How to build a supercomputer

Abstract

Supercomputer performance has been improved by increasing the clock rate of individual central processing units (CPUs) and scaling out to growing numbers of multiple high-performance processors arranged in clusters. The most powerful supercomputers are now comprised of thousands of processors operating in parallel. Such architectures are hitting a performance barrier imposed by constraints on power consumption, footprint and the demands of cooling a machine comprising large numbers of tightly packed, high-frequency electronic components. Sustained performance is also being constrained by the growing disparity between processor and memory access times, which has led to increasingly inefficient utilisation of the processing power available as the CPUs cycle idly, waiting for data to be read or written. The IBM Research team is attacking these problems with Blue Gene/L, one of the first architectures to be developed as part of its Blue Gene project. The Blue Gene approach to resolving this disparity is to employ larger numbers of moderate-frequency processors, in the domain of hundreds of MHz, allowing for a more balanced memory architecture. Systems with more than 100000 moderate-frequency processors are currently planned-interconnected via a massively powerful communication fabric. The potential performance gains of this system are many orders of magnitude larger than what can be obtained with a conventional shared memory machine.