05 Fakultät Informatik, Elektrotechnik und Informationstechnik

Permanent URI for this collectionhttps://elib.uni-stuttgart.de/handle/11682/6

Browse

Filters

2024 - 20252

Settings

Institute: search.filters.UBSInstitut.Institut für Parallele und Verteilte SystemeInstitute: search.filters.UBSInstitut.Stuttgarter Zentrum für Simulationswissenschaften (SC SimTech)Start date: 2020Publication Type: search.filters.UBSTyp.Abschlussarbeit (Bachelor)

Search Results

Now showing 1 - 2 of 2
  • Thumbnail Image
    ItemOpen Access
    Comparison of different Hyperparameter-Tuners for Support Vector Machines : an analysis using Parallel Least-Squares SVM Library on GPU
    (2024) Dzubba, Yannick Marian
    Working with large datasets requires sophisticated tools. One such tool developed for classification is the Support Vector Machine (SVM). As with any ML algorithm, the user has to set several different Hyper Parameter (HP) to run a SVM. Finding the optimal choice of HPs is important for model performance and it is highly dependent on the dataset. Given the number of different HPs, a search space might be massive, so optimization methods have been developed, to automate this search. This work aims to compare three popular choices: The Grid Search, the Random Search and Bayesian Model Search. They are compared in different metrics, such as performance, runtime and energy. Optuna [ASY+19] was used as optimizer backend, it implements all three optimizer types, it implements Tree-Parzan Estimator (TPE) as Bayesian Search algorithm. It was connected to Parallel Least-Squares Support Vector Machine (PLSSVM) [VCBP22] as SVM implementation. PLSSVM can efficiently exploit parallel compute cores. The optimizers have been tested on a selection of different search spaces and datasets with PLSSVM running on Graphic Processing Unit (GPU).
  • Thumbnail Image
    ItemOpen Access
    Real-time visualized and GPU-accelerated lattice Boltzmann simulations
    (2025) Graf, Marcel
    In a preceding project, four lattice Boltzmann algorithms were implemented on the CPU using HPX. Building up on this project, the goal of this work is to implement the two most suitable algorithms on the GPU together with a framework permitting real-time visualization. In the propaedeuticum, fundamental concepts of GPU programming are elaborated, and based on the insights gained, the aptitude of the four algorithms from the project for a portation to the GPU is investigated. The two-lattice and the swap algorithm were identified as the most promising candidates. The visualization framework was designed using the Dear ImGui and ImPlot APIs. In the bachelor thesis, the swap algorithm and multiple variants of the two-lattice algorithm were realized using AdaptiveCpp, which is one of two major implementations of the SYCL standard. Kármán vortex streets were chosen as a scenario demonstrating the capabilities of the proposed simulations. Since all algorithms update the lattice faster than the frontend can accept new frames, all of them are suitable for fulfilling the objective under the limitations imposed by the visualization framework. Similarly to the project, a simple and mostly runtime-coordinated two-lattice variant was recognized as the most convenient and, at the same time, very competitive implementation. Out of the data layouts proposed by Mattila et al., the bundle layout is well suited for devices with small caches, while the stream layout uses the memory bandwidth more efficiently. The optimal work group size and subdomain shape also depend on the targeted hardware.