Collaborative Research: CNS Core: Medium: Learning to Cache and Caching to Learn in High Performance Caching Systems

合作研究：CNS 核心：中：学习缓存以及在高性能缓存系统中学习缓存

• 批准号: 1955777
• 负责人: Vijay Subramanian
• 金额: $ 17.5万
• 依托单位: Regents of the University of Michigan - Ann Arbor
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2023-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1955777&HistoricalAwards=false
• 关键词: Collaborative; Research; CNS; Core; Medium

Caching is fundamental to cloud computing and content distribution, and is important to the vast number of applications and services they support. Crucial performance metrics of a caching algorithm are its ability to quickly and accurately learn a changing popularity distribution. However, there is a serious disconnect between empirical studies using real-world traces that account for popularity changes, and analytical performance analysis results that assume a fixed popularity. A basic goal of this project is to develop a methodology based on online learning and reinforcement learning for caching algorithm design with provable performance guarantees. This enables the systematic design of caching algorithms that can be tailored to a variety of application contexts. The use-case of these algorithms is in high performance caching networks that support large-scale cloud applications and services. Emulation of high-performance caching systems to leverage and to empirically evaluate the online learning algorithms developed supports this goal, and provides a real-world context for the methodology developed. The results will also enhance the performance of content distribution platforms. At the same time the project develops fundamental theories that pertain to the area of machine learning, specifically to online learning. This project aims at optimally utilizing locally available memory and computing resources of caches, while ensuring provably good performance via fast and accurate learning of content popularity. This requires the conjunction of several mathematical tools to analyze online learning algorithms, as well as strong systems development skills to make the algorithms a reality. The project addresses these key challenges in two main themes. The first theme focuses on systematic design of distributed online learning in networks of caches using collaborative filtering for distributed identification of popular content, and multi-agent reinforcement learning for joint learning and content placement. The second theme focuses on building high performing caching systems using the algorithms developed in the first theme, and quantifying the impacts of the algorithms on real-world applications such as Hipster Shop, an open-source e-commerce website, and Spark data-analytics job pipelines. The immediate impact of this project is in creating high performance caching schemes that apply to cloud computing and content distribution networks. This project also advances the fundamental theory of online learning. The project includes an education plan focusing on machine learning and caching, and outreach in the form of summer camps and seminars for high school students.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.

缓存是云计算和内容分发的基础，对于它们支持的大量应用程序和服务也很重要。 缓存算法的关键性能指标是其快速准确地学习不断变化的流行度分布的能力。然而，使用现实世界痕迹来解释流行度变化的实证研究与假设流行度固定的分析性能分析结果之间存在严重脱节。 该项目的基本目标是开发一种基于在线学习和强化学习的方法，用于具有可证明的性能保证的缓存算法设计。 这使得缓存算法的系统设计能够适应各种应用程序环境。这些算法的用例是支持大规模云应用程序和服务的高性能缓存网络。 模拟高性能缓存系统以利用和实证评估所开发的在线学习算法支持这一目标，并为所开发的方法提供真实世界的背景。 结果还将提高内容分发平台的性能。 与此同时，该项目开发了与机器学习领域（特别是在线学习）相关的基础理论。该项目旨在最佳地利用本地可用的内存和缓存的计算资源，同时通过快速准确地了解内容流行度来确保可证明的良好性能。这需要结合多种数学工具来分析在线学习算法，以及强大的系统开发技能来使算法成为现实。该项目通过两个主题解决这些关键挑战。第一个主题侧重于缓存网络中分布式在线学习的系统设计，使用协作过滤来分布式识别流行内容，以及用于联合学习和内容放置的多智能体强化学习。第二个主题侧重于使用第一个主题中开发的算法构建高性能缓存系统，并量化算法对现实应用程序（例如 Hipster Shop、开源电子商务网站和 Spark 数据分析）的影响作业管道。该项目的直接影响是创建适用于云计算和内容分发网络的高性能缓存方案。该项目还推进了在线学习的基础理论。该项目包括一项专注于机器学习和缓存的教育计划，以及针对高中生的夏令营和研讨会形式的推广。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评估进行评估，被认为值得支持。影响审查标准。

项目成果

Rarest-First with Probabilistic-Mode-Suppression (RFwPMS)

具有概率模式抑制的稀有优先 (RFwPMS)

• 发表时间: 2024-01
• 期刊: IEEE transactions on information theory
• 影响因子: 2.5
• 作者: Nouman Khan;Mehrdad Moharrami;Vijay G. Subramanian
• 通讯作者: Vijay G. Subramanian

Bayesian Persuasion in Sequential Trials

序贯试验中的贝叶斯说服

• DOI: 10.1007/978-3-030-94676-0_2
• 发表时间: 2022-01
• 期刊: Web and Internet Economics. WINE 2021. Lecture Notes in Computer Science(
• 作者: Su, ST.;Subramanian, V.G.;Schoenebeck, G.
• 通讯作者: Schoenebeck, G.

Learning-Based Optimal Admission Control in a Single-Server Queuing System

单服务器排队系统中基于学习的最优准入控制

• 发表时间: 2023-12
• 期刊: Stochastic systems
• 作者: Asaf Cohen;Vijay Subramanian;Yili Zhang
• 通讯作者: Yili Zhang

A Strong Duality Result for Cooperative Decentralized Constrained POMDPs

合作分散约束 POMDP 的强对偶结果

• DOI: 10.1109/cdc49753.2023.10383989
• 发表时间: 2023-12-13
• 期刊: 2023 62nd IEEE Conference on Decision and Control (CDC)
• 作者: Nouman Khan;Vijay G. Subramanian
• 通讯作者: Vijay G. Subramanian

Bayesian Learning of Optimal Policies in Markov Decision Processes with Countably Infinite State-Space

可数无限状态空间马尔可夫决策过程中最优策略的贝叶斯学习

• 发表时间: 2023-12
• 期刊: Advances in Neural Information Processing Systems 36 (NeurIPS 2023
• 作者: Saghar Adler;Vijay Subramanian
• 通讯作者: Vijay Subramanian