1 / 10

Communication Libraries for Parallel-Programming-Model Runtime Systems

Communication Libraries for Parallel-Programming-Model Runtime Systems. Andy Stone CS653. AndyStone -- CS653 Mini Research Exam Presentation , November 16, 2010 Slide 1. What I'll Talk About. Three communication libraries. Communication Operations.

jalia
Télécharger la présentation

Communication Libraries for Parallel-Programming-Model Runtime Systems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Communication Libraries for Parallel-Programming-Model Runtime Systems Andy Stone CS653 AndyStone -- CS653 Mini Research Exam Presentation , November 16, 2010 Slide 1

  2. What I'll Talk About Three communication libraries Communication Operations How they've been tailored for a machine AndyStone -- CS653 Paper Review Presentation , Sept 21, 2010 Slide 2

  3. Context UEs (task/processes) need to communicate Why? Synchronize Computation Pass Data In a programming model communication can be implicit or explicit Example: Set array A on process 1 to array B on process 2 MPI: Explicit Co-Array Fortran: Implicit if(id == 1) { MPI_Recv(A, n, MPI_DOUBLE, 2, 0, MPI_COMM_WORLD); } else if(id ==2) { MPI_Send(B, n, MPI_DOUBLE, 1, 0, MPI_COMM_WORLD); } if(this_image() == 1) then A=B[2] end if AndyStone -- CS653 Paper Review Presentation , Sept 21, 2010 Slide 3

  4. Models and Communication Libraries Co-Array Fortran GlobalArrays CHAPEL AndyStone -- CS653 Paper Review Presentation , Sept 21, 2010 Slide 4

  5. Communication Operations Classification of communication operations: Blocking versus non-blocking Two-sided versus one-sided 2-sided: send, receive 1-sided: put, get, accumulate What ARMCI/GASNET has that MPI doesn't: - Stronger focus on 1-sided communication - Active Messages (RPC mechanism) - I/O vector and strided messages - ARMCI: Sum-and-scale in one accumulate operation AndyStone -- CS653 Paper Review Presentation , Sept 21, 2010 Slide 5

  6. Implementation Optimizations AndyStone -- CS653 Paper Review Presentation , Sept 21, 2010 Slide 6

  7. Implementation Optimizations J. Nieplocha and B. Carpenter. Armci: A portable remote memory copy library for distributed array libraries and compiler run-time systems. In Parallel and Distributed Processing, volume 1586 of Lecture Notes in Computer Science, pages 533{546. Springer Berlin / Heidelberg, 1999. Compare MPICH against IBM Experimentally determine the behavior of ISEND/IRECV Suggest optimization: Add in wait or test commands Show 27% reduction in communication time for SAMR example AndyStone -- CS653 Paper Review Presentation , Sept 21, 2010 Slide 7

  8. Implementation Optimizations D. Bonachea, P. H. Hargrove, M. Welcome, and K. Yelick. Porting gasnet to portals: Partitioned global address space (pgas) language support for the crayXT. In Cray Users Group, 2009. Modifying GASNet to use Cray XT's native communication library: Portals AndyStone -- CS653 Paper Review Presentation , Sept 21, 2010 Slide 8

  9. Implementation Optimizations J. Nieplocha and B. Carpenter. Armci: A portable remote memory copy library for distributed array libraries and compiler run-time systems. In Parallel and Distributed Processing, volume 1586 of Lecture Notes in Computer Science, pages 533{546. Springer Berlin / Heidelberg, 1999. Accumulate is implemented differently on different machines IBM SP: Owner-computes Cray T3E: Source-computes AndyStone -- CS653 Paper Review Presentation , Sept 21, 2010 Slide 9

  10. Conclusions Libraries very similar in terms of one-sided communication and synchronization operations they include. GASNET/ARMCI Used by programming model implementors Differences: GASNET/ARMCI: I/O Vector, Strided ARMCI: Scale-and-sum accumulate operation Performance of library implementation dependent, implementations are often tailored to work well with a specific machine. AndyStone -- CS653 Paper Review Presentation , Sept 21, 2010 Slide 10

More Related