Single-Sided Statistic Multiplexed High Performance Computing
Justin Y. Shi, Yasin Celik · 2016
For the last three decades, end-to-end computing has been the de facto paradigm for distributed and parallel computing. MPI (Message Passing Interface), RPC (Remote Procedure Call), OpenMP (share memory) and RMI (Remote Method Invocation) are all end-to-end protocols from the same computing paradigm. A myth persisted since 1990's that while the data communication community is well served by the end-to-end protocols, direct use of the end-to-end protocols in distributed computing was considered a fallacy.This paper explores the possibility of acquiring higher performance and better reliability at the same time when expanding the application processing infrastructure by avoiding the well-known fallacy. The key instrument is a single-sided statistic multiplexed computing (SMC) paradigm. It reports a computation performance study for two single-sided SMC prototypes: Synergy and AnkaCom against an MPI application in bare metal HPC environments. Computational experiments confirmed the investigation objectives by exposing the hidden deficiencies of end-to-end computing protocols. Reliability tests are also included.