基于多级散列的动态流量调度方法  

Dynamic Traffic Scheduling Method Based on Multilevel Hashing

作　　者：徐泽 汪学舜 戴锦友[2] 吴小锋 XU Ze;WANG Xueshun;DAI Jinyou;WU Xiaofeng(Wuhan Research Institute of Posts and Telecommunications,Wuhan,430074,China;Fiberhome Telecommunication Technologies Co.,Ltd.,Wuhan,430074,China)

机构地区：[1]武汉邮电科学研究院,武汉430074 [2]烽火通信科技股份有限公司,武汉430074

出　　处：《网络新媒体技术》2024年第3期64-72,共9页Network New Media Technology

基　　金：科技部重大研发专项多模态网络控制调度系统技术(编号:2022YFB2901200)。

摘　　要：为保证数据中心场景流量转发的整体质量,在数据中心交换机上需要一个能够区分大象流和老鼠流的流量调度方法。目前,在设备上并没有对其区分,而是当作一种流量进行转发,无法保证用户的体验。本文提出一种基于大象流和老鼠流识别的调度实现方案(MHS),将链路提前规划为低时延链路和高吞吐链路,数据中心交换机将流量通过多级不平等散列表的方式进行记录,设置流量的阈值筛选出大象流,借助调度策略重定向到高吞吐链路,减少大象流和老鼠流相互影响。在可编程交换机上进行试验,结果表明该方法可以在数据中心高性能网络达到较好的效果。本方法与等价多路径路由(ECMP)在高速网络中对比,性能提升明显,队列长度比ECMP降低16%,时延降低20%。To ensure the overall quality of traffic forwarding in a data center scenario,a traffic scheduling method that can differentiate between elephant flows and mice flows is required on data center switches.Currently,there is no distinction made on the devices,and all traffic is treated equally,which cannot guarantee a satisfactory user experience.This paper proposes an implementation scheme for elephant flow and mice flow identification called Multi-hash Scheduling(MHS).The links are pre-planned as low-latency links and high-throughput links.The data center switches record the traffic using a multi-level,unequal hash table,and set a threshold to filter out the elephant flows.With the help of the scheduling strategy,these flows are redirected to high-throughput links,reducing the mutual interference between elephant flows and mice flows.Experiments conducted on programmable switches demonstrate that this method achieves better results in high-performance data center networks.When compared to Equal-Cost Multi-Path Routing(ECMP)in high-speed networks,this method shows significant performance improvements with a 16%reduction in queue length and a 20%reduction in latency.

关 键 词：流量调度 流识别 可编程 多级不平等散列 数据中心 

分 类 号：TP393.06[自动化与计算机技术—计算机应用技术]