Multi-threaded MPI Remote Memory Mini-applications at Scale

ACES researchers have used mini-applications utilizing next generation communication techniques, remote memory access with multi-threading, the only RMA-MT mini-applications currently available and demonstrated their performance at full scale on Trinity. Initial results are promising and are scaling well. Improvements have been made to Open MPI in response to these mini-applications and initial scaling results that are showing promise for future testing on real applications. Initial results on smaller systems show that Open MPI is performing well in comparison to other MPI implementations.

RMA-MT mini-app run time overhead compared to traditional send/recv
RMA-MT mini-app run time overhead compared to traditional send/recv

 

Contact
Ryan Grant, regrant@sandia.gov

July 1, 2016