Abstract

Traditionally graphics clusters have been employed in real-time visualization of large geometric models (many millions of 3D points). Data parallel approaches have been the obvious choice when it comes to breaking up the computations over multiple processors. In the recent years, programmable graphics processing hardware units enabling a myriad of specialized algorithms to be programmed in the graphics unit have gained widespread acceptance. Today, every processing node in a graphics cluster has two powerful and fully programmable processors; a CPU (Central Processing Unit) and a GPU (Graphics processing unit), providing us with the opportunity of organizing the distribution of graphics computations over a cluster in more specialized ways to suit an application's needs. As part of our research we have studied and explored in detail functionality distribution in large graphics applications. Our thesis is that existing data parallel distribution approaches can be further enhanced by the addition of application specific functionality distribution achieved through suitable programming of the graphics hardware. To demonstrate this, we have chosen the upcoming application domain of point based rendering and implemented a functionality distributed point based rendering pipeline. The performance improvements are impressive. To the best of our knowledge, ours is the first attempt anywhere to devise a functionality distribution scheme for point based rendering and demonstrate the significant performance improvement achieved. We discuss the merits and limitations of such a distribution scheme by comparing it against traditional data parallel and single node schemes.