Revisiting Persistent Communication in MPI

  • Yutaka Ishikawa
  • Kengo Nakajima
  • Atsushi Hori
Part of the Lecture Notes in Computer Science book series (LNCS, volume 7490)

Abstract

The implementation of persistent communication provided in MPI is reconsidered to provide low latency and true overlapping communication and computation. In the persistent communication facility, the end-points of both the sender and the receiver are set up by issuing MPI Send init and MPI Send recv primitives prior to actual communication triggered by the MPI Start or MPI Startall primitive. The same communication pattern is reused without reissuing the initialization. Thus, at the start of actual communications in persistent communication, the runtime system already knows all the communication patterns, i.e., peers and message sizes if both sender and receiver have issued persistent communication primitives.

References

Copyright information

© Springer-Verlag Berlin Heidelberg 2012

Authors and Affiliations

  • Yutaka Ishikawa
    • 1
    • 2
    • 3
  • Kengo Nakajima
    • 2
  • Atsushi Hori
    • 3
  1. 1.Department of Computer ScienceUniversity of TokyoJapan
  2. 2.Information Technology CenterUniversity of TokyoJapan
  3. 3.Advanced Institute of Computational ScienceRIKENJapan

Personalised recommendations