Fujitsu designs leaner supercomputer with fewer switches
By Tim Hornyak, IDG News Service | July 16th, 2014
A new communications algorithm and multilayer full-mesh topology promise power savings.
Fujitsu has developed an approach to cluster supercomputers that reduces the number of network switches by 40 percent without sacrificing performance.
The approach centers on using a new communications algorithm that efficiently controls data transmissions as well as deploying a multilayer full-mesh topology in the arrangement of the network.
Compared to a three-layer “fat-tree” network topology, which employs a tree-like structure of connections, the multilayer full-mesh topology eliminates a layer of switches through more efficient mapping.
Meanwhile, scheduling data transfers avoids data collisions along the same paths when each server is communicating with every other server.
A cluster supercomputer system of 6,000 servers could use hundreds or thousands of network switches, with the networking accounting for 20 percent of the system’s electricity needs.
When applied to a system of several thousand units, the new approach reduces network switches by 40 percent while maintaining the level of performance of a conventional system.
“As a result, cost reductions can be achieved across the board, from number of components used and power consumption to installation space and maintenance,” a Fujitsu spokesman wrote in an email.
The technology could speed the use of powerful supercomputers used for the analysis of earthquakes and weather data as well as drug discovery, according to Fujitsu. Cluster supercomputers have been used to design everything from smartphones to cars.
Fujitsu Laboratories is presenting the technology later this month at the Summer United Workshops on Parallel, Distributed and Cooperative Processing 2014 (SWoPP 2014) in Niigata City in Japan.
The lab plans to produce a practical version of the approach during the year ending March 31, 2016, and continue research into reducing the number of network switches in cluster supercomputers.