Networking

Here’s where the magic happens, when all the nodes speaking the same language work in harmony. Apart from the software that we just discussed, the individual nodes need to be physically connected in a network.

If the network is thought of as a series of pipes linking the nodes, the bandwidth is the diameter of the pipes and the latency is the length. The speed of the network is determined by both parameters; if both the bandwidth and the latency are high, a large amount of data can flow but it does so slowly, as though over a longer distance.

The kinds of scientific problems that require supercomputers are those dealing with massive amounts of data, on the order of petabytes or more for the fastest supercomputers. At this scale, simply connecting two nodes over the internet is too slow. Instead, high bandwidth, low latency networks using InfiniBand or Ethernet interconnects are used.

Big league supercomputing

The ‘supercomputer’ we’ve just built is clearly nowhere near the likes of the K supercomputer or other machines on the TOP500 list, but it does have the advantage of being significantly cheaper. To solve problems that supercomputers were designed to tackle, however, your best bet is probably to use your institution’s or country’s national resource.

This article was first published in the print version of Supercomputing Asia, July 2017.

———

Copyright: Asian Scientist Magazine.
Disclaimer: This article does not necessarily reflect the views of AsianScientist or its staff.