Task-based Visualization Methods for Scalable Analysis of Large Data Sets

除了理论和实验之外，技术和自然现象的模拟已经成为现代科学和工程的第三大支柱。使用科学的可视化技术分析结果模拟数据集是该方法的重要组成部分。大规模并行高性能计算机作为一种仿真平台，其执行单元（内核）数量不断增加。随着所产生的数据量与可用的计算能力成比例地增长，开发可扩展的并行可视化技术变得越来越重要。在商品硬件（pc）中也可以观察到类似的发展；因此，在中长期内，这些体系结构也需要相应的方案。迄今为止，对可视化算法并行化的研究主要集中在单个方法上。对于一些单独的技术，在可伸缩性和效率方面已经显示出良好的结果。相反，可视化的实际应用通常需要多种方法的组合。关于这种组合的效率的陈述，特别是在涉及不同并行化范式的情况下，目前是不可能的。这对选择合适的算法是一个很大的障碍，特别是当一个不合适的选择可能导致显著甚至令人望而却步的低效率时。近年来，基于任务的并行化已经被广泛应用。这里，算法被表述为一组任务，每个任务代表计算的一个原子步骤。任务之间的依赖关系是显式建模的。通过这种方法，只要不违反依赖关系，就可以选择一个基本任意且并发的任务执行序列，以优化计算。拟议的项目旨在调查已建立的可视化技术的基于任务的公式。总体目标是提高可视化在当前和未来架构上对大型数据集的适用性和实用性。在这方面，将考虑不同类别的可视化算法对基于任务的公式的一般适用性，以及由此产生的效率和运行时行为。特别是，不同技术的组成，因为它经常在实践中应用，将进行审查。申请人的初步工作表明，这种方法是有希望的。

In addition to theory and experiment, simulation of technical and natural phenomena has become the third pillar of modern science and engineering. The analysis of resulting simulation data sets using scientific visualization techniques is an essential component of this approach. As a platform for simulation, massively-parallel high-performance computers are employed with a steadily increasing number of execution units (cores). As the resulting amount of data is growing proportionally to available computing power, the development of scalable, parallel visualization techniques is of ever increasing importance. A similar development can be observed in commodity hardware (PCs); thus corresponding schemes will also be needed on these architectures in the medium to long term future. Research into parallelization of visualization algorithms has thus far focused mostly on individual approaches. For several individual techniques, good results regarding scalability and efficiency have been shown. In contrast, real-world applications of visualization often require a combination of methods. Statements about the efficiency of such a combination, especially where different parallelization paradigms are concerned, are currently not possible. This is a significant hindrance towards the choice of suitable algorithms, especially as an unsuitable choice may result in significant or even prohibitive inefficiencies.In recent years, task-based parallelization has been established as useful across a wide range of applications. Here, an algorithm is formulated as a set of tasks, each of which represents an atomic step of the computation. Dependencies between tasks are modeled explicitly. Through this, as long as dependencies are not violated, a mostly arbitrary and concurrent sequence of execution of the tasks can be chosen in order to optimize the computation.The proposed project aims at the investigation of task-based formulations of established visualization techniques. The overarching goal is to increase applicability and utility of visualization for large data sets on contemporary and future architectures. In this regard, the general suitability of different classes of visualization algorithms towards a task-based formulation will be considered, as well as the resulting efficiency and runtime behavior. In particular, the composition of different techniques, as it is often applied in practice, will be examined. Initial work by the applicants indicates that this approach is promising.