Out-of-core rendering of large, unstructured grids

Abstract

We address the problem of rendering large unstructured volumetric grids on machines with limited memory. This problem is particularly interesting because such datasets are likely to come from computations generated on supercomputers, that is, machines with superior resources than even the most powerful workstations. Here, we present a set of techniques which can be used to render arbitrarily large datasets on machines with very little memory. In particular, we present two techniques which vary in rendering speed, disk and memory usage, ease of implementation, and preprocessing costs. The first technique is completely disk-based, and requires a small amount (actually, constant) main memory. It works by performing one scan over the file containing the unstructured grid (assuming this file has been normalized as a pre-processing step), one externalmemory sort, and a final accumulation scan which computes the image. The second technique is based on our ZSWEEP algorithm, and it is more involved both in its preprocessing, implementation, and main memory requirements, but it is substantially faster, in some cases up to an order of magnitude faster. We have implemented both techniques, and we present results on rendering a few large datasets under different conditions (image resolutions, main memory configurations, etc), and discuss the tradeoffs of using the techniques presented in this paper.

Dynamic Chunking for Out-of-Core Volume Visualization Applications

“…Many other techniques use chunks as a unit of storage for data [5] or as a unit of communication between a server that contains the data and clients that visualize the data [6]. A technique similar with chunking but applied to irregular grids is the meta-cell technique [7,8,9]. Meta-cells contain several spatially close cells, have fixed size (several disk blocks) and are loaded from disk as a unit which enables them to be stored in a space saving format.…”

A unified infrastructure for parallel out-of-core isosurface extraction and volume rendering of unstructured grids

“…In the run-time phase, we parallelize the bottleneck operations in the out-of-core isosurface extraction technique of [9] and in the outof-core ZSweep algorithm [17]: for the former, these include reading the active meta-cells from disk and scanning through the active meta-cells to generate isosurface triangles; for the latter, these include reading the meta-cells from disk in a desired front-to-back order and performing in-core ZSweep on the meta-cells read. We develop self-scheduling schemes [39] for both our parallel algorithms, which ensure load balancing among processor nodes and are very simple to implement.…”

“…We parallelize the out-of-core isosurface extraction algorithm of [9] and the out-of-core ZSweep technique [17] for direct volume rendering, using the meta-cell technique as a unified underlying building block.…”

“…Similar to the out-of-core isosurface extraction technique [9], we build an indexing data structure, namely the binary-blocked I/O interval tree (BBIO tree), which is a B-tree-like interval tree and is entirely kept in disk, to index the meta-cells, so that given an isovalue the active meta-cells that are intersected by the isosurface can be located in an I/O-optimal way. In addition, similar to the out-of-core ZSweep algorithm [17], we build a bounding-box file, which is also kept in disk and contains the bounding box and other auxiliary information for each meta-cell, to facilitate the (in-core) ZSweep algorithm [16] (which is based on sweeping a plane in the © -direction) in an I/Oefficient way. The meta-cells, the BBIO tree, and the bounding-box file only need to be constructed once, and this construction process can be performed efficiently even without the need of keeping the entire dataset in main memory; the running time for this preprocessing is the same as running external sorting a few times.…”

“…We parallelize the out-of-core isosurface extraction algorithm of [9] and the out-of-core ZSweep technique [17] for direct volume rendering, using the meta-cell technique as a unified underlying building block.Our one-time preprocessing first partitions the dataset into metacells that are stored in disk. From the meta-cells, we build a BBIO tree in disk, which can be used to speed up isosurface extraction, and a bounding-box file in disk, which is used for direct volume rendering.…”