Implementation Techniques Building Scalable Cluster-Based 3D Graphics Engines

近年来，高性能计算中最明显的体系结构趋势是出现了(现在)工作站集群或网络体系结构，在这种体系结构中，通过高速系统区域互连连接的一组现成的高性能台式机可以高效地执行并行程序，有时性能优于定制的多处理器或多计算机，后者的成本降低了数量级。集群计算背后的关键驱动力是微处理器性能的不可阻挡的进步，而这反过来又受到个人电脑行业的规模经济的推动。我们正在3D图形硬件领域见证一场类似的技术革命。在过去五年中，低成本3D显卡的性能正在以每年80%的惊人速度提高，如果继续保持这种速度，预计将在2003年超过高端图形引擎，如Sili-con Graphics的那些引擎。尽管有这些令人难以置信的性能进步，但这些显卡中的大多数仍然相对便宜。本项目旨在回答的中心研究问题是，基于集群体系结构和3D显卡构建并行3D图形引擎是否可行，其整体性能是否可以随着CPU/图形处理器的数量而扩展，如果可行，使能算法和体系结构技术是什么。该项目将通过为基于集群的并行3D图形系统SUnder设计、实现和评估高效软件来解决这些问题，该系统构建在Linux和Mesa之上。

The most apparent architectural trends in high-performance computing in the recent years is the emergence of cluster or network of workstations (NOW) architectures, where a set of off-the-shelf high-performance desktop machines connected by a high-speed system-area interconnect can efficiently execute parallel programs and sometimes out-perform custom-built multiprocessors or multicomputers that axe orders of magnitude more expensive. The key driving force behind cluster computing is the unstoppable advances in microprocessor performance, which in turn is enabled by the economies of scale in the PC industry. We are witnessing a similar technological revolution in 3D graphics hardware arena. The performance of low-cost 3D graphics cards is improving at an astonishing rate of 80% per year within the last five years, and if continuing at this rate, is expected to out-pace high-end graphics engines such as those from Sili-con Graphics in 2003. Despite these incredible performance progresses, most of these graphics cards remain relatively inexpensive. The central research questions that this project aims to answer are whether it is feasible to build parallel 3D graphics engines that are based on the clustering architecture and 3D graphics cards and whose overall performance can scale with the number of CPUs/graphics processors, and if so, what the enabling algorithmic and architectural techniques axe. The project will address these questions by designing, implementing, and evaluating efficient software for a cluster-based parallel 3D graphics system called Sunder, which is built on top of Linux and Mesa.