Feature-Based Adaptive Block Partition Method for Data Prefetching in Streamline Visualization

With the increasing size of flow field, a challenge in streamline visualization arises that the memory of calculation node cannot accommodate the entire required data. To solve this problem, out-of-core technique divides the flow field into blocks and read block on demand of computing. Data prefetching is a frequent out-of core method to reduce the affection of the gap between I/O and calculation speed, while the performance is coherent with prefetching hit rate. In this paper, we focus on how to improve the prefetching hit rate to increase the data prefetching efficiency by changing the style of flow field partitioning, and present a novel feature-based dynamic block partition method that divides data to blocks of different sizes. The key of our method is first to compute the feature attributes of the field, and then determine the partitioning points by specific operations to divide feature regions more finely. It is easy to apply our approach to replace block partition part of all state-of-the-art prefetching algorithms. Experimental results show that the major quality measurement of our partitioning strategy for prefetching is much better than the traditional methods, with an increase of about 10%in both prefetch hit rate and effective rate.