LEGATO: LEarninG how to enhAnce neTwork prOtocols

对Internet通信性能的不断追求导致网络协议及其机制变得越来越复杂。即使是看似简单的机制，如HTTP/2中的资源优先级划分，结果也具有非常复杂的参数相互依赖关系。今天的协议性能不仅取决于其在客户端和服务器端的配置，还取决于各自网络堆栈中其他协议层的配置，使得协议裁剪成为一个纵向和横向的跨层优化问题。传统的跨层优化工作通过人工分析影响和依赖关系，找出优化策略。然而，日益增长的复杂性让参数空间不断增长，以至于很难找到合适的策略。因此，参数空间通常被减小，然而，这限制了整体视图，并且因此可能降低有效的性能收益。例如，对于资源优先级排序，与其复杂性相比，这些策略只实现了平庸的加速。因此，该机制基本上处于闲置状态。这种影响并不是HTTP/2独有的，而是在许多类似的情况下都会发生。机器学习领域的技术可以处理较大的参数空间。它们已经用于协议优化，使用这些技术的方法提供了如何使用它们的见解，强调简单地应用它们不是正确的方式，因为验证需要领域知识。然而，这些方法以不可知的方式查看层，而忽略了跨层组件。因此，如何将这些技术应用于跨层优化领域是未知的。我们建议创建一种涉及机器学习的跨层协议优化方法，其中我们回答了如何使用它来捕获互联网协议环境的复杂相互依赖关系。我们的目标不仅限于优化网络协议。相反，应该深入了解哪些方法是合适的，如何限制它们，它们做出什么决定，以及如何将获得的策略反馈到协议开发中。

The constant pursuit for performance in Internet communications leads to network protocols and their mechanisms getting more and more involved and hence complex. Even seemingly simple mechanisms such as resource prioritization in HTTP/2 turn out to have very complex parameter interdependencies. Today’s protocol performance does not only depend on its configuration on client- and server-side, but also on the configuration of the other protocol layers in the respective network stack, rendering protocol tailoring into a vertical and horizontal cross-layer optimization problem. Traditional cross-layer optimization work analyses the influences and dependencies manually and finds optimization strategies. However, the increasing complexities let the parameter space grow, such that it is hard to find appropriate strategies. Thus, the parameter space is typically reduced, which, however, limits the holistic view and can hence reduce the effective performance gains. E.g., for resource prioritization, the strategies achieve only mediocre speedups in comparison to their complexities. Hence, the mechanism is mainly left unused. This effect is not exclusive to HTTP/2, but occurs in many similar cases.Techniques from the area of machine learning can handle big parameter spaces. They are already used for protocol optimization and the approaches using these techniques give insights on how to use them, emphasizing that simply applying them is not the right way, as domain knowledge is needed for validation. However, the approaches view the layers agnostically ignoring the cross-layer component. Hence, it is unknown how to apply these techniques in the realm of cross-layer optimization. We propose to create a methodology for cross-layer protocol optimization involving machine learning, where we answer how to use it for capturing the complex interdependencies of the Internet protocol landscape. Our goals are not limited to only optimizing network protocols. Instead, insights should be given in regard to which approaches are suitable, how to constrain them, what decisions they make and how to feed the gained strategies back into protocol development.