Identifying Performance Challenges in Smoothed Particle Hydrodynamics Simulations

Smoothed particle hydrodynamics (SPH) is a commonly used technique for the numerical simulation of fluids in astrophysics and many computational science and engineering fields. SPH is a gridless, fully Lagrangian, particle method where the simulated system is discretized in a series of interpolation points (also called SPH particles). The physical properties of the simulated astrophysical system are evaluated as a weighted interpolation over close neighboring particles. SPH simulations are computationally-intensive and not trivial to parallelize. SPHYNX is a state-of-the-art astrophysics SPH code that includes many recent upgrades, such as adaptive interpolating sinc kernels, an integral approach to calculate derivatives, and generalized volume elements. The goal of this work is to study the performance of SPHYNX and identify its performance challenges on high performance computing (HPC) systems. To this end, certain performance issues are identified, such as load imbalance at process and thread levels and non-overlapping communication and computation phases, among others. A number of improvements are proposed to address the identified challenges, and experiments on two HPC systems are performed to evaluate their effectiveness. The performance results show improved application performance and better scaling profile. A number of improvements are proposed to address the identified challenges, and experiments on two HPC systems are performed to evaluate their effectiveness. The performance results show improved application performance and better scaling profile. A number of improvements are proposed to address the identified challenges, and experiments on two HPC systems are performed to evaluate their effectiveness. The performance results show improved application performance and better scaling profile.