N-Neuron Simulation Using Multiprocessor Cluster

Abstract

AbstractA growing number of research effort has been made to make the Simulation of the astonishing biophysical activities of the massively interconnected organic neurons on various computational platforms such as Field programmable Gate Arrays (FPGAs) and General Purpose Graphics Processing Unit (GP-GPU). Nonetheless, to make efficient and realistic simulation of this biophysical activities, a considerable amount of processing power and memory space is incurred. Although computational platforms come with an advanced solutions such as high performance cluster, and super computers, they come with high establishment cost. Hence, biologically realistic and computationally efficient simulation on consumer level cluster of microprocessor nodes would be helpful, as they give a room for a better performance of compute intensive applications in a relatively cheaper cost as compared to dedicated High performance computing facilities. Accordingly, in this work, large scale simulation of spiking neural network (SNN) on a cluster of 8 physical cores enabled with hyper threading (16 logical cores) is presented. The neural network is composed of the biologically plausible and computationally efficient Izhikevich single neuron model. To improve the performance of the simulation and effectively exploit the computational capacity of the cluster, we have used two parallel programming techniques: distributed parallel programming using Message Passing Interface (MPI) library and distributed shard (hybrid) parallel programming using MPI in tandem with Open Multi-Processing (OpenMP) library. Moreover, to harness the combined memory and computation power of the cluster the neurons were distributed across the nodes using static load balancing mechanism. Hence, we were able to simulate up to 160,000 neurons and 3.2M synapses connection per neuron. Performance evaluation for different configuration of the SNN with a purely MPI and Hybrid Parallelization method was presented. Our performance result show that for 160K neurons with 200 synapses connections, using purely MPI parallelization with 16 MPI processes the sequential simulation has improved by 43.12% and using the hybrid parallel programming the sequential simulation has improved by 69.58%. Hence, comparing the performance results the hybrid parallelization approach demonstrated to be a good programming solution for simulation of SNNs on a cluster of consumer level multiprocessors.KeywordsSpiking neural networkMPIOpenMPHybrid parallelization