Abstract of PICOPT2 Project

This project focuses on the optimization of gyrokinetic PIC (particle-in-cell) codes on both conventional and accelerated partitions of the Marconi-Fusion platform, as well as GPU-enabling of applications. The previous proposal was focusing on the KNL processors, but Intel has decided to discontinue this type of architecture. In another recent development, a new GPU-equipped partition has been installed on the Marconi-Fusion platform: DAVIDE. Therefore, there is a strong and rather urgent need to enable fusion applications to this type of architecture. The ‘conventional' (Skylake) partition should not be forgotten, and further effort geared at improving the efficiency of PIC codes is also part of this project, with particular attention to enhanced vectorization. For the porting and optimizing on GPU, two codes are targeted. The ORB5 code has been recently ported to GPU-accelerated architecture (Intel Haswell + Nvidia P100), and this project will search ways to further optimize it, in particular for the target architecture of DAVIDE (OpenPower NVIDIA NVlink). The VENUS-LEVIS particle code, which is part of the SCENIC package and is mainly targeting fast ion physics in 3D configurations, will be considered for adaptation to GPU. Inspired by the works done on ORB5, the OpenACC programming standard will be used. In this first year, it is expected that the core kernels of VENUS-LEVIS will be considered. It is realistic to assume that porting the whole code to GPU is a multi-year effort. For the conventional CPU architecture, the general trend of Intel processors is to rely more and more on vectorization, and these aspects will be considered in this project, targeting the EUTERPE code first. It is expected that all these efforts will be mutually beneficial, and that the techniques developed here will be useful for other particle-based fusion codes as well.