ICE-T:RI: Towards End-to-End Resource Optimization for Time-Critical Computing Using Reinforcement Learning and Program Analysis

数据密集型、时间关键型应用程序会生成需要快速分析的海量数据。时间关键型计算的资源优化面临着许多挑战，包括对编程技能的高要求，难以确定合适的并行度，以及考虑多个优化目标的资源分配的高度复杂性。为了帮助设计更高效的应用程序，该项目使用强化学习和程序分析技术对时间关键型计算的端到端资源优化进行了研究。该方法结合了程序分析的资源请求优化和考虑时间关键特性的强化学习的资源调度。该项目将通过强化学习和语义感知的程序分析来加深我们对解决问题需求所涉及的挑战的理解。本项目试图做出以下新的贡献：(1)为数据密集型应用设计语义感知优化，包括两个阶段。离线阶段使用静态程序分析来分析大数据系统原语和用户定义函数，以生成参数化数据框架，并通过基于规则和基于成本的模型修复部分性能缺陷。在线阶段使用动态程序分析，实例化基于执行度量的参数化框架以修复性能问题；(2)探索扩大并行度和最小化跨计算节点的数据量之间的权衡；以及(3)设计基于强化学习的资源分配模型。该项目启动了中佛罗里达大学和荷兰阿姆斯特丹大学之间的研究合作。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Data-intensive, time-critical applications generate an enormous amount of data that needs to be analyzed quickly. Resource optimization for time-critical computing faces many challenges including high demand on programming skills, difficulty in determining suitable parallelism degree, and great complexity in making resource allocation considering multiple optimization targets. To help in designing more efficient applications, this project investigates end-to-end resource optimization for time-critical computing using reinforcement learning and program analysis techniques. The approach integrates both resource request optimization by program analysis, and resource scheduling by reinforcement learning with consideration of time-critical features. The project will enhance our understanding of the challenges involved in addressing problem demand with reinforcement learning and semantics-aware program analysis.This project seeks to make the following novel contributions: (1) designing a semantics-aware optimization for data-intensive applications including two stages. The offline stage uses static program analysis for analyzing big data system primitives and user-defined functions to generate a parameterized data framework and fix partial performance flaws by rule-based and cost-based models. The online stage uses dynamic program analysis that instantiates a parameterized framework based on execution metrics to repair performance problems; (2) exploring the trade-off between enlarging parallelism degree and minimizing the amount of data shuffling across computing nodes; and (3) designing a reinforcement learning based model for resource allocation.This project initiates a research collaboration between the University of Central Florida and the University of Amsterdam, Netherlands.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.