Student cluster competition 2017, team Tsinghua University: Reproducing vectorization of the tersoff multi-body potential on the Intel Skylake and NVIDIA Volta architectures

A paper of SC '16 entitled “The Vectorization of the Tersoff Multi-Body Potential: An Exercise in Performance Portability” Höhnerbach et al. (2016) [1] implemented reduced precision calculation and cross-platform vectorization for Tersoff potential, which the authors claimed as accurate, efficient, scalable and portable. In this report, we focus on recently released computing architectures, Intel Skylake and NVIDIA Volta, to present our results and compare them with the Tersoff paper. With new input provided by Porter et al. (1997) [2], we run the given testcases on our cluster and obtain results that not consistent with the performance improvements and scalability claimed in the original publication. Deeper analysis demonstrate that it is the communication bottleneck caused by special characteristics of the new input data that limit the reproducibility.