Performance of the Particle-in-Cell Method with the Intel (Broadwell, KNL) and IBM Power9 Architectures

The particle-in-cell method is one of the efficient methods of plasma simulation. But the principal limitation of the method is its resource intensity. As the motion computation of a large number of model particles can be performed using the parallel algorithms, the computational node architecture is very important. In this paper comparison is made of the performance of the particle-in-cell method between the classic Intel Broadwell architecture, Intel Xeon Phi accelerators, and IBM Power9 processor. As the number of computing operations per one particle depends on its form-factor, in the given paper comparison has been made of the form-factors from the first to the fourth order. The computational experiments have shown that the IBM and Intel Broadwell nodes are similar in performance, and the use of Intel KNL in case of higher-order form-factors can promote essential performance growth. Moreover, in order to effectively use the modern processors it is necessary to take into account their key features such as the memory access, vector instructions, etc.