CRII: CNS: NeTS: Adaptive Cache Dimensioning in Cloud CDNs: Foundations and Practice

目前的互联网基础设施提供广泛的服务，如音乐和电影传送、信息传递、视频会议和软件下载。通过互联网传输的与这些服务相对应的数据被称为“内容”。随着越来越多的用户希望通过互联网提供这样的服务，称为内容提供商的公司正在致力于开发系统，以确保以高质量的体验即最小的延迟获得所需的服务。这是通过“服务放置”的过程实现的，即将流行服务的副本放置在称为缓存的小型服务器上，并将流行服务的副本放置在终端用户附近，再将额外的副本放置在互联网更深的较大服务器上。如果用户请求的内容在缓存中可用，则会立即交付。否则，必须将请求转发到更远的服务器，从而增加延迟。然而，配置这些小型和大型服务器的成本都很高。在云计算模式下，云提供商提供可供租用的服务器资源，并允许动态调整缓存大小，这称为缓存大小调整。这意味着内容提供商的成本可以显著降低。该项目开发了用于处理不同类型服务的高速缓存大小的方法。一个重大的挑战在于，服务的受欢迎程度随着时间的推移而变化，因此学习、规模确定和服务放置必须不断发生。解决方案的方法是通过机器学习，该项目有助于随着时间的推移从一系列样本中学习的基本原理，名为在线学习。该项目还包括编写关于联网、分布式系统和机器学习的教材。该项目考虑云内容分发网络(CDN)中的缓存大小问题，其目标是决定在网络中的每个位置放置多少存储空间。该项目解决了在云CDN中提供自适应缓存大小区分服务的理论基础、实用在线算法和低复杂性实现所必需的关键问题。这需要结合几个数学工具来分析在线算法，从而导致系统开发使算法成为现实。本项目开发了一个基于社会福利最大化的框架，用于在云CDN中提供自适应缓存规模区分服务。该项目被组织成三个相互依存的推力。第一个重点是基于生存时间(TTL)近似类比的分析，通过动态调整计时器值来分离不同内容的行为，从而最大化社会福利。第二个重点是基于在线优化的分析，通过利用在线学习来设计新的在线反应算法，这些算法能够感知非平稳的受欢迎程度和流量变化。第三个重点是公有云基础设施的实施和评估。该项目的直接影响是帮助设计下一代云CDN，从而提高企业工作效率和用户满意度。具体的少数群体融入活动、侧重于缓存和机器学习的教育计划以及以高中生夏令营形式开展的外联活动，都加强了这一影响。同时，该项目开发了与机器学习领域相关的基本理论，特别是与在线学习相关的理论。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。

The current Internet infrastructure provides a wide range of services such as music and movies delivery, messaging, video-conferencing and software download. The data transmitted across the Internet corresponding to these services is dubbed as “content”. As an increasing number of users desire such services over the Internet, firms called content providers are engaged in developing systems that ensure that the demanded services are available at high quality of experience, i.e., with minimal delay. This is achieved by the process of “service placement”, which places replicas of popular services near end users at small servers called caches, coupled with additional copies at larger servers deeper in the Internet. If the content requested by a user is available at a cache, it is promptly delivered. Otherwise, the request must be forwarded to servers that are further, hence increasing delay. However, provisioning these small and large servers is expensive. Under the cloud computing paradigm, cloud providers make server resources available for rent and allow dynamic sizing of caches, referred to as cache dimensioning. This implies that costs for the content provider may be significantly reduced. This project develops methodologies on cache dimensioning for handling different types of services. A significant challenge lies in the fact that popularity of services changes with time, and hence learning, dimensioning and service placement must happen continually. The solution approach is via machine learning, and the project contributes to the fundamentals of learning from a sequence of samples over time, entitled online learning. The project also includes the development of educational materials on networking, distributed systems and machine learning. This project considers the cache dimensioning problem in cloud content distribution networks (CDNs), where the objective is to decide how much storage to place at each location in the network. This project addresses key issues essential to developing theoretical foundations, practical online algorithms and low-complexity implementation for providing adaptive cache dimensioning differentiated services in cloud CDNs. This requires the conjunction of several mathematical tools to analyze online algorithms, leading to systems development to make the algorithms a reality. This project develops a social welfare maximization-based framework for providing adaptive cache dimensioning differentiated service in cloud CDNs. The project is organized into three interdependent thrusts. The first thrust focuses on a Time-to-Live (TTL) approximation analogy-based analysis to decouple the behaviors of different contents by means of dynamically adapting the timer values to maximize the social welfare. The second thrust focuses on online optimization-based analysis by leveraging online learning to design new online reactive algorithms that are aware of non-stationary popularity and traffic variations. The third thrust focuses on implementation and evaluation on public cloud infrastructure. An immediate impact of this project is to help design next-generation cloud CDNs leading to greater enterprise productivity and user satisfaction. The impact is enhanced by specific minority inclusion activities, an education plan focusing on caching and machine learning, as well as outreach in the form of summer camps for high school students. At the same time the project develops fundamental theories that pertain to the area of machine learning, specifically to online learning.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.