Serverless computing systems are becoming very popular. Large corporations such as Netflix, Airbnb, and Coca-Cola use such systems for running their websites and IT systems. The advantages of such systems include superior support for auto-scaling, load balancing, and fast distributed processing. These are multi-QoS systems where different classes of applications have different latency and jitter (variation in the latency) requirements: we consider a mix of latency-sensitive (LS) and latency-desirable (LD) applications. Ensuring proper schedulability and QoS enforcement of LS applications is non-trivial. We need to minimize the jitter without increasing the response latency of LS applications, and we also need to keep the degradation of the response latency of LD applications in check. This is the first paper in this domain that achieves a trade-off between the jitter suffered by LS applications and the response latency of LD applications. We minimize the former with a bound on the latter using a reinforcement learning (RL) based scheme. To design such an RL scheme, we performed detailed characterization studies to find the input variables of interest, defined novel state representations, and proposed a bespoke reward function that allows us to achieve this trade-off. For an aggressive use case comprising five popular LS and LD applications each, we show a reduction in response time variance and mean latency of 50.31% and 27.4%, respectively, for LS applications. The mean degradation in the execution latency of LD applications was limited to 19.88%.
翻译:无服务器的计算系统正在变得非常流行。大型公司,如Netflix、Airbnb和Coca-Cola,在运行其网站和IT系统时使用这种系统。这些系统的优点包括:对自动缩放、负负平衡和快速分布处理的高度支持。这些是多QOS系统,不同类别的应用程序的延迟和急度(悬浮)要求不同:我们考虑的是延缓敏感(LS)和延缓(LD)应用程序的混合。确保LS应用程序的适当弹性和QoS执行是非三角性的。我们需要在不提高LS应用程序的应对时间长度的情况下,最大限度地减少音量,同时还要控制LD应用程序的反应时间。这是这一领域中第一个在LS应用程序受到的急度和LD应用程序的响应惯性之间实现交换。我们把LS应用程序与后一种基于强化学习(RL)的有限差异应用捆绑在一起。我们为RS应用程序设计了50个新的递增时间,我们为RL功能设计了一个新的递减(L),我们为RL的递减了对RL功能的每个变量进行了详细的递减。</s>