II-NEW: GEARS - An Infrastructure for Energy-Efficient Big Data Research on Heterogeneous and Dynamic Data

大数据技术已经成功地应用于许多学科的知识发现和决策，但大数据范式的进一步发展和采用面临着几个关键的挑战。首先，难以满足现代大数据问题的性能需求，如异构和不精确数据的学习，这些问题本身难度更大，对动态数据的实时分析等性能要求也更严格。其次，功耗正在成为大数据系统及其支持的应用进一步扩展的严重限制因素。这些挑战需要一种新型的大数据系统，这种系统需要结合能够加速数据处理和访问的非常规硬件，同时降低系统功耗。因此，该项目正在开发所需的计算基础设施，以支持GEARS（节能大数据研究系统），利用异构计算和存储资源研究异构和动态数据。GEARS是一种独一无二的、高效节能的大数据研究基础设施，基于协同设计的软件和硬件组件。它使异构和动态数据的各种重要研究成为可能，并推动了计算机科学以及其他数据驱动学科的科学知识。它通过支持独特的研究和教育活动，加强了对大量本科生和研究生的培训，其中包括许多来自代表性不足的群体。最后，它还通过提供新的开源解决方案和潜在的商业应用程序来支持异构和动态数据分析，从而使社会受益。GEARS的硬件包括一个数据节点集群，这些节点配备了异构处理器和存储设备以及细粒度的电源管理功能。该软件基于广泛使用的大数据框架开发，支持跨cpu、gpu和fpga的统一编程，以及集成DRAM、NVM、SSD和HDD的深度存储层次的透明数据访问。GEARS还为学习异构和动态数据提供了新的系统和算法研究，包括：(1)使用异构加速器和优化大数据任务的性能和能效的新算法划分和调度方案；(2)新的I/O调度和数据分级策略，以提高大数据深层存储层次的性能和能效；(3)大规模张量的多相、堆芯外分解技术；(4)实时视觉分析系统，将流媒体与模拟连接起来，进行预期分析；(5)异构社会数据的多模态深度学习方法；(6)利用社交媒体实时分析社会动荡的新型计算工具；(7)利用网络大数据设计高绩效团队的可扩展、自适应、交互式团队检测与组装系统；(8)罕见类分析和异构学习算法，用于在海量异构社会数据下快速准确地发现罕见事件；(9)新的分布式机器学习框架，用于从带有不完整/噪声文本注释的web级图像/视频中学习语义知识。所有项目成果将通过项目网站（http://gears.asu.edu）与更广泛的社区共享。出版物将在网站上列出，并附有其出版商的链接。数据和软件下载将在网站上列出，并附有使用说明。源代码将托管在GitHub上，项目网站上也会列出到存储库的直接链接。

Big data technologies have been successfully applied to many disciplines for knowledge discovery and decision making, but the further growth and adoption of the big data paradigm face several critical challenges. First, it is challenging to meet the performance needs of modern big data problems which are inherently more difficult, e.g., learning of heterogeneous and imprecise data, and have more stringent performance requirements, e.g., real-time analysis of dynamic data. Second, power consumption is becoming a serious limiting factor to the further scaling of big data systems and the applications that it can support. These challenges demand a new type of big data systems that incorporate unconventional hardware capable of accelerating data processing and accesses while lowering the system's power consumption. Therefore, this project is developing the needed computational infrastructure to support GEARS (an enerGy-Efficient big-datA Research System) for studying heterogeneous and dynamic data using heterogeneous computing and storage resources. GEARS is a one-of-kind, energy-efficient big-data research infrastructure based on cohesively co-designed software and hardware components. It enables a variety of important studies on heterogeneous and dynamic data and advances the scientific knowledge in computer science as well as other data-driven disciplines. It enhances the training of a large body of undergraduate and graduate students, including many from underrepresented groups, by supporting unique research and education activities. Finally, it also benefits the society by contributing new open-source solutions and with potential commercial applications in support of heterogeneous and dynamic data analysis. The hardware of GEARS includes a cluster of data nodes equipped with heterogeneous processors and storage devices and fine-grained power management capability. The software is developed upon widely-used big data frameworks to support unified programming across CPUs, GPUs, and FPGAs and transparent data access across a deep storage hierarchy integrating DRAM, NVM, SSD, and HDD. GEARS also enables novel systems and algorithms research on learning heterogeneous and dynamic data, including (1) new algorithm partitioning and scheduling schemes for using heterogeneous accelerators and optimizing the performance and energy efficiency of big data tasks; (2) new I/O scheduling and data staging strategies for performance and energy efficiency of the deep big-data storage hierarchy; (3) multi-phase, out-of-core decomposition techniques for large-scale tensors; (4) real-time visual analytics system that links streaming media with simulations for anticipatory analytics; (5) multi-modal deep learning methods with heterogeneous social data; (6) new computational tools for real-time analysis of social unrest using social media; (7) scalable, adaptive, and interactive team detection and assemble system for designing high-performing teams using big network data; (8) rare category analysis and heterogeneous learning algorithms for fast and accurate rare event discoveries with large and heterogeneous social data; and (9) new distributed machine learning framework for learning semantic knowledge from Web-scale images/videos with incomplete/noisy textual annotations. All project results will be shared with the broader community via the project website (http://gears.asu.edu). Publications will be listed on the website with links to their publishers. Data and software downloads will listed on the website with instructions on how to use them. Source code will be hosted on GitHub and a direct link to the repository will also be listed on the project website.