Implementierung exemplarischer, paralleler, numerischer Verfahren aus dem CFD-Bereich auf Grafikkarten unter Verwendung von OpenCL

Kurzfassung

This master thesis examines whether the linear module of the flow solver TRACE (Turbo machinery Research Aerodynamic Computational Environment) can be accelerated by the use of GPUs. The target platforms of the current implementation are CPU Clusters and employs MPI (Message Passing Interface) for parallelization. Different data formats are examined in this thesis. The basics of the General-Purpose Computation on Graphics Processing Unit programming (GPGPU) with the use of Open Computing Language (OpenCL) and Compute Unified Device Architecture CUDA are described. The first step is to analyze the flow solver TRACE. Which algorithm is used? Which parts of the algorithm are candidates to be executed by GPU? Which data format is used? TRACE uses the Blocked Compressed Sparse Row format (BCSR) with a block size of 5x5 values. It is examined if this format is suitable for processing by the GPU or if another format works better on this architecture. For testing different formats several data converters have been implemented. TRACE uses the Generalized Minimum Residual algorithm (GMRES). GMRES contains a sparse matrix-vector multiplication (SpMV); this will be ported onto the GPU. The SpMV is implemented with the use of different data formats. By trying different optimization approaches it is tested if a better performance is possible. The various implementations with OpenCL and CUDA are compared on a test system with the MPI version of TRACE. The performance of the implemented SpMV kernels is measured and compared. It is found that block variants of the ELL format give the best results.