Visualization and parallel I/O at extreme scale

In our efforts to solve ever more challenging problems through computational techniques, the scale of our compute systems continues to grow. As we approach petascale, it becomes increasingly important that all the resources in the system be used as efficiently as possible, not just the floating-point units. Because of hardware, software, and usability challenges, storage resources are often one of the most poorly used and performing components of today's compute systems. This situation can be especially true in the case of the analysis phases of scientific workflows. In this paper we discuss the impact of large-scale data on visual analysis operations and examine a collection of approaches to I/O in the visual analysis process. First we examine the performance of volume rendering on a leadership-computing platform and assess the relative cost of I/O, rendering, and compositing operations. Next we analyze the performance implications of eliminating preprocessing from this example workflow. Then we describe a technique that uses data reorganization to improve access times for data-intensive volume rendering.