云南龙网

  1. 首页 > 云知道 > >

网络流量控制 网络流量监控工具( 四 )

云知道


https://en.wikipedia.org/wiki/TCP_congestion_control关于不同拥塞算法对比 。
拥塞算法网络反馈需要修改点应用场景公平性准则(New) RenoLoss——DelayVegasDelaySenderLess lossProportionalHigh SpeedLossSenderHigh bandwidthBICLossSenderHigh bandwidthCUBICLossSenderHigh bandwidthC2TCP[10][11]Loss/DelaySenderUltra-low latency and high bandwidthNATCP[12]Multi-bit signalSenderNear Optimal PerformanceElastic-TCPLoss/DelaySenderHigh bandwidth/short & long-distanceAgile-TCPLossSenderHigh bandwidth/short-distanceH-TCPLossSenderHigh bandwidthFASTDelaySenderHigh bandwidthProportionalCompound TCPLoss/DelaySenderHigh bandwidthProportionalWestwoodLoss/DelaySenderLJerseyLoss/DelaySenderLBBR[13]DelaySenderBLVC, BufferbloatCLAMPMulti-bit signalReceiver, RouterVMax-minTFRCLossSender, ReceiverNo RetransmissionMinimum delayXCPMulti-bit signalSender, Receiver, RouterBLFCMax-minVCP2-bit signalSender, Receiver, RouterBLFProportionalMaxNetMulti-bit signalSender, Receiver, RouterBLFSCMax-minJetMaxMulti-bit signalSender, Receiver, RouterHigh bandwidthMax-minREDLossRouterReduced delayECNSingle-bit signalSender, Receiver, RouterReduced loss这里还要提一下TCP流的公平性问题，有的拥塞算法可能会存在带宽抢占，基于延时相关的算法就很容易出现，因为是基于RTT来抢占带宽，RTT越小占用则更多，比如有两条TCP流，TCP流中延迟更小的会抢占更多的带宽，甚至完全占有所有带宽，其他流都无法正常工作 。比如reno和bic都存在抢占问题，cubic是通过3次曲线来探测理想带宽，和RTT无关联，公平性做得较好 。

网络流量控制 网络流量监控工具


如上图使用reno算法，使用iperf打流到两个不同服务端，stream2延时低直接抢占了大部分带宽

网络流量控制 网络流量监控工具


如上图使用cubic算法，同样使用iperf打流到两个不同服务端，steam2延时低抢占带宽则没有那么严重
Linux内核中拥塞算法Linux 2.6.x内核版本具体可参考
https://linuxgazette.net/135/pfeiffer.html进行拥塞算法选择，最新的linux 5.x内核多支持了几种，内核代码Kconfig中也有简要描述，顺便提一下mptcp内核也包含了多路径场景的几种拥塞算法 。
Linux 4.x内核中支持的tcp拥塞算法，默认使用cubic：
config DEFAULT_TCP_CONGstringdefault "bic" if DEFAULT_BICdefault "cubic" if DEFAULT_CUBICdefault "htcp" if DEFAULT_HTCPdefault "hybla" if DEFAULT_HYBLAdefault "vegas" if DEFAULT_VEGASdefault "westwood" if DEFAULT_WESTWOODdefault "veno" if DEFAULT_VENOdefault "reno" if DEFAULT_RENOdefault "dctcp" if DEFAULT_DCTCPdefault "cdg" if DEFAULT_CDGdefault "bbr" if DEFAULT_BBRdefault "cubic"Linux mptcp v0.95内核支持的mptcp拥塞算法：
config TCP_CONG_LIAtristate "MPTCP Linked Increase"depends on MPTCPdefault n---help---MultiPath TCP Linked Increase Congestion ControlTo enable it, just put 'lia' in tcp_congestion_controlconfig TCP_CONG_OLIAtristate "MPTCP Opportunistic Linked Increase"depends on MPTCPdefault n---help---MultiPath TCP Opportunistic Linked Increase Congestion ControlTo enable it, just put 'olia' in tcp_congestion_controlconfig TCP_CONG_WVEGAStristate "MPTCP WVEGAS CONGESTION CONTROL"depends on MPTCPdefault n---help---wVegas congestion control for MPTCPTo enable it, just put 'wvegas' in tcp_congestion_controlconfig TCP_CONG_BALIAtristate "MPTCP BALIA CONGESTION CONTROL"depends on MPTCPdefault n---help---Multipath TCP Balanced Linked Adaptation Congestion ControlTo enable it, just put 'balia' in tcp_congestion_controlconfig TCP_CONG_MCTCPDESYNCtristate "DESYNCHRONIZED MCTCP CONGESTION CONTROL (EXPERIMENTAL)"depends on MPTCPdefault n---help---Desynchronized MultiChannel TCP Congestion Control. This is experimentalcode that only supports single path and must have set mptcp_ndiffportslarger than one.To enable it, just put 'mctcpdesync' in tcp_congestion_controlFor further details see:http://ieeexplore.ieee.org/abstract/document/6911722/https://doi.org/10.1016/j.comcom.2015.07.010

推荐阅读


上一篇：1500左右手机推荐 1500左右配置最高手机

下一篇：十二星座专属手机 十二星座专属手机品牌