A GALS Infrastructure for a Massively Parallel Multiprocessor
IEEE Design & Test
NOCS '08 Proceedings of the Second ACM/IEEE International Symposium on Networks-on-Chip
Understanding the interconnection network of SpiNNaker
Proceedings of the 23rd international conference on Supercomputing
Event-Driven Configuration of a Neural Network CMP System over a Homogeneous Interconnect Fabric
ISPDC '09 Proceedings of the 2009 Eighth International Symposium on Parallel and Distributed Computing
INSEE: an interconnection network simulation and evaluation environment
Euro-Par'05 Proceedings of the 11th international Euro-Par conference on Parallel Processing
Scalable communications for a million-core neural processing architecture
Journal of Parallel and Distributed Computing
Hi-index | 0.00 |
The SpiNNaker system is a biologically-inspired massively parallel architecture of bespoke multi-core System-on-Chips. The aim of its design is to simulate up to a billion spiking neurons in (biological) real-time. Packets, in SpiNNaker, represent neural spikes and these travel through the two-dimensional triangular torus network that connects the over 65 thousand nodes housed in the largest size of SpiNNaker. The research question that we explore is the impact that spatial locality, temporal causality and burstiness of the traffic have on the performance of such interconnection network. Given the limited knowledge of neuron activity patterns, we propose and use synthetic traffic patterns which resemble biological neural traffic and allow tuning of spatial locality. Causality is explored by means of temporal patterns that maintain a specified overall network load while allowing at the node level autonomous causal traffic generation. Part of the traffic is generated automatically, but the remaining traffic is triggered by a spike arrival in the form of a packet or a burst of packets; as neural stimuli do. In this way, we generate non-uniform traffic patterns with an evolving concentration of activity at nodes which contain more active parts of the spiking neural network. Given the application domain, the simulation-based study focuses on the real-time behavior of the system rather than focusing on standard HPC network metrics. The results show that the interconnection network of SpiNNaker can operate without dropping packets with traffic loads that exceed more than 3.5 times those required to simulate 109 spiking neurons, despite using non-local traffic. We also find that increments in the degree of traffic causality do not affect the performance of the system, but burstiness in the traffic can hurt performance.