Isosurface extraction in large scientific visualization applications using the I/O-filter technique

For large scientiic visualization applications, it is often impossible to hold the entire datasets in main memory, even on supercomputers. Previously, we proposed the I/O-lter technique, which is the rst I/O-optimal method for the problem of isosurface extraction in scientiic visualization. I/O-lter works by indexing and reorganizing the datasets in disk, so that isosurface can be extracted with a very small amount of disk I/O's. The main advantage of this approach is that datasets much larger than main memory can be visualized very eeciently, possibly even on low-end machines. The original I/O-lter technique uses the I/O-optimal interval tree of Arge and Vitter as the indexing data structure, together with the isosurface engine from Vtk (one of the currently best visualization packages). The main shortcoming of this approach was the overheads of the disk scratch space and the preprocessing time necessary to build the data structure, and of the disk space needed to hold the data structure. In this paper, we improve the rst version of I/O-lter by reducing its data structure size, the disk scratch space and the preprocessing time, while keeping the isosurface query time the same. We achieve the improvements through a new implementation of I/O-lter, by replacing the interval tree with the metablock tree of Kanellakis et al., which is also I/O-optimal for query and space. In the process, we propose two simple preprocessing algorithms for static metablock tree that are I/O-optimal. We give the rst implementation of metablock tree, and compare the practical performance between metablock tree and interval tree under the same framework of I/O-lter with real-world test data. Our experiments provide detailed quantitative evaluation of the implementations of the two trees, and lead to new insights to their properties. As mentioned, the metablock-tree implementation results in performance improvements of the I/O-lter method. Moreover, for our test datasets larger than main memory, the isosurface queries in both implementations of I/O-lter are about two orders of magnitude faster than those in Vtk's original implementation, showing that I/O-lter is an excellent isosurface technique for large scientiic visualization applications. With this technique, we achieve fast isosurface generation for large datasets at a disk memory cost (in dollars) that is 10-20% of the cost associated with having enough main memory to hold the entire dataset.