SC²S Colloquium - October 21, 2016
|Date:||October 21, 2016|
|Time:||3:00 pm, s.t.|
Tobias Neuhauser: A Method for Simulation-based Parameter Optimization of Autonomous Emergency Braking
The parameters of an Autonomous Emergency Braking (AEB) system are tuned manually on a test-track up to now. This approach is time-consuming, costly and yields suboptimal performance of the AEB system. To overcome these disadvantages, a method for simulation-based parameter optimization of autonomous emergency braking systems is presented and implemented within this master’s thesis. First of all, contrary cost functions are modelled, which specify the target behavior of the AEB system. Their objectives are on the one hand the effectiveness of the AEB system and on the other hand the customer-acceptance. Moreover, an optimization strategy is presented, which is required to tackle the high-dimensional optimization problem consisting of more than one hundred parameters. This includes also a sensitivity analysis to screen for non-influential parameters. Different optimization methods are investigated for the high-dimensional optimization: direct, metamodel-based and hybrid optimization. The hybrid optimizer “Efficient Global Optimization” proves as the best one for this task. Finally, the so-optimized parameters are fine-tuned with the help of a gradient-based optimization
Benedikt Kucis: A GPU-accelerated Chebyshev Function Approximation on an Adaptive Tree Structure
This bachelor’s thesis describes the process of designing and implementing a GPU-accelerated Chebyshev function approximation on an Adaptive Tree Structure. The adaptive trees are used for storing Chebyshev coefficients computed using an input function for e.g. vorticity fields or Gaussian functions. Chebyshev points are optimal interpolation points to approximate smooth functions over a specified domain. The evaluation of the Chebyshev interpolant for an arbitrary point is computationally expensive. Because the Chebyshev approximation can be heavily parallelized, the goal of this work was to speed up the evaluation by utilizing the massively parallel architecture of GPUs. Existing CPU implementations in the PVFMM and TbSLAS libraries are explained and the process of implementing and optimizing the GPU version of the evaluation is described. The largest achieved speedup was about 2X the CPU performance.
Michael Obersteiner: Parallel Implementation of the Fast Multipole Method
In this thesis we develop an MPI parallelization for the Fast Multipole Method in the Molecular Dynamics software MarDyn. Different optimizations to the implementation were investigated to minimize communication overhead. By restructuring of the standard Fast Multipole traversal and the usage of non-blocking communication routines, an over- lap between communication and computation is created. Furthermore, synchronized local reduce operations are used to avoid collective operations which significantly improves parallel efficiency for large scale simulations. Moreover, we discuss novel applications of the zonal methods for parallelization of long range interactions in the context of the Fast Multipole Method. Therefore, a new adaptation of the NT method was designed which re- duces the communication partners to 6 in the local tree part and to 31 in the global tree part for send as well as receive operations. In addition, import loads are reduced significantly for the global tree part and for up to three levels in the local tree part. In this way, a parallel efficiency of 67% with a speedup of 347 can be obtained even for small simulations with 2 local levels and 512 processors. For larger processor ranges relative speedups of 5.7 for 512 to 4096 processors and 3.6 for 4096 to 32768 processors could be achieved. Moreover, the implementation is compared to the state-of-the-art Fast Multipole library ExaFMM.