摘 要：延迟是数据中心网络及其承载应用的一个关键性能度量指标，越来越受到学术界和工业界的关注。文章提出一种综合的协调机制，把负载均衡和拥塞控制统一起来，在多条等价路径上均匀分布流量，利用ECN 并结合流的优先级自适应地对拥塞做出反应。实验结果显示，该机制与目前采用的TCP 和基于哈希的多路径传输机制相比，对延迟敏感的流，应用吞吐率提高了30%，流完成时间的99th 分位数降低了90%；对背景流量，流的平均完成时间降低50% ～ 80%。
关键词：拥塞控制；负载均衡；多路径传输；数据中心网络；显式拥塞通告；时限
中图分类号：TP393 文献标识码：A 文章编号：2096-4706（2020）23-0044-07
Study on a Kind of Network Flow Latency Framework of Minimizing Data Center
ZHAO Zhengwei，REN Zhenqin，ZHAO Xuge
（School of Information Technology，Luoyang Normal University，Luoyang 471934，China）
Abstract：Latency is a key performance measurement index of data center network and it’s load bearing application，which is getting more and more attention from academic circles and industrial circles. This pager proposes an integrated coordinated mechanism to unite load balancing and congestion control，and evenly distribute flow in several equivalent paths，uses ECN and combining priority level of flow to adaptively respond to congestion. Experimental results show，comparing to existing multi-path transmission mechanism based on TCP and Hash，the mechanism achieves an improvement of 30% in application throughput rate and an reduction of 90% in 99th percentile of flow completion time for latency-sensitive flow，and achieve a reduction of 50% ～ 80% in average completion time of flow for background flow.
Keywords：congestion control；load balance；multi-path transmission；data center network；explicit congestion notification；time limit
基金项目：河南省高等学校重点科研项目（18A520042）
参考文献：
[1] HOFF T. Latency is Everywhere and it Costs You Sales-How to Crush it [EB/OL].（2009-07-25）.http://highscalability.com/blog/2009/7/25/latency-is-everywhere-and-it-costs-you-sales-how-tocrush-it.html.
[2] KOHAVIR R，LONGOTHAM R. Online experiments：Lessons learned [J] Computer ，2007，40（9）：103-105.
[3] BRUTLAG J. Speed matters for Google web search [EB/OL].[2020-10-15].https://services.google.com/fh/files/blogs/google_delayexp.pdf.
[4] GAO P X，NARAYAN A，KARANDIKAR S. Networkrequirements for resource disaggregation [C]//Proceedings of the 12thUSENIX conference on Operating Systems Design and Implementation.Berkeley：USENIX Association，2016：249-264.
[5] SHAN Y Z，ZHANG Y Y，CHEN Y L，et al. LegoOS：a disseminated，distributed OS for hardware resource disaggregation [C]//13th USENIX Symposium on Operating Systems Design andImplementation.Berkeley：USENIX Association，2018：69-87.
[6] KUMAR G，DUKKIPATI N，JANG K，et al. Swift：Delayis Simple and Effective for Congestion Control in the Datacenter [C]//Proceedings of the Annual conference of the ACM Special InterestGroup onData Communication on the applications，technologies，architectures，and protocols for computer communication.New York：Association for Computing Machinery，2020：514-528.
[7] VASUDEVAN V，PHANISHAYEE A，SHAH H，et al.Safe and Effective Fine-grained TCP Retransmissions for DatacenterCommunication [J].ACM SIGCOMM Computer Communication Review，2011,39（4）：303-314.
[8] CHEN Y P，GRIFFITH R，LIU J D. Understanding TCPincast throughput collapse in datacenter networks [C]//Proceedings ofthe 1st ACM SIGCOMM 2009 Workshop on Research on EnterpriseNetworking.Barcelona：Association for Computing Machinery，2009：73-82.
[9] ALIZADEH M，GREENBERG A G，MALTZ D A，et al. Data center TCP （DCTCP）[J].ACM SIGCOMM Computer Communication Review，2010，40（4）：63-74.
[10] WILSON C，BALLANI H，KARAGIANNIS T，et al.Better Never than Late：Meeting Deadlines in Datacenter Networks [J].ACM SIGCOMM Computer Communication Review，2012，41（4）：50-61.
[11] OUSTERHOUT J K，AGRAWAL P，ERICKSON D，et al.The case for RAMClouds：Scalable high-performance storage entirelyin DRAM [J].ACM SIGOPS Operating Systems Review，2009，43（4）：92-105.
[12] ZATS D，DAS T，MOHAN P，et al. DeTail：reducing theflow completion time tail in datacenter networks [J].ACM SIGCOMMComputer Communication Review，2012，42（4）：139-150.
[13] GUO C X，LU G H，LI D，et al. BCube：A HighPerformance，Server-centric Network Architecture for Modular DataCenters [J].ACM SIGCOMM Computer Communication Review，2009，39（4）：63-74.
[14] GUO C X，WU H T，TAN K，et al. DCell：A scalable andfault-tolerant network structure for data centers [J].ACM SIGCOMMComputer Communication Review，2008，38（4）：75.
[15] GREENBERG A G，HAMILTON J R，JAIN N，et al.VL2：A Scalable and Flexible Data Center Network [J].Communicationsof the ACM，2009，54（3）：95-104.
[16] MYSORE R N，PAMBORIS A，FARRINGTON N，etal. PortLand：A Scalable Fault-Tolerant Layer 2 Data Center NetworkFabric [J].ACM SIGCOMM Computer Communication Review，2009，39（4）：39-50.
[17] Cisco.Cisco Data Center Infrastructure 2.5 Design Guide [EB/OL].[2020-10-15].http://www.cisco.com/univercd/cc/td/doc/solution/dcidg21.pdf.
[18] AL-FARES M，RADHAKRISHNAN S，RAGHAVAN B，et al. Hedera：dynamic flow scheduling for data center networks [C]//Proceedings of the 7th USENIX Symposium on Networked SystemsDesign and Implementation.San Jose：USENIX Association，2010：19.
[19] RAICIU C，BARRÉ S，PLUNTKE C，et al. Improving Datacenter Performance and Robustness with Multipath TCP [J].ACM SIGCOMM Computer Communication Review，2011，41（4）：266-277.
[20] BENSON T，ANAND A，AKELLA A，et al. MicroTE：Fine Grained Traffic Engineering for Data Centers [C]//Conferenceon emerging Networking Experiments and Technologies.New York：Association for Computing Machinery，2011：1-12.
[21] CURTIS A R，MOGUL J，TOURRILHES J. DevoFlow：Scaling Flow Management for High-Performance Networks [J].ACMSIGCOMM Computer Communication Review，2011，41（4）：254-265.
[22] GREENBERG A G，LAHIRI P，MALTZ D A，et al.Towards a next generation data center architecture [C]//Proceedings of the ACM SIGCOMM 2008 Workshop on Programmable Routers for Extensible Services of Tomorrow.New York：Association for Computing Machinery，2008：57-62.
[23] SINHA S，KANDULA S，KATABI D. Harnessing TCPs Burstiness using Flowlet Switching [C]//Hot Topics in Networks HotNets- Ⅲ .San Diego：2004.
[24] Cisco. Per-packet load balancing [EB/OL].[2020-10-15].https://www.cisco.com/c/en/us/td/docs/ios-xml/ios/ipswitch_cef/configuration/15-s/isw-cef-15-s-book/isw-cef-load-balancing.html#GUIDC725A4B8-9263-4D2C-95FB-F31D14E477C4.
[25] GEOFFRAY P，HOEFLER T. Adaptive Routing Strategies for Modern High Performance Networks [C]//2008 16th IEEE Symposium on High Performance Interconnects.Stanford：IEEE，2008：165-172.
[26] HONG C Y，CAESAR M，GODFREY P B. Finishing Flows Quickly with Preemptive Scheduling [J].ACM SIGCOMM Computer Communication Review，2012，42（4）：127-138.
[27] DIXIT A，PRAKASH P，HU Y C，et al. On the impact of packet spraying in data center networks [C]//2013 Proceedings IEEE INFOCOM.Turin：IEEE，2013：2130-2138.
[28] DUKKIPATI N. Rate control protocol （rcp）：congestioncontrol to make flows complete quickly [D].Stanford：StanfordUniversity，2007.
[29] KOHAVI R，LONGBOTHAM R，SOMMERFIELD D，etal. Controlled experiments on the Web：survey and practical guide [J].Data Mining and Knowledge Discovery，2009，18：140-181.
[30] APPENZELLER G，KESLASSY I，MCKEOWN N.Sizing Router Buffers [J].ACM SIGCOMM Computer CommunicationReview，2004，34（4）：281-292.
作者简介：
赵正伟（1982—），男，汉族，河南渑池人，讲师，博士，主要研究方向：数据中心网络；
任祯琴（1983—），女，汉族，河南焦作人，讲师，博士，主要研究方向：非线性系统；
赵旭鸽（1992—），女，汉族，河南汝州人，助教，硕士，主要研究方向：大数据。