LOD Based Parallel Volume Rendering

Volume rendering is one of the main methods in scientific visualization field. The intrinsic information of volume dataset can be visualized interactively by volume rendering. Therefore the users can get information from the rendered image. Volume data can be acquired from different sources, such as data from Computed Tomography (CT), Magnetic Resonance Imaging (MRI) scanners or computational fluid dynamics (CFD).With the development of science and technology, the size of generated volume data has been getting larger and larger. Such large scale datasets pose immense challenges to volume rendering. There are two important and popular techniques for real time visualization of large datasets, parallel volume rendering and multiresolution rendering, which could be combined as parallel multiresolution volume rendering. Level of detail (LOD) technology can supply different level resolution data for different need. As visualization is an iterative and exploratory process, rendering a lower resolution of data sometimes is sufficient for the purpose of obtaining an overview of the data before the user can query further details in selected regions. This method can improve rendering efficiency.Parallel volume rendering solves the large data visualization problem by splitting both the data and rendering task among cluster nodes. In this paper, we adopt sort-last parallel volume rendering technology and propose a series methods to improve rendering efficiency. Memory mapped files technology is used to speed up disk access. After each nodes complete their rendering, we adopt binary-swap technology to composite all sub images in parallel. We use slice template technology to render any brick by simply computing the offset along the view direction and then translating the template slice vertices along the direction opposite of the view direction.To achieve good overlapping between rendering and communication, each node runs three threads. The first thread handles rendering, the second thread handles compositing and the third handles data loading. However, when using LOD method in parallel rendering, load unbalance becomes a challenging issue. We solve this issue by using kd tree data structure to dynamically distribute the data among the rendering nodes according to the load of the previous frame. The result show that good load balancing can be achieved.Finally, an efficient parallel volume rendering framework has been carefully designed and developed. Our framework can obtain good rendering speed and high resolution image showed in large scale display system after testing by Visible Human Image Data Set.