improvement from 114% to 321% for transferring messages which size is between 4KB and 2MB.
  • Keywords
    message passing; middleware; multi-threading; MPI library; MPI program accelerator; MPIActor; MVAPICH2; OSU PINGPONG benchmark; middleware; point-to-point communication; threaded MPI; Acceleration; Educational institutions; High performance computing; Libraries; Middleware; Moore´s Law; Multicore processing; Parallel programming; Performance gain; Yarn; MPI Accelerator; MPIActor; Threaded MPI;
  • fLanguage
    English
  • Publisher
    ieee
  • Conference_Titel
    Quality of Service (IWQoS), 2010 18th International Workshop on
  • Conference_Location
    Beijing
  • ISSN
    1548-615X
  • Print_ISBN
    978-1-4244-5987-2
  • Type

    conf

  • DOI
    10.1109/IWQoS.2010.5542703
  • Filename
    5542703