Accelerating large-scale simulations of cortical neuronal network development
Abstract
Cultured dissociated cortical cells grown into networks on mult-electrode arrays are used to investigate neuronal network development, activity, plasticity, response to stimuli, the effects of pharmacological agents, etc. We made computational models of such neuronal networks and studied the interplay of individual neuron activity, cell culture development, and network behavior. For small networks (100 neurons in a 10x10 arrangement), we concluded that our simulations' behaviors were dominated by their limited size. However, increasing network size required huge computational resources: for a single-threaded simulator, a 100x100 neuron simulation would take at least 2,000 hours (83 days). To tackle this problem, we ported the network simulator to the GPU. A first, naive implementation performed about 2.4 times faster than the single threaded simulator. By progressively modifying the simulator structure, we achieved about 23 times performance gain compared with the single threaded simulator, bringing large-scale simulations into the realm of feasibility. We executed a set of simulations of networks of 100x100 arrangements on GPU. We made statistical analyses of bursts generated by simulations, and found basic relationship between simulation parameters (independent variables), network structure (connectivity), and burst proles (emergent properties).