Performance Optimization of the XGC code on KNL architecture*

Gyrokinetic particle-in-cell (PIC) simulations have played an important role to the fundamental understanding of plasma edge turbulence, which is one of the barriers to a successful tokamak operation. In order to fully resolve the Vlasov transport equation the ion and electron species must be treated separately due to their different velocities. The electron push for kinetic electrons occurs many times per ion time step, becoming a dominant consumer of computer time. This makes the electron push a prime candidate for vectorization using modern computer architectures, such as Intel Knights Landing architecture used in many high performance computing settings. Such improvements include the proper selection of Array of Structures or Structure of Arrays datatypes, the optimal combination of nested openMP and vector processing, and the optimization of the memory access pattern within the electron push kernel. Results demonstrating the performance improvement of the electron push kernel on such architectures will be presented.