Conference Paper

Application Enabling in DEISA: Petascaling of Plasma Turbulence Codes.

Conference: Parallel Computing: Architectures, Algorithms and Applications, ParCo 2007, Forschungszentrum Jülich and RWTH Aachen University, Germany, 4-7 September 2007
Source: DBLP


The ITER (International Thermonuclear Experimental Reactor) experiment must be accompanied by advanced plasma turbulence simulations. Due to the high demands for compute power and memory, simulations must be capable of using thousands or tens of thousands of processor-cores simultaneously. Highly scalable applications are mandatory. Through a joint effort of application specialists from DEISA (Distributed European Infrastructure for Supercomputing Applications) and scientists engaged in the theory support for ITER, two important European simulation codes for core turbulence, ORB5 and GENE, have been adapted for portable usage within the heterogeneous DEISA infrastructure. Moreover, the codes were thoroughly analyzed, bottlenecks were identified and removed, and, most importantly, the scalability of the codes could be significantly enhanced. Through application of the domain cloning concept, the PIC code ORB5 was enabled for high scalability. Efficient usage of ORB5 code could be demonstrated up to 8k processors, both on a Cray XT3 and on an IBM BlueGene/L system. GENE was parallelized through domain decomposition of the five-dimensional problem grid to such a high degree that close to loss-free efficiency on up to 32k processors of an IBM BlueGene/L machine was achieved. Results combined from both strong and weak scaling measurements indicate an even higher scalability potential for GENE. Extrapolations suggest an efficient usage on up to the order of 1M processor-cores of a similarly scalable future HPC architecture is possible, representing a milestone on the way towards realistic core turbulence simulations of future fusion devices.

Download full-text


Available from: Hermann Lederer, Jan 07, 2014
  • Source
    • "In gyrokinetics the resolution of the gyration in time is dropped so that the development of turbulence in time can be directly observed. But even with its mature parallelization performance [2] [3], GENE is facing scalability problems due to the curse of dimensionality, since even moderate resolutions in each dimension lead to tremendous numbers of degrees of freedom in the simulations. One remedy to that could be the application of sparse grids [4], since they reduce the amount of degrees of freedom used to retrieve a solution. "
    [Show abstract] [Hide abstract]
    ABSTRACT: Using the five-dimensional gyrokinetic equations for simulations of hot fusion plasmas requires discretizations with a lot degrees of freedom due to the curse of dimensionality. The sparse grid combination technique could be one remedy to this problem, since it is dividing the original problem into smaller subproblems, which can be solved independently. For this study, the gyrokinetic code GENE has been used for solving the gyrokinetic eigenvalue problem using the combination technique. The performance of different combination schemes is used on a test problem and evaluated with respect to accuracy in retrieving an eigenvalue and the computational effort required. This evaluation suggests, that the sparse grid combination technique is a feasible method to compute eigenvalues of the gyrokinetic eigenvalue problem.
    Procedia Computer Science 01/2013; 18:449–458. DOI:10.1016/j.procs.2013.05.208
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: In this report we present the necessary steps to install and run the code GENE into MareNostrum. Several profiling analysis and performance test were realized in order to prove the code behaviour. We report some code's difficulties found in our architecture.
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Plasma microinstabilities are one of the key physics problems on the way to efficient power plants based on nuclear fusion. They cause anomalous heat and particle transport which significantly degrades the plasma confinement quality, thus preventing self-sustaining plasma burning in present-day experiments. Hence, extensive experimental studies are dedicated to understanding and predicting turbulence features. They are accompanied by numerical simulations which are typically based on the gyrokinetic theory. While experimental diagnostics are about to address the role of fine-scale turbulence within a bath of large-scale turbulence, nonlinear gyrokinetic codes are already able to investigate turbulent transport at a wide range of wave numbers simultaneously. However, such simulations covering several space and time scales self-consistently are computationally extremely demanding and thus need to be massively parallelized.
    12/2009: pages 491-501;
Show more