CAREER: Communication-efficient and topology-aware designs for geo-spatial analytics on heterogeneous platforms

地理空间数据集的数量、复杂性和异质性都在不断增长。当前的地理空间大数据时代与数据科学的进步相结合，使公共和私人组织能够获得新的见解并建立新的能力。各机构和科学界依靠空间数据管理和分析来获得见解并制定可行的计划。该项目将为空间数据分析设计高效算法，作为解决不同科学领域广泛研究问题的关键工具。这些算法和实现将利用高性能计算（HPC）来加速使用空间数据的领域中的计算和数据密集型工作负载。所选择的计算几何算法广泛适用于利用空间数据的各个领域，包括社会学、流行病学、病理学、气候科学、太阳物理学等。本科生和研究生将被训练进行高性能计算研究。将通过平行计算和空间计算领域的教育讲习班编写和传播与研究议程有关的教育材料。地理空间分析文献中的大多数现有工作都局限于利用基本的线程级和数据并行性。现有的I/O高效算法主要基于共享内存并行性和磁盘上的数据。由于处理器之间的通信而产生的数据移动是运行在高性能计算（HPC）系统上的应用程序产生的主要成本。我们提出了基于异构平台的地理空间分析的高效通信设计。第二个研究领域是空间计算系统的拓扑感知设计，它可以无缝地利用数据处理单元（dpu）。DPU是一种新型协处理器，是可编程智能网络接口卡的后继产品。该项目将为基于过滤和细化的空间分析内核开发分布式内存并行算法。由于计算节点的异构性，为地理空间应用程序的设计空间探索提供了新的机会。将研究基于通信与计算权衡的新并行化技术。新设计将利用dpu中的网络拓扑感知空间数据分区和远程直接内存访问（RDMA）功能来实现分层过滤和细化技术。并行算法的实现将被纳入公开可用的MPI GIS软件。这些库将在异构计算系统中实现可扩展。为了产生更广泛的影响，新的软件库将与CyberGIS地理空间软件包集成。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Geospatial datasets are growing in volume, complexity, and heterogeneity. The current era of geospatial big data combined with advancements in data science is enabling public and private organizations to gain new insights and build new capabilities. Various agencies and scientific communities rely on spatial data management and analysis to gain insights and produce actionable plans. This project will design efficient algorithms for spatial data analytics that can serve as a crucial tool for solving a wide set of research problems from different scientific areas. The algorithms and implementations will leverage high performance computing (HPC) to speedup compute and data intensive workloads in domains that employ spatial data. The computational geometry algorithms chosen are broadly applicable in various domains that make use of spatial data, including sociology, epidemiology, pathology, climate science, solar physics, etc. Undergraduate and graduate students will be trained to carry out high performance computing research. Educational materials relevant to the research agenda will be developed and disseminated through educational workshops in the parallel computing and spatial computing domains.Most of the existing work in the literature for geospatial analytics is limited to exploiting basic thread-level and data parallelism. Existing work on I/O efficient algorithms are mostly based on shared memory parallelism and data on disk. Data movement due to communication among processors is a dominant cost incurred by an application running on a high performance computing (HPC) system. We propose communication efficient designs for geospatial analytics on heterogeneous platforms. A second area of research is topology aware designs for spatial computing systems that can seamlessly leverage Data Processing Units (DPUs). A DPU is a new class of coprocessor that evolved as a successor to programmable smart network interface cards. This project will develop distributed memory parallel algorithms for filter and refine-based spatial analytics kernels. Motivated by the heterogeneity in a compute node, there is a new opportunity for design space exploration for geospatial applications. New parallelization techniques based on communication vs. computation tradeoffs will be investigated. The new design will leverage network topology aware spatial data partitioning and remote direct memory access (RDMA) capability in DPUs for implementing hierarchical filter and refine techniques. The implementation of parallel algorithms will be incorporated into publicly available MPI GIS software. These libraries will be implemented to be scalable in heterogeneous computing systems. For broader impact, the new software libraries will be integrated with CyberGIS geospatial packages.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.