PREDICTING PERFORMANCE OF SMOOTHED PARTICLE HYDRODYNAMICS CODES AT LARGE SCALES

We present performance prediction studies and trade-offs of Smoothed Particle Hydrodynamics (SPH) codes that rely on a Hashed OctTree data structure to efficiently respond to neighborhood queries. We use the Performance Prediction Toolkit (PPT) to (i) build a loop-structure model (SPHSim) of an SPH code, where parameters capture the specific physics of the problem and method controls that SPH offers, (ii) validate SPHSim against SPH runs on mid-range clusters, (iii) show strong-and weak-scaling results for SPHSim, which test the underlying discrete simulation engine, and (iv) use SPHSim to run design parameter scans showing trade-offs of interconnect latency and physics computation costs across a wide range of values for physics, method and hardware parameters. SPHSim is intended to be a computational physicist tool to quickly predict the performance of novel algorithmic ideas on novel exascale-style hardware such as GPUs with a focus on extreme parallelism.