Performance analysis of neuronal network connectivity creation from complex rules

Abstract

This thesis investigates the performance of neuronal network connectivity instantiation during neuronal simulation. It is important to try and get the simulation times of neuronal networks closer to the actual biological time as this will significantly increase the use cases of computational neural network simulations. Since network connectivity creation phase makes up a large fraction of the total time take during network simulation, it is important to try and reduce it (Morrison et al. 2005; Ippen et al. 2017). In this thesis, we perform an investigative analysis of the utilization of compute resources of different neuronal networks during parallel simulations on a supercomputer. This is done in order to identify how key properties of neuronal networks formed from different connection rules contribute to compute-time consumption during network connectivity creation. Due to the connectivity rules, neuronal network models incorporating combinations of different network properties were formed. How these properties affect the consumption of hardware resources is investigated here. We approach this investigation by performing weak and strong scaling experiments involving the simulation of different neuronal networks. These neuronal networks are constructed using NEST, and are performed on a supercomputer using beNNch as a benchmarking tool. The plots gotten from this experimentation are discussed extensively in this paper.