CAREER: Robust Processing of Data Streams in Real Time

该项目研究了医疗传感器数据流集合的处理调度问题，其目标是提供高可信度的每数据包服务保证，这些服务保证对流生成中的可变性和处理流的分布式资源集的负载变化具有鲁棒性。考虑到最优流调度的NP-Hard复杂性和处理随机流的需要，服务保证是概率性的。该方法利用统计和机器学习技术，利用独立于应用程序和依赖于应用程序的混合特性，在具有动态利用率配置文件的资源集合上自适应地编排流处理，同时保留在高负载下优先处理多个并发应用程序的能力。来自临床和辅助生活环境的数据用于评估解决方案的有效性。医疗保健和国土安全可以从这项研究以及实验科学中受益。医疗保健成本的飙升与医疗和辅助生活环境中电子监控设备的普及相吻合，这些设备会产生患者数据流。及时监测和分析这些流可以及早发现紧急情况并改善患者的预后，但失败可能是致命的。自动化医疗数据流处理的效率和健壮性的提高转化为更低的成本和更好的结果。在国土安全领域也有类似的机会，必须对化学和生物传感器数据进行实时处理，以进行威胁评估。作为这项研究的一部分，开源软件可以在任意数量的机器上配置，以处理各种设置下的大量数据流，这降低了需要在应用程序中处理观测数据的科学家的入门门槛。该项目还将为学生提供教育机会，并开展旨在改善美洲土著学生数学概念同化的中学外展活动。

This project investigates the problem of scheduling the processing of collections of streams of medical sensor data, with a goal of providing high-confidence per-packet service guarantees that are robust to variability in the stream generation and concomitant changes in the loads at the distributed set of resources where streams are processed. Given the NP-Hard complexity of optimal stream scheduling and the need to handle streams that are stochastic in nature, the service guarantees are probabilistic. The approach makes use of statistical and machine learning techniques, harnessing a mix of application-independent and application-dependent features to adaptively orchestrate stream processing over a collection of resources with dynamic utilization profiles while retaining the ability to prioritize processing under heavy load, for multiple concurrent applications. Data from clinical and assisted living settings are used to evaluate the efficacy of solutions.Health care and homeland security can benefit from this research, as well as experimental science. Skyrocketing healthcare costs have coincided with the proliferation of electronic monitoring devices in medical and assisted living environments, which generate data streams of patient data. Timely monitoring and analysis of these streams can detect emergencies early and improve patient outcomes, but failure can be fatal. Improvements in the efficiency and robustness of automated medical data stream processing translate to lower costs and improved outcomes. There are analogous opportunities in homeland security, where chemical and biological sensor data must be processed in real-time for threat evaluation. Open-source software produced as part of this research, which can be configured over an arbitrary number of machines to process a large number of streams in a variety of settings, lowers entry barriers for scientists who need to process observational data in their applications. The project will also provide educational opportunities for students, and middle school outreach activities targeted at improving assimilation of mathematical concepts among Native American students.