数据中心以太网拥塞控制

One of current trends is to construct lossless, low latency and high throughput unified switch fabric by utilizing congestion control in data center Ethernet. This research improves the fairness and the robustness of congestion control mechanism: quantify system robustness with the H∞ norm, overcome the problem that the performance of the standard link layer congestion control algorithm QCN is parameters and environment sensitive, and satisfy the new requirement of simultaneously ensuring fairness among NICs of virtual machines and among different flows via the cooperation of scheduling and congestion control. Design scheme which scales congestion control to ultra-high bandwidth such as 40/100Gbps: decouple and improve the rapidity and the smoothness of convergence process under the direction of variable structure control theory, such that the congestion control system reaches the stable state before the finishing of short flows, and there is no packet loss due to queue oscillation; as well as reduce the impacts of delay. Develop proper method to deal with the burst traffic: ensure no packet loss by storing burst traffic in the switch buffers, and avoid the increment of queuing delay by fast response to burst traffic, referring to the core ideas of both backpressure and hop-by-hop bandwidth reservation. Construct the NS2-based simulation platform and the NetFPGA-based experimental platform to test and complete the designed mechanisms and algorithms.

通过数据中心以太网拥塞控制，构建低延时、无丢失和高吞吐量的一体化交换网络是当前的发展趋势。本项目改进以太网拥塞控制机制的鲁棒性和公平性：使用H∞范数刻画系统的鲁棒程度，克服标准链路层拥塞控制算法QCN的性能对参数和环境变化敏感的问题, 通过调度和拥塞控制协作来满足数据中心网络中同时确保虚拟机网卡出口间公平和每流公平的新需求。设计适应40/100Gbps超高带宽的拥塞控制策略：在变结构控制理论的指导下，解耦系统收敛过程的快速性和平滑性，同时加快收敛速度和降低缓存队列的震荡幅度，既确保系统在短流完成前进入稳定状态，又避免丢包；削弱延时的影响。开发恰当应对突发流量的方案：融合逐跳预约和后压拥塞状态的思想，既在交换机中恰当地缓存突发数据从而确保无丢失，又加快系统响应速度来避免排队延时的增大。搭建基于NS2的仿真平台和基于NetFPGA的实验平台，测试和完善设计的机制和算法。

通过数据中心以太网拥塞控制，构建低延时、无丢失和高吞吐量的一体化交换网络是当前的发展趋势。本项目设计了快速简单的拥塞控制算法FSQCN，简化了标准链路层拥塞控制算法QCN，克服了其复杂机制带来的对参数和环境变化敏感的问题。设计了基于延时的滑模拥塞控制机制DSM, 能够在更大的延时带宽积环境中保持鲁棒稳定，适应于40/100Gbps超高带宽的网络环境。综述了应对突发流量、构建无丢失网络的方案，建模分析了键值存储系统中的延时分布。针对调研和分析发现的纯短流情形下系统瓶颈为硬盘I/O而不是网络的情形，设计了键值数据库系统副本选择算法L2、Tars和TAP，将流量更加均匀分配给不同的服务器，降低了键值存取操作的尾延时。