第171回 第2部

Today’s extremely scalable simulation technology for spiking neuronal networks enables the representation of models of more than a billion of neurons and their connections using the entire K computer. However, the runtimes of the largest possible simulations carried out so far were too long to allow for observations of the network dynamics over long periods of time, and also small to medium-scale simulations typically run in far more than real-time. The performance challenges for spiking neuronal network simulators such as NEST on general purpose computers arise from the inherent sparse but broad connectivity between neurons and from the unpredictable neuronal spiking activity. In distributed simulations of spiking networks, this requires frequent communication of spike data, and on each compute node routing of the incoming spikes to the local targets. This entails irregular memory access and hence constitutes a major performance bottleneck, which is a problem that I will address in my talk. I will present recent developments in simulation technology that aim at meeting such performance challenges.