Why is MPI so slow?
| Autor: | Alexander Sannikov, Sangmin Seo, Yanfei Guo, Ken Raffenetti, Paul Fischer, Tomislav Janjusic, Thilina Rathnayake, Michael Alan Blocksome, Jithin Jose, Matthew Otten, Hajime Fujita, Sergey Oblomov, Sayantan Sur, Masamichi Takagi, Pavan Balaji, Masayuki Hatanaka, Misun Min, Abdelhalim Amer, Paul Coffman, Wesley Bland, Akhil Langer, Michael Chuvelev, Dmitry Durnov, Charles J. Archer, Min Si, Lena Oden, Gengbin Zheng, Xin Zhao |
|---|---|
| Rok vydání: | 2017 |
| Předmět: |
020203 distributed computing
Network architecture business.industry Computer science 010103 numerical & computational mathematics 02 engineering and technology 01 natural sciences Variety (cybernetics) Software Embedded system 0202 electrical engineering electronic engineering information engineering 0101 mathematics business PATH (variable) |
| Zdroj: | SC |
| Popis: | This paper provides an in-depth analysis of the software overheads in the MPI performance-critical path and exposes mandatory performance overheads that are unavoidable based on the MPI-3.1 specification. We first present a highly optimized implementation of the MPI-3.1 standard in which the communication stack---all the way from the application to the low-level network communication API---takes only a few tens of instructions. We carefully study these instructions and analyze the root cause of the overheads based on specific requirements from the MPI standard that are unavoidable under the current MPI standard. We recommend potential changes to the MPI standard that can minimize these overheads. Our experimental results on a variety of network architectures and applications demonstrate significant benefits from our proposed changes. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|