Scalable rendering on PC clusters

Abstract

This case study presents initial results from research targeted at the development of cost-effective scalable visualization and rendering technologies. The implementations of two 3D graphics libraries based on the popular sort-last and sort-middle parallel rendering techniques are discussed. An important goal of these implementations is to provide scalable rendering capability for extremely large datasets (>> 5 million polygons). Applications can use these libraries for either run-time visualization. by linking to an existing parallel simulation. or for traditional postprocessing by linking to an interactive display program. The use of parallel, hardware-accelerated rendering on commodity hardware is leveraged to achieve high performance. Current performance results show that, using our current hardware (a small 16-node cluster), we can utilize up to 850/. of the aggregate graphics performance and achieve rendering rates in excess of 20 million polygons/second using OpenGL@ with lighting, Gouraud shading. and indi\iduallj specified triangles (not [-stripped).

Application of interactive parallel visualization for commodity-based clusters using visualization APIs

“…The latest Chromium paper [10] reports on the implementation of components (stream processing units, or SPUs) that can be used in sort-last parallel graphics applications. More on the sort-last parallel rendering approach can be found in [13] and [23].…”

GEVP: Grid Enabled Visualization Pipeline

“…Several methods and techniques were introduced in the past few years to tackle the problem of providing computational power required for visualization operations. The introduced techniques result most often in building a cluster of nodes to provide the necessary computational power [1], [2], [3], [4], [5], [6], [6]. Other implementation focuses on remote visualization such as [7], [8], [9], [10], [11], [12].…”

A divided-screenwise hierarchical compositing for sort-last parallel volume rendering

“…In particular, some recent works [12,22] have challenged to real-time rendering of at least 512 3 voxel volumes. The above previous methods are classified as sort-last methods [11], which partition the volume into subvolumes and distribute the subvolumes to each processor.…”

“…methods [2,4,7,8,12,13,16,18,20,22,23]. In particular, some recent works [12,22] have challenged to real-time rendering of at least 512 3 voxel volumes.…”

“…The BS method completes the image compositing with log n stages and transmits at most 2.43n 1/3 p pixels, where p is the number of pixels in the final image. Many sort-last methods [10,16,22,23] use this method.…”