SGER: Modeling Memory Access Patterns of Geometry Processing Algorithms

SGER：几何处理算法的内存访问模式建模

• 批准号: 0738401
• 负责人: Gopi Meenakshisundaram
• 金额: $ 6.31万
• 依托单位: University of California-Irvine
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2007
• 资助国家: 美国
• 起止时间: 2007-10-01 至 2008-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0738401&HistoricalAwards=false
• 关键词: SGER; Modeling; Memory; Access; Patterns

Current graphics and visualization systems have to be built such that they can handle gigantic data sets. Such data sets include large scientific simulations such as nuclear and power simulations, data relevant to national priority and homeland security, and digital models of defense and commercial equipments such as tanks, aircraft, ships, and power plants. Such large data sets cannot fit into the main memory of the machines. Hence, the performance of the visualization systems depends on how efficiently they can process this data segments and still provide a holistic visualization for efficient and correct decision making. This project involves fundamental research in the analysis of methods that process these large geometry data sets for computer graphics and visualization applications. Using this analysis we model the data access pattern of common geometry processing algorithms. Such models can be used to organize data coherently in the secondary storage so that the access time of the data can be reduced. This will improve the performance of the graphics and visualization systems. The coarse data analysis systems derive aggregate information from the data and hence are useful in streaming applications and out-of-core implementations; fine data analysis systems are interested in individual data points and their performance is dictated by data access patterns. In this research, we investigate if there is any natural grouping or partitioning of primitives that describes the data access patterns of most common geometry processing algorithms. We explore the existence of a function that would optimize the grouping of primitives and thus benefit a large class of geometry processing algorithms. This study will enable us to suggest an optimal layout for geometric data that would work best for common geometric algorithms.

当前的图形和可视化系统必须能够处理庞大的数据集。此类数据集包括大型科学模拟，例如核和电力模拟、与国家优先事项和国土安全相关的数据，以及坦克、飞机、船舶和发电厂等国防和商业设备的数字模型。如此庞大的数据集无法装入机器的主存储器。因此，可视化系统的性能取决于它们处理这些数据段的效率，并且仍然为有效和正确的决策提供整体可视化。该项目涉及处理这些大型几何数据集用于计算机图形和可视化应用的方法分析的基础研究。利用这一分析，我们建立了常见的几何处理算法的数据访问模式。这样的模型可以用于在辅助存储器中连贯地组织数据，使得可以减少数据的访问时间。这将提高图形和可视化系统的性能。 粗数据分析系统从数据中获得聚合信息，因此在流应用和核外实现中很有用;细数据分析系统对单个数据点感兴趣，其性能由数据访问模式决定。在这项研究中，我们调查是否有任何自然的分组或分区的图元，描述了最常见的几何处理算法的数据访问模式。我们探讨存在的一个功能，将优化分组的图元，从而有利于一大类的几何处理算法。这项研究将使我们能够提出一个最佳的布局几何数据，将最好的工作，为共同的几何算法。