NeTS: Medium: Managing Datacenter Traffic Bursts, Fast and Slow

作为互联网的现代“工厂”，互联网中心为我们日益数字化的世界提供了燃料。它们承载着当今大多数应用程序，从电子邮件到电子医疗保健解决方案。这些应用程序具有极低的延迟要求。由于网络拥塞，满足这些要求具有挑战性。传统的拥塞管理方法已不适用于当今日益复杂和高度动态的网络。特别是，它们无法对网络流量中普遍存在的短暂峰值做出足够快的反应。本研究的目标是解决这一挑战，开发新的模式，快速拥塞识别和管理数据中心网络。如果成功，该项目将提高基于云的应用程序的速度和效率。该项目假定理想的拥塞控制取决于拥塞事件的时间尺度。例如，实时和本地拥塞管理技术，如数据包偏转，可以有效地管理微秒级的拥塞。然而，如果在大规模拥塞期间部署，相同的技术可能会导致不稳定、吞吐量崩溃，并且具有讽刺意味的是，会增加延迟。在当今动态、大规模和复杂的网络中，基于拥塞的规模进行识别和反应是一项挑战。为了解决这些挑战，本项目进行了两条调查路线：（a）拥挤事件在交通中心是否规模化以及发生频率如何？拥堵事件的结构特征和成因是什么？以及（B）我们能否设计多尺度拥塞管理系统并使各种技术能够互操作？该项目将建立一个整体的测量框架，系统地自动分析从部署在主机的各个层和网络中不同拓扑位置的多个Vantage位置的拥塞的缩放。该项目还将开发多尺度拥塞管理技术，并引入新的抽象概念，使这些技术能够通过根据拥塞规模将拥塞处理任务划分为不同技术来实现互操作。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

As the modern “factories” for the Internet, datacenters fuel our increasingly digital world. They host the majority of today’s applications ranging from email to e-healthcare solutions. These applications have exceedingly low latency requirements. Meeting these requirements is challenging due to network congestion. Traditional approaches for managing congestion are not well-suited for today’s increasingly complex and highly dynamic networks. In particular, they fail to react fast enough to pervasive short-lived spikes in network traffic. The goal of this research is to address this challenge by developing new paradigms for fast congestion identification and management in datacenter networks. If successful, the project will improve the speed and efficiency of cloud-based applications. It can also enable future latency-sensitive applications.The project posits that ideal congestion control depends on the time scale of the congestion event. For example, real-time and local congestion management techniques such as packet deflection can effectively manage microsecond-scale congestion. However, the same techniques can cause instability, throughput collapse and, ironically, increased latency if deployed during widescale congestion. Identifying and reacting based on the scale of congestion is challenging in today’s dynamic, large-scale, and complex networks. To resolve these challenges, this project pursues two lines of inquiry: (a) do congestion events scale in datacenters and how frequently? What are the structural properties and origins of congestion events? and (b) can we design multi-scale congestion management systems and enable various techniques to interoperate? This project will build a holistic measurement framework that systematically and automatically analyzes the scaling of congestion from multiple vantage points deployed in various layers of hosts and different topological locations in the network. This project will also develop multi-scale congestion management techniques and introduce novel abstractions that allow these techniques to interoperate by dividing the task of congestion handling among different techniques based on the scale of congestion.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.