Y. Hong, T. Peterka, and H.-W. Shen, "Histogram-Based I/O Optimization for Visualizing Large-Scale Data," Preprint ANL/MCS-P1698-1209, December 2009. [pdf]
We present an I/O optimization method for parallel volume rendering based on visibility and spatial locality. The combined metric is used to organize the file layout of the dataset on a parallel file system. This reduces the number of small, noncontiguous I/O operations and improves load balance among I/O servers. The net result is reduced I/O time. Since large-scale visualization is data-intensive, overall visualization performance improves using this method. This paper explains the preprocessing of data blocks to compute feature vectors and the storage organization based on them. Run-time performance is analyzed with a variety of transfer functions, view directions, system scales, and datasets. Our results show significant performance gains over file layouts based on space-filling curves.