FEATURE

Predicting the performance of extreme-scale supercomputer networks | A modern supercomputer is the Internet in a microcosm, with tens of thousands of nodes— computers not much dierent from the one you may be using to read this article—all hooked together via a high-speed network. However, while computers on the Internet operate largely independently of each other, supercomputers regularly harness the power of thousands to many tens of thousands of nodes at once to run a single application signicantly faster than any lone computer could. Coordinating the eorts of so many nodes requires massive amounts of communication, making the design of the interconnection network critical to the performance of the supercomputer as a whole. In this article we present technology we are developing to predict the impact of various network-design alternatives on the overall performance of supercomputing applications before the supercomputer is even built. This is important because a large supercomputer can easily cost tens to hundreds of millions of dollars (and in the case of Japan’s K supercomputer, over a billion dollars). Being able to evaluate network technologies during their design phase helps ensure that the supercomputer will provide as much performance as possible to applications