Abstract of GOKE Project

Predicting the performance of fusion plasmas in terms of amplification factor, namely the ratio of the fusion power over the injected power, is among the key challenges in fusion plasma physics. In this perspective, turbulence and heat transport need being modeled within the most accurate theoretical framework, using first-principle non-linear simulation tools. Using the non-linear global full-f gyrokinetic 5D code GYSELA, a simulation very close to ITER-size plasmas with kinetic ions and adiabatic electrons has been performed on 16384 cores during 15 days. This type of simulation is at the edge of current research in fusion plasma modeling. In each part of the code GYSELA, parallel algorithms have been designed in order to scale up to thousands of cores using a hybrid MPI/OpenMP approach. Each year, new methods and techniques are conceived in order to tackle performance and scalability bottlenecks related to computations, communications and disk input/outputs. We have for example achieved a relative efficiency of 91% on the full Juqueen machine composed of 458,752 cores [1].

Intel MIC architecture and NVIDA GPGPUs are steadily being adopted in clusters. The current generation of MIC coprocessor, the Xeon Phi, provides a highly multi-threaded environment on which regular programming models such as MPI/OpenMP can be used. This specific hardware offers both large memory bandwidth and computing resources compared to standard computes nodes. GPGPUs are also interesting alternatives but require the use of other programming models such as OpenCL, CUDA or OpenACC.

Former HLST projects (MICPORT in 2013, GOMIC in 2014) have demonstrated that porting a large scientific application on MIC is really difficult. Targeting GPUs is expected to be at least as challenging. This project therefore rather aims to focus on short computation kernels that can be studied in depth. This should allow to finely tune the code so as to leverage the hardware capabilities as much as possible. Even if this approach does not immediately yield production-ready applications for the new hardware, it enables to gather precious information on the best approach to port the full code in a second step. This project will thus study and tune small kernels inspired from those of Gysela on Intel KNL and NVIDIA K80 GPUs.

[1] http://www.fz-juelich.de/ias/jsc/EN/Expertise/High-Q-Club/_node.html