Data movement options in accelerated clusters

Abstract

The concurrency galore is currently defining computing at all levels. At the high end, it is leading to clusters composed of hundreds or thousands of nodes, each integrating 10-100s of computing cores, and fed with instructions by up to two orders of magnitudes more threads. Technology constraints prohibit a reversal of this trend, and the still unsatisfied need for more computing power has led to a pervasive use of accelerators to speed up computations. Both in terms of energy and time, communication is more expensive than computation. Being amplified by the advent of Big Data, we are observing a fundamental transition to communication-centric systems composed of heterogeneous computing units. This talk will review current techniques for data movements in clusters. This is followed by an introduction to the Global GPU Address Spaces (GGAS) project, which we use to explore data movement optimizations in heterogeneous environments. The talk will conclude with a summary of synergies regarding related research projects. © 2014 Springer-Verlag Berlin Heidelberg.