Parallel visualization and compute environments for graphics clusters [article]

Magnus Strengert, Universität Stuttgart, Universität Stuttgart
2010
Parallelism evolved into the primal driving force for progressing the performance of nearly all of today's processing platforms, ranging from mobile devices, over personal computers, to the world's most powerful supercomputers. Whether it is simultaneously executing multiple threads on a modern central processing unit (CPU), or utilizing the massively parallel processor arrays of recent graphics processing units (GPUs); whether it is scaling performance by employing a multiplicity of each of
more » ... h processor in a single system, or using large cluster environments composed of hundred of thousands of such processing units, parallelism forms the hardware foundation to approach almost all of today's most demanding compute and visualization challenges. However, making highly efficient use of all the theoretically available distributed processing power remains the key challenge to successfully scale the actual performance to solve a given task. This thesis addresses the challenges of efficiently leveraging all levels of parallelism inherent to graphics cluster systems - an interconnected compound of processing nodes equipped with GPUs - for the purpose of distributed visualization as well as parallel general purpose computing. Various algorithms and techniques have been developed that are well suited for parallel execution and also serve as building blocks for utilizing the next higher levels of parallelism. The introduced single-pass volume ray casting technique allows for near-optimal usage of the parallel processor array of a single GPU while still maintaining a high level of flexibility. GPU-based pyramidal filtering techniques enable a new, highly adaptive volume sampling algorithm, which is well suited for high output resolutions. Both techniques allow to offload visualization workload from the CPU, which in turn opens up the CPU's parallel processor cores to be leveraged for other tasks, such as data compression or image processing. For that purpose, a highly optimized alpha compositing operator has been introduce [...]
doi:10.18419/opus-2684 fatcat:wdguxcppgrgvlo6bsjz3zrtvhu