Collaborative Research: PPoSS: Planning: Extreme-scale Sparse Data Analytics

图形数据结构用于存储和操作关系数据。张量是二维矩阵表示的高阶推广。图和张量都用于探索性和自动化数据分析。应用领域包括网络安全、复杂系统分析和个性化医疗保健。在这些领域中存在用于典型数据分析任务的无数已知算法。例如，图中的社区识别问题，指的是自动识别图中的良好连通的顶点组，有几十种算法。类似于矩阵中的奇异值分解，存在几种具有不同用例的张量分解。图算法和张量因子分解都使用受矩阵计算启发的计算机存储格式。该项目的重点是数据分析用例，这些用例导致大规模的图形和张量，需要并行和分布式处理。该项目的新颖之处在于识别和开发跨越多个图计算和张量因子分解的统一并行算法设计原则。在规划阶段，几个重点研究任务将探索八个统一的主题。该项目旨在为端到端流数据分析系统奠定基础，该系统的性能可与当前高端工作站和超级计算机上高度调整的静态图分析基准相媲美。研究人员的多学科专业知识涵盖高性能计算，理论和算法，计算机体系结构，编程语言和编译器。交叉研究的目标包括可概括的原则，协调内部和节点间的通信，利用层次并行，本地化增强策略，自动性能调整的多种方法。规划阶段的软件工件可以形成新数据分析基准的基础。研究人员将把研究结果纳入他们教授的课程中。在规划阶段让来自国家实验室和行业的专家参与，将有助于巩固未来的大规模努力。研究人员将利用现有的机构计划，扩大参与计算研究。这个奖项反映了NSF的法定使命，并已被认为是值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估的支持。

The graph data structure is used for storing and manipulating relational data. Tensors are a higher-order generalization of the two-dimensional matrix representation. Both graphs and tensors are used in exploratory and automated data analysis. Applications areas include cybersecurity, complex system analysis, and personalized healthcare. There exist a myriad of known algorithms for typical data analysis tasks in these areas. For instance, the problem of community identification in graphs, referring to automatically identifying well-connected groups of vertices in graphs, has dozens of algorithms. Analogous to the singular value decomposition in matrices, several tensor factorizations exist with diverse use-cases. Both graph algorithms and tensor factorizations use computer storage formats inspired by matrix computations. This project focuses on data analysis use-cases that result in large-scale graphs and tensors, necessitating parallel and distributed processing. The project's novelties are in identifying and developing unifying parallel algorithm design principles that span multiple graph computations and tensor factorizations. In the planning stage, several focused research tasks will explore eight unifying themes.The project aims to develop the foundations for an end-to-end streaming data analytics system with performance comparable to highly tuned static graph analysis benchmarks on current high-end workstations and supercomputers. The investigators' multi-disciplinary expertise span high-performance computing, theory and algorithms, computer architecture, and programming languages and compilers. The cross-cutting research aims include generalizable principles to orchestrate intra- and inter-node communication, multiple approaches for exploiting hierarchical parallelism, locality-enhancing strategies, and automatic performance tuning. The software artifacts from the planning stage could form the basis for new data analytic benchmarks. The investigators will incorporate research findings into the courses they teach. Engaging experts from the national laboratories and the industry in the planning stage will help solidify future large-scale efforts. The investigators will leverage and contribute to existing institutional programs that broaden participation in computing research.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.