Research On Parallel Visualization For Large Scale Dataset

Parallel Visualization is one of the most significant techniques for analysing large scale data sets, and it play an important role in science simulation and computation science. Nowadays, many authorities in the world take parallel visualization as an important research fields, and they spend much time to develop parallel visualization technique.However, there still many problems which need to resolve in parallel visualization, for example, the preprocessing will cost too much time and the rendering frame rates still not meet the needs of real-time interaction.With the developments of the computer science and GPU technique, high-performance GPU cluster has been a powerful tool for visualizing large scale data sets, and it gives an unprecedented opportunity for developing parallel visualization. In this paper, we focus on studying the key techniques of parallel visualization, the main research achievements are detailed as follows:(1)To solve the problem that the construction of binary interval tree is very timeconsuming, we proposed a construction algorithm of binary interval tree nodes, and it is based on lattice partitioning and applied to isosurface extraction. Our algorithm will reduce the complex of sorting extremely in preprocessing phase, and the experiment show that the preprocessing time will reduce 50% approximately, while the difference between optimal algorithm and ours on active-cell searching time could be ignored. So our algorithm is practical, especially for the real-time visualization application system.(2)To fully explore the potentials of multi-GPU system and improve the visualization performance, an asynchronous parallel algorithm was proposed for parallel volume rendering on multi-GPU system. Based on the hybrid subdivision of the rendering tasks,a novel GPU-based data structure is proposed and a smart DMA asynchronous transfer scheme is implemented. Compared with the traditional sort-last parallel rendering,the rendering methods proposed in this paper remove the synchronous composition stage and overlap the compositing communication with rendering. The theoretical analysis and experiments demonstrate that our approach is practical and scalable for large data sets rendering and high resolution display.(3)To improve the image compositing performance on multi-GPU platform, we thoroughly investigate the NUMA-aware image compositing problem, which is the key final stage in sort-last parallel rendering. Based on a proven radix-k strategy, We find the optimal compositing algorithm, which takes advantage of NUMA architecture on the multiGPU platform. We quantitatively analyze different image compositing modes for practical image compositing, taking into account peer-to-peer communication costs between GPUs. Our experiments on various datasets show that our image compositing method is very fast, an image of 4 megapixels can be composited in about 5ms by four GPUs.(4)To reduce the overhead of network transmission in image compositing on GPU cluster, we proposed a new GPU based image encoding method, which can compress images quickly and efficiently. Compared with the RLE and ROI methods, our method was more suitable for GPU’s SIMD execution model. The experiments showed that the cost time of our method was reduced up to 55% over the RLE method and 64% over the ROI method.(5)To minimize the communication cost for direct send, we also presented a greedy approximate algorithm based on our image encoding method. We implemented the method on a 24-node GPU cluster, and the results demonstrate that our image compositing method is very fast, an image of four megapixels can be composited in about 15 ms by 24 rendering nodes, which reduced 40% time shorter than the traditional direct send.