SC16 Salt Lake City, UT

03. Acceleration of All-to-All Communication on Multi-Layer Full Mesh, Low-Cost Connectable Network Topology

Authors: Toshihiro Shimizu (Fujitsu Ltd)Masahiro Miwa (Fujitsu Ltd)Kohta Nakashima (Fujitsu Ltd)

Abstract: Recently, due to the increasing scale of computation, more and more servers are used to make calculations simultaneously and communicate with each other. These servers are connected by links and switches. Since the cost of the switch is relatively high, reduction of the number of switches is desirable to realize a larger cluster system cost-efficiently. We have already proposed the multilayer full mesh (MLFM) topology in this purpose and showed that MLFM topology can connect more servers than conventional fat tree topology and can achieve congestion-free all-to-all communication using all servers. In this poster we propose a method of congestion-free all-to-all communication using part of the servers on MLFM. This is necessary because users of supercomputer systems typically use part of the servers rather than all of them. Our experimental results show 2.2 times higher throughput compared to the conventional communication pattern.

Poster: pdf
Two-page extended abstract: pdf

Poster Index