The Performance Prediction and Improvement of SPH with the Interaction-List-Sharing Method on PEZY-SCs

The demands for the optimization of particle-based methods with short-range interaction forces such as those in smoothed particle hydrodynamics (SPH) is increasing, especially for many-core architectures. However, because particle-based methods require large amount of memory access, it is challenging to obtain high efficiency for low-byte/FLOP many-core architectures. Hence, an efficient technique, the so-called “multiwalk” method, was developed in an N-body gravitational field. The key of the multiwalk method is in sharing of the interaction lists with multiple particles to offer an efficient use of the cache memory in the double-loops operation for calculating the interactions and reducing the main memory access. However, such performance improvement is not clear for the problems with short-range interaction forces such as those in SPH. In this paper, we proposed a theoretical performance model to examine the tradeoff relations between the memory and the cost of floating point operations to optimise the SPH code. We also validated the model with the wall-clock time spent on the PEZY-SCs (SC1 and SC2).