Search

Modeling turbulent transport is a major goal in order to predict confinement performance in a tokamak plasma. The gyrokinetic framework considers a computational domain in five dimensions to look at kinetic issues in a plasma; this leads to huge computational needs. Therefore, optimization of the code is an especially important aspect, especially since coprocessors and complex manycore architectures are foreseen as building blocks for Exascale systems. This project aims to evaluate the applicability of two auto-tuning approaches with the BOAST and StarPU tools on the GYSELA code in order to circumvent performance portability issues. A specific computation intensive kernel is considered in order to evaluate the benefit of these methods. StarPU enables to match the performance and even sometimes outperform the hand-optimized version of the code while leaving scheduling choices to an automated process. BOAST on the other hand reveals to be well suited to get a gain in terms of execution time on four architectures. Speedups in-between 1.9 and 5.7 are obtained on a cornerstone computation intensive kernel.

In magnetically confined plasmas used in Tokamak, turbulence is respon-sible for specific transport that limits the performance of this kind of reactors. Gyroki-netic simulations are able to capture ion and electron turbulence that give rise to heat losses, but require also state-of-the-art HPC techniques to handle computation costs. Such simulations are a major tool to establish good operating regime in Tokamak such as ITER, which is currently being built. Some of the key issues to address more re- alistic gyrokinetic simulations are: efficient and robust numerical schemes, accurate geometric description, good parallelization algorithms. The framework of this work is the Semi-Lagrangian setting for solving the gyrokinetic Vlasov equation and the Gy-sela code. In this paper, a new variant for the interpolation method is proposed that can handle the mesh singularity in the poloidal plane at r = 0 (polar system is used for the moment in Gysela). A non-uniform meshing of the poloidal plane is proposed instead of uniform one in order to save memory and computations. The interpolation method, the gyroaverage operator, and the Poisson solver are revised in order to cope with non-uniform meshes. A mapping that establish a bijection from polar coordinates to more realistic plasma shape is used to improve realism. Convergence studies are provided to establish the validity and robustness of our new approach.

Gyrokinetic modeling is appropriate for describing Tokamak plasma turbulence, and the gyroaverage operator is a cornerstone of this approach. In a gyrokinetic code, the gyroaveraging scheme needs to be accurate enough to avoid spoiling the data but also requires a low computation cost because it is applied often on the main unknown, the 5D guiding-center distribution function, and on the 3D electric potentials. In the present paper, we improve a gyroaverage scheme based on Hermite interpolation used in the Gysela code. This initial implementation represents a too large fraction of the total execution time. The gyroaverage operator has been reformulated and is now expressed as a matrix-vector product and a cache-friendly algorithm has been setup. Different techniques have been investigated to quicken the computations by more than a factor two. Description of the algorithms is given, together with an analysis of the achieved performance.

This work describes the challenges presented by porting parts of the Gysela code to the Intel Xeon Phi coprocessor, as well as techniques used for optimization, vectorization and tuning that can be applied to other applications. We evaluate the performance of some generic micro-benchmark on Phi versus Intel Sandy Bridge. Several interpolation kernels useful for the Gysela application are analyzed and the performances are shown. Some memory-bound and compute-bound kernels are accelerated by a factor 2 on the Phi device compared to Sandy architecture. Nevertheless, it is hard, if not impossible, to reach a large fraction of the peak performance on the Phi device, especially for real-life applications as Gysela. A collateral benefit of this optimization and tuning work is that the execution time of Gysela (using 4D advections) has decreased on a standard architecture such as Intel Sandy Bridge.

Reproducibility of results is a strong requirement in most fields of research for experimental results to be called science. For results obtained through simulation software using high performance computing (HPC) this translates as code quality requirements. While there are many works focusing on software quality, these typically do not take the specificities of HPC scientific simulation softwareinto account. This paper presents an approach to introduce quality procedures in HPC scientific simulation softwarewhile remaining the less invasive as possible so as to ease its adoption. The approach relies on quality procedures including human code review and automated testing and offers a dedicated procedure to help correct defects found this way. These procedures are integrated in a development work-flow designed to improve the traceability of defects. By implementing this approach for the development of the Gysela code, we show that it is indeed viable and that the return on investment is positive. We also identify multiple reusable elements developed for this experiment that should reduce the cost of adopting the approach for other codes as well as some aspects that can still be improved to ensure a widespread propagation of the approach in the community.

Approximations for closing the kinetic equation for the one particle distribution function are calculated by using propagators. These provide the formal structure of the collision term in the Landau approximation. The method allows one to investigate the effect of inhomogeneities at the Debye scale and to analyse magnetised collisions, when the Larmor radius is smaller than the Debye length. This method also allows one developing a simple renormalisation scheme to derive the Lenard-Balescu collision operator.