On Generating Multicast Routes for SpiNNaker, a Massively-Parallel System for Neural Net Simulation
J Navaridas, M Lujan, L Plana, S Temple and S Furber
Abstract
The human brain is an immense biological neural network characterized by high degrees of connectivity among neurons. Any system designed to simulate biologically-plausible spiking neuronal networks needs to support such connectivity and the associated communication traffic in the form of spike events. This paper demonstrates the adequacy of multicast communications to achieve such a demanding goal and introduces a collection of algorithms to generate multicast routes. These algorithms target the SpiNNaker interconnect; a two dimensional triangular toroidal mesh with support for selective multicast. As generating multicast routes is a NP-complete problem, these algorithms are an essential ingredient for an efficient operation of SpiNNaker. Although multicast networks have been studied in the literature, existing algorithms cannot be applied efficiently to SpiNNaker. A comprehensive evaluation analyzing the largest configuration of the SpiNNaker system (over 1 million ARM cores) shows that each algorithm provides diverse benefits and drawbacks which can be exploited to avoid possible bottlenecks. Results show that the communication infrastructure of SpiNNaker will be able to support the high communication pressure exerted by simulating in real-time biologically plausible spiking neural applications