The International Conference for High Performance Computing, Networking, Storage and Analysis
Performance Model for Large-Scale Neural Simulations with NEST.
Authors: Wolfram Schenck (Juelich Research Center), Andrew V. Adinetz (Juelich Supercomputing Center), Yury V. Zaytsev (Juelich Research Center), Dirk Pleiter (Juelich Supercomputing Center), Abigail Morrison (Juelich Research Center)
Abstract: NEST is a simulator for large networks of spiking point neurons for neuroscience research. A typical NEST simulation consists of two stages: first the network is wired up, and second the dynamics of the network is simulated. Our work aims at developing a performance model for the second stage, the simulation stage, by a semi-empirical approach. We collected measurements of the runtime performance of NEST under varying parameter settings on the JUQUEEN supercomputer at Forschungszentrum Jülich, and subsequently fitted a theoretical model to this data. This performance model defines the simulation time as weighted sum of algorithmic complexities which have been identified in the NEST source code. After parameter fitting, the coefficient of determination on the training data is close to 1.0, and the model can be used to successfully extrapolate NEST simulation times. Furthermore, recommendations for algorithmic improvements of the NEST code can be derived from the modeling results.