《Protocol-MPTCP原版完整课件.pptx》由会员分享，可在线阅读，更多相关《Protocol-MPTCP原版完整课件.pptx（33页珍藏版）》请在taowenge.com淘文阁网|工程机械CAD图纸|机械工程制图|CAD装配图下载|SolidWorks_CaTia_CAD_UG_PROE_设计图分享下载上搜索。
1、Design,implementation and evaluation of congestion control for multipath TCPDamon Wischik,Costin Raiciu,Adam Greenhalgh,Mark HandleyNSDI 2011Presented by Ye Tian for Course CS05112OutlineBackgroundDesign problem for MPTCP congestion controlCongestion control at multihomed serverCongestion control in
2、 data centerCongestion control in wireless networkBackgroundMPTCP design objectiveCapable of using multiple network paths for a single connection.Be able to use the available network paths at least as well as regular TCP,but without starving TCP.Be as usable as regular TCP for existing applications(
3、no need to change ends).Protocol layerMPTCP acts as a shim layer between the socket interface and one or more TCP subflows SignalingUse the three-way handshake with TCP options to negotiate its usage.MP_CAPABLE in SYN,SYN+ACK,ACKConnection and subflowTo use another interface,MPTCP uses a three-way h
4、andshake to establish one subflow over this interface.MPTCP needs to be able to link each subflow to an existing MPTCP connection.For this,MPTCP assigns a locally unique token to each connection.The MP_JOIN option of the SYN segment contains the token of the associated MPTCP connection.MPTCP in WiFi
5、/4GThe make-before-break handover can be used if the smartphone can predict that one interface will disappear soonMultipath TCP since iOS 7Primary TCP connection over WiFiBackup TCP connection over cellular dataFailoverIf WiFi becomes unavailable iOS 7 will use the cellular data connectionFor destin
6、ations controlled by AppleE.g.,SiriMPTCP in DatacenterServer have multiple NICsImprove availability as well as throughputLinux kernel implementationhttps:/multipath-tcp.orgOutlineBackgroundDesign problem for MPTCP congestion controlCongestion control at multihomed serverCongestion control in data ce
7、nterCongestion control in wireless networkRegular TCPw can be viewed as packet on the fly(sent out but not yet acked)RTT is the time for a packet to be sent and acked.So,in an RTT,if w packets are sent and acked,window size in increased by w*1/w=1.Increase the window by one packet per RTT.Called AIM
8、D:Additive Increase/multiplicative DecreaseMPTCP congestion controlProposed MPTCP congestion controlEach subflow rR maintains its own congestion window wrFairnessOn each path,subflow gets window size proportional to a2If regular TCP on a same path gets window size as 1.Fairness with TCPSubflows are
9、independentWhen a subflow is congested,the other subflow doesnt become more aggressive.2-subflow MPTCP can not have higher throughput than a single TCP flow Problem:choosing efficient pathEWTCP would not make very efficient use of the networkThree links each has 12MbpsLink bandwidth evenly split by
10、three identical subflows(although they from different connections)If split traffic evenly in two paths,each subflow has 4Mbps,total 8MbpsShould shift all traffic to the least-congested path,get 12Mbps.Choosing efficient pathp is packet loss probability on path rSolve the equation:Choosing efficient
11、pathwtotal doesnt depend on number of pathsSubflow on path with higher pr will see more decreases.The equilibrium window of a path with higher loss probability pr pmin has a size wr=0.The algorithm always prefers least congested path.EWTCP vs.Coupled EWTCP,subflows in a same link have equal throughp
12、utFlow A:5+6;flow B:6+5;flow C:5+3Coupled,a flow uses a path only if that path has the lowest loss rate.If multiple paths used,it must be that each link have a same loss prob.,so each flow has a same throughput()Flow A=flow B=flow C=10Adapting to load changes“Only use the least congested paths”needs
13、 to be balanced by an opposing consideration,“Always keep sufficient traffic on other paths,as a probe,so that you can quickly discover when they improve.”Window sizeKeep moderate traffic on each pathCompensating RTT mismatchIncrease at most 1/wr,so will not be more aggressive than regular TCPa cont
14、rols aggressivenessFinal MPTCP algorithm is the generalization of the algorithmOutlineBackgroundDesign problem for MPTCP congestion controlCongestion control at multihomed serverCongestion control in data centerCongestion control in wireless networkStatic load balancing Five bottleneck links in a to
15、rusReduce capacity on link CCOUPLED is very good at balancing congestion,EWTCP is bad,and MPTCP is in between.Jains fairness index is 0.99 for the flow rates with COUPLED,0.986 for MPTCP and 0.92 for EWTCP.Dynamic load balancing Two links as in Fig.9,both of capacity 100Mb/s and buffer 50 packetsOn
16、the top link there is an additional bursty CBR flow which sends at 100Mb/s for a random duration of mean 10ms,then is quiet for a random duration of mean 100ms.The multipath flow ought to use only the bottom link when the CBR flow is present,and it ought to quickly take up both links when the CBR fl
17、ow is absent.COUPLED is worstOutlineBackgroundDesign problem for MPTCP congestion controlCongestion control at multihomed serverCongestion control in data centerCongestion control in wireless networkTopologyTwo topologiesFatTree:128 single-interface host and 80 eight-port switchBCube:125 three-inter
18、face host and 25 five-port switchCompareTCP+ECMP8 random paths for multipathTraffic patternTP1:each host opens a flow to a single destination chosen uniformly at randomTP2:each host opens 12 flows to 12 destinationsFatTree:destinations are chosen at randomBCube:hosts neighbors in the three levelsTP3
19、:30%of the hosts open one flow to a single destination chosen uniformly at random.FatTree FatTreeMPTCP does a better job of allocating throughput fairly than EWTCP,BCubeboth multipath algorithms allow a host to use all three of its interfaces whereas single-path TCP can use only oneOutlineBackground
20、Design problem for MPTCP congestion controlCongestion control at multihomed serverCongestion control in data centerCongestion control in wireless networkSingle flow use a laptop equipped with a 3G USB interface and a 802.11 network adapterWiFi provides much higher throughput and shorter RTT,higher l
21、oss rate15 tests5 with single-path TCP on WiFi,5 with single-path TCP on 3G,5 with MPTCPThe average throughputs(with standard deviations)were 14.4(0.2),2.1(0.2)and 17.3(0.7)Mb/s respectively.Competing flowCompeting flowonly MPTCP gives the multipath flow a fair total throughputClose to WiFi in this
22、caseMobile experiment0-9 min.:3G path has less congestion,but Wi-Fi path hasmore bandwidthOn 3G path,subflow competes with TCP flow9-11 min.:go downstair for coffee,out of Wi-Fi coverage,but has better 3G coverageReviewWhat is the benefit of MPTCP,how it works?Concerns of MPTCP congestion control.Fa
23、irness with TCP flow.Adapt to load change.Discover and make use of less congested path.How EWTCP,Coupled,and MPTCP behavesEWTCP:subflows are independent,one flow congested,other subflows will not be more aggressive.Coupled:only use the least congested path,can not discover other paths when they are good.MTTCP