Collaborative Research:XPS:CLCCA: Performance Portable Abstractions for Large-Scale Irregular Computations

正在进行的技术趋势正在加速科学发现，使研究人员能够在从计算生物学到社会网络的各个领域产生大量数据。迫切需要使用适当的抽象和并行运行时来轻松、快速地从这些数据中提取有用的内容。在该项目下开展的工作旨在使具有动态结构和不规则依赖关系的应用程序更容易地使用“大数据”计算，从而使科学计算特别是计算生物学方面的进步成为可能。该项目通过开发编程抽象来扩展科学计算的最新水平，以暴露--并在运行时优化--在大型、不规则的应用程序中可用的并行性。并行对于从日益增长的科学数据中提取有用信息是必不可少的，但许多问题领域的数据结构和访问的不规则性使得高效的并行化变得困难。在编程模型的级别上，该项目通过确定重要的新应用程序类的设计模式来解决不规范的挑战--特别是那些使用树和图来表示和访问数据但在遍历中展示了一些结构的应用程序。在运行时系统级别，它正在开发支持和利用新模式的计算引擎，利用公开的结构自动和动态地将计算任务映射到硬件节点。

Ongoing technology trends are accelerating scientific discovery by allowing researchers to generate enormous quantities of data, in domains ranging from computational biology to social networks. There is an urgent need to make it easy and fast to extract useful content from this data using appropriate abstractions and parallel runtimes. Work conducted under this project aims to make "big data" computing more readily available to applications with dynamic structure and irregular dependencies, thereby enabling advances in scientific computing in general and computational biology in particular.This project extends the state of the art in scientific computing by developing programming abstractions to expose -- and run-time optimizations to exploit -- the parallelism available in large, irregular applications. Parallelism is essential for the extraction of useful information from ever increasing volumes of scientific data, but the irregularity of data structure and access in many problem domains makes efficient parallelization difficult. At the level of the programming model, the project addresses the challenge of irregularity by identifying design patterns for important new classes of applications -- in particular, those that use trees and graphs for data representation and access but demonstrate some structure in the traversal. At the level of the run-time system, it is developing computational engines that support and exploit the new patterns, leveraging the structure exposed to automatically and dynamically map computational tasks to hardware nodes.