Breaking the Transience-Equilibrium Nexus: A New Approach to Datacenter Packet Transport
Breaking the Transience-Equilibrium Nexus: A New Approach to Datacenter Packet Transport
复制标题
DOI：
--
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Shiyu Liu;Ahmad Ghalayini;Mohammad Alizadeh;B. Prabhakar;M. Rosenblum;Anirudh Sivaraman
中科院分区：
文献类型：
--
作者：
Shiyu Liu;Ahmad Ghalayini;Mohammad Alizadeh;B. Prabhakar;M. Rosenblum;Anirudh Sivaraman
最近的数据中心传输协议严重依赖来自网络的丰富拥塞信号，阻碍了它们在公共云等环境中的部署。在这篇文章中，我们解释了这一趋势，表明在没有丰富的拥塞信号的情况下，在分组传输的均衡性和瞬变性能之间存在着强烈的权衡。然后，我们提出了一种简单的方法来解决这种紧张，而不会使传输协议复杂化，也不需要来自网络的丰富的拥塞信号。我们的方法将运输分为两个独立的部分，用于平衡和瞬时处理。为了实现均衡处理，我们继续使用现有的拥塞控制协议。对于暂态，我们开发了一种新的底层算法，入站-RAMP，该算法在暂态过载期间在网络边缘拦截和保持任何协议的分组。入口斜坡使用单向延迟的精确测量来检测瞬时过载，这是通过最近开发的时间同步算法在软件中实现的。在Google Cloud平台上，In-Ramp将Cubic流量的99%请求完成时间(RCT)提高了2.8x，BBR提高了5.6x。在裸机云(CloudLab)中，入口斜坡将立方体的RCT提高了4.1倍。在ns-3仿真中，它模拟了更有效的基于NIC的入口斜坡实现，入口斜坡根据工作负载在不同程度上改善了DCQCN、DCTCP和HPCC的RCT。在所有三种环境中，onramp还改进了非广播后台流量的流完成时间。在Facebook的一项评估中，入站显著减少了计算流量的延迟，同时确保存储流量的吞吐量不受影响。
Recent datacenter transport protocols rely heavily on rich congestion signals from the network, impeding their deployment in environments such as the public cloud. In this paper, we explain this trend by showing that, without rich congestion signals, there is a strong tradeoff between a packet transport’s equilibrium and transience performance. We then propose a simple approach to resolve this tension without complicating the transport protocol and without rich congestion signals from the network. Our approach factors the transport into two separate components for equilibrium and transient handling. For equilibrium handling, we continue to use existing congestion control protocols. For transients, we develop a new underlay algorithm, On-Ramp, which intercepts and holds any protocol’s packets at the network edge during transient overload. On-Ramp detects transient overloads using accurate measurements of one-way delay, made possible in software by a recently developed time-synchronization algorithm. On the Google Cloud Platform, On-Ramp improves the 99th percentile request completion time (RCT) of incast traffic of CUBIC by 2.8× and BBR by 5.6×. In a bare-metal cloud (CloudLab), On-Ramp improves the RCT of CUBIC by 4.1×. In ns-3 simulations, which model more efficient NIC-based implementations of On-Ramp, On-Ramp improves RCTs of DCQCN, TIMELY, DCTCP and HPCC to varying degrees depending on the workload. In all three environments, OnRamp also improves the flow completion time of non-incast background traffic. In an evaluation at Facebook, On-Ramp significantly reduces the latency of computing traffic while ensuring the throughput of storage traffic is not affected.