Scalable cortical simulations on Beowulf architectures

Abstract

Biologically motivated simulation of large scale neural networks is a computationally costly task. In this paper, a commodity 8-node Beowulf architecture is proposed as a scalable low cost environment for studies of cortical dynamics. By means of a distributed message-based event-driven framework, the size of memory-limited tractable problems increased 8-fold, resulting in a mere 8.3% increase in elapsed CPU time, attributable to inter-process communication overhead. The attainable network size reached over 10 6 neurons and 2.5×10 8 synapses, with a typical performance of 900 s , Beowulf processing time, per simulated second.