In this work, we provide the design and implementation of a switch-assisted congestion control algorithm for data center networks (DCNs). In particular, we provide a prototype of the switch-driven congestion control algorithm and deploy it in a real data center. The prototype is based on few simple modifications to the switch software. The modifications imposed by the algorithm on the switch are to enable the switch to modify the TCP receive-window field in the packet headers. By doing so, the algorithm can enforce a pre-calculated (or target rate) to limit the sending rate at the sources. Therefore, the algorithm requires no modifications to the TCP source or receiver code which considered out of the DCN operators' control (e.g., in the public cloud where the VM is maintained by the tenant). This paper describes in detail two implementations, one as a Linux kernel module and the second as an added feature to the well-known software switch, Open vSwitch. Then we present evaluation results based on experiments of the deployment of both designs in a small testbed to demonstrate the effectiveness of the proposed technique in achieving high throughput, good fairness, and short flow completion times for delay-sensitive flows.
翻译:在这项工作中,我们为数据中心网络(DCNs)提供开关辅助拥堵控制算法的设计和实施。特别是,我们提供了开关驱动拥堵控制算法的原型,并将其安装在真正的数据中心。原型基于对开关软件的少量简单修改。开关的算法的修改是使开关能够修改包头中的TCP接收风窗字段。通过这样做,算法可以执行预先计算(或目标速率)来限制源的发送率。因此,算法不需要修改从DCN操作员控制中考虑过的TCP源代码或接收器代码(例如,在VM由租户维护的公共云层中)。本文详细描述了两个执行过程,一个是Linux内核模块,第二个是众所周知的软件开关的附加特征,Open vSwitch。然后根据在小型测试台里应用两种设计试验的结果,以显示拟议技术在高通量、公平性、短流完成延迟流中的效率。