Parallelizing biochemical stochastic simulations: A comparison of GPUs and Intel Xeon Phi processors

Stochastic simulations of biochemical reaction networks can be computationally expensive on Central Processing Units (CPUs), especially when a large number of simulations is required to compute the system states distribution or to carry out advanced model analysis. Anyway, since all simulations are independent, parallel architectures can be exploited to reduce the overall running time. The purpose of this work is to compare the computational performance of CPUs, general-purpose Graphics Processing Units (GPUs) and Intel Xeon Phi coprocessors based on the Many Integrated Core (MIC) architecture, for the execution of Gillespie’s Stochastic Simulation Algorithm (SSA). To this aim, we consider an ad hoc implementation of SSA on GPUs, while exploiting the peculiar capability of MICs of reusing existing CPUs source code. We measure the running time needed to execute several batches of simulations, for various biochemical models of increasing size. Our results show that in all tested cases GPUs outperform the other architectures, and that reusing available code with the MICs does not represent a clever strategy to fully leverage Xeon Phi horsepower.