MPI-Based Asynchronous Simulation of Spiking Neural Networks on the Grid

Abstract

Brain microcircuits exhibit an almost chaotic behavior. This is of high interest in developing new computational models or designing high capacity storage systems. Therefore, the simulation of such microcircuits must preserve the brain dynamic behavior. But investigating the dynamics analysis of such systems is a complex computational task due to the large number of neurons and synapses in the network, the large number of simulation scenarios that need to be computed, and to their model representation. To address the first challenge, we propose in this paper an MPI-based parallelization scheme of the asynchronous spiking neural network simulation algorithm. Due to the partitioning method, we can compute scenarios with more than 50.000 neurons and 300 millions synapses. The proposed solution has been evaluated on production HPC systems, employing up to 512 parallel processes. We contribute to the second challenge by extending the fACIBiNET framework with client-side capabilities for the Globus Online service. As such, scenarios with both high-throughput and high-performance computing requirements are managed in an efficient manner from within the framework, using grid technologies.