Optimizing the Nearest-Neighbor Communication Using One-Sided Communication
Submitted by hebertem on Wed, 2016-11-09 11:36
Colloq: Speaker Institution:
Lawrence Berkeley National Laboratory
Colloq: Date and Time:
Thu, 2016-11-10 10:00
Building 5700, Room L202
Colloq: Host Email:
Nearest-neighbor communication is the dominant communication in many scientific applications. It is often implemented using MPI two-sided communication and communication data between two neighbors and are aggregated for performance concerns. As a result, there is at most one message between any pair of processes between synchronizations. This high-frequent synchronization poses a great challenge for one-sided communication. In this talk, I am going to examine the performance and programming differences between MPI two-sided, MPI one sided, and UPC++ (a PGAS implementation for one-sided communication) for the nearest neighbor communication. The performance results indicate that MPI one-sided can deliver similar performance with MPI two-sided. But UPC++ can perform significantly better by applying features such as message pipelining and active messages.
Colloq: Speaker Bio:
Hongzhang Shan specializes in parallel programming paradigms, performance tuning, and modeling. He has published over 60 papers in premier international conferences and journals. One of them was given the Gordon Bell Award at SC2008. He also got the Best Paper award at IPDPS2007 and the Best Student Paper award at SC2000.