A Baker's dozen of TCP

download A Baker's dozen of TCP

If you can't read please download the document

description

Linux Plumbers Conference discussion of TCP and Bufferbloat

Transcript of A Baker's dozen of TCP

2. AgendaWhat is TCP congestion control?What are some of the alternatives?Problems in congestion controlSpeculation 3. TCP congestion controlis the result of years ofresearch. 4. AcronymnsRTT Round Trip TimeBDP Bandwidth Delay ProductAIMD Additive Increase Multiplicative Decrease 5. Goals of Congestion control?UtilizationFairnessAdaptability What about latency? 6. Evolution of TCP Congestion Control 7. TCP Congestion Control TypesLoss basedDelay basedRENOVegasBIC, CUBICHTCP Hybrid Illinois Veno Yeah 8. Traditional TCP (Reno)http://www.deneholme.net/tom/ 9. Whats wrong with RENO?Rate Recovery Time @ 200ms RTT1Mbps1.7s10Mbps 10.7s100Mbps2min1Gbps28min10Gbp4 hrs 43minhttp://www.deneholme.net/tom/ 10. http://www.slac.stanford.edu/comp/net/wan-mon/thru-vs-loss.html 11. CUBIC 12. Title:cubic2.epsCreator:gnuplot 4.0 patchlevel 0CreationDate:Sun Jul 29 00:11:16 2007King of the hill problem 13. HTCPUse Epoch as estimate of BDP 14. Performance with small QueuesTitle:util.epsCreator:gnuplot 4.0 patchlevel 0CreationDate:Tue Jul 31 13:59:03 2007 15. Vegas 16. Delay based IssuesSample sizeDelayed ACKShort lived flowsRTT Data qualityPacket trainsTimer resolutionNot aggressive Works great in well behaved network 17. General IssuesMiddleboxes: broken, cheatingMost connections too shortNo persistenceSender chooses, receiver loses 18. CC Rap SheetReno Too slow to recoverBIC Unfair at short RTT or low speedCUBIC Slow to convergeHTCP Slower with small queuesVegas Too wimpy and sensitiveHybrid Reverts to delay based 19. IdeasResearchHybrid congestion controlRefine delay based portionDetect mega queuesBest CC discoveryTestingMore real world testingGooglish random testing