CIC: EAGER: Scalable Algebraic Visualization in the Cloud

对大型科学可视化系统和大型科学数据库的需求正在融合。可视化系统正在配备基本的查询处理能力，观察到简单地通过图形管道“抛出数据集”忽略了现实应用程序所需的可伸缩重组、操作和过滤。此外，越来越重视现场可视化-在单一平台上执行可视化和数据处理，以避免数据传输成本并提供新的优化。特别是，云计算作为大规模数据处理平台之外的大规模可视化平台的作用还没有得到充分的发挥。针对这些压力，PI提出了一种新的方法来解决可伸缩可视化问题，该方法的灵感来自于数据库社区开发的代数查询处理技术，以及数据密集型可伸缩计算的最新进展，如MapReduce、Dryad和它们的同代人。具体地说，PI正在开发一种可伸缩可视化操作符的代数，专门用于操作和可视化网格结构的数据集。PI以前在有限元模拟中发现的非结构化网格数据集的代数上的工作提供了基础，但不支持高效的并行处理，并且不能表达某些常见的可视化任务。其他现有的系统更注重深度而不是广度，专注于针对特定硬件上的特定可视化算法进行优化，而不是一个通用的视觉分析平台，该平台可以在通常在云中找到的不共享的商用计算机集群上运行。新方法在Windows Azure平台上开发和部署，提供了一组核心的可伸缩原语，用于使用无共享体系结构操作网格数据集，能够表达各种可视化算法，并服从代数推理和优化。有关更多信息，请参阅项目网页：http://visdb.cs.washington.edu

The requirements for large-scale scientific visualization systems and large-scale scientific databases are converging. Visualization systems are being equipped with rudimentary query processing capabilities, observing that simply "throwing datasets" through the graphics pipeline ignores the scalable restructuring, manipulation, and filtering required by realistic applications. Further, there is increasing emphasis on in situ visualization --- execution of visualization and data processing on a single platform to avoid data transfer costs and afford new optimizations. In particular, the role of cloud computing as a large-scale visualizaton platform in addition to a large-scale data processing platform is underexplored.In response to these pressures, the PI proposes a novel approach to the problem of scalable visualization, one informed by the algebraic query processing techniques developed by the database community coupled with recent advances in data-intensive scalable computing latforms such as MapReduce, Dryad, and their contemporaries. Specifically, the PI is developing an algebra of scalable visualization operators specialized for manipulating and visualizing mesh-structured datasets. The PI's previous work on an algebra for unstructured grid datasets found in finite element simulations provides a foundation, but does not support efficient parallel processing and cannot express certain common visualization tasks. Other existing systems favor depth over breadth, focusing on optimizations for specific visualization algorithms on specific hardware rather than a generic platform for visual analytics that can run on the shared-nothing clusters of commodity computers typically found in the cloud. The new approach, being developed and deployed on the Windows Azure platform, provides a core set of scalable primitives for manipulating mesh datasets using shared-nothing architectures, capable of expressing a variety of visualization algorithms, and amenable to algebraic reasoning and optimization.For further information see the project web page: http://visdb.cs.washington.edu