Congestion Control in TCP

Transcription

1 Congestion Control in TCP Antonio Carzaniga Faculty of Informatics Università della Svizzera italiana November 16, 2018

2 Outline Intro to congestion control Input rate vs. output throughput Congestion window Congestion avoidance Slow start Fast recovery

3 Arouterbehavesalotlikeakitchensink Understanding Congestion

4 Understanding Congestion Arouterbehavesalotlikeakitchensink maxrate=r

5 Understanding Congestion Arouterbehavesalotlikeakitchensink λ 1 =R/2 maxrate=r throughput = R/2

6 Understanding Congestion Arouterbehavesalotlikeakitchensink λ 1 =R/2 λ 2 =R/2 maxrate=r throughput = R

7 Understanding Congestion Arouterbehavesalotlikeakitchensink λ 3 =R/2 λ 1 =R/2 λ 2 =R/2 maxrate=r throughput = R

8 Understanding Congestion Arouterbehavesalotlikeakitchensink λ 3 =R/2 λ 1 =R/2 λ 2 =R/2 maxrate=r throughput = R

9 Understanding Congestion Arouterbehavesalotlikeakitchensink λ 3 =R/2 λ 1 =R/2 λ 2 =R/2 maxrate=r throughput = R

10 Understanding Congestion Arouterbehavesalotlikeakitchensink λ 3 =R/2 λ 1 =R/2 λ 2 =R/2 maxrate=r throughput = R

11 Understanding Congestion Arouterbehavesalotlikeakitchensink λ 3 =R/2 λ 1 =R/2 λ 2 =R/2 maxrate=r throughput = R

12 Queuing Delay

13 Queuing Delay Totallatencyisthesumoflinklatency,processingtime,andthetimethata packet spends in the input queue L =d TX +d CPU +d q whered q = q /R

14 Queuing Delay Totallatencyisthesumoflinklatency,processingtime,andthetimethata packet spends in the input queue L =d TX +d CPU +d q whered q = q /R Ideal case: constant input data rate λ in <R Inthiscasethed q =0,because q =0 (idealinputflow)

15 Queuing Delay Totallatencyisthesumoflinklatency,processingtime,andthetimethata packet spends in the input queue L =d TX +d CPU +d q whered q = q /R Ideal case: constant input data rate λ in <R Inthiscasethed q =0,because q =0 (idealinputflow) Extreme case: constant input data rate λ in >R Inthiscase q = (λ in R)tandtherefore d q = λ in R t R

16 Queuing Delay

17 Queuing Delay Steady-state queuing delay d q = { 0 λin <R λ in R t λ R in >R

18 Queuing Delay Steady-state queuing delay d q = { 0 λin <R λ in R t λ R in >R d q λ in R ideal input flow λ in constant

19 Queuing Delay Steady-state queuing delay d q = { 0 λin <R λ in R t λ R in >R d q d q λ in R ideal input flow λ in constant λ in R realistic input flow λ in variable

20 Queuing Delay

21 Queuing Delay Conclusion:astheinputrateλ in approachesthemaximumthroughputr, packets will experience very long delays

22 Queuing Delay Conclusion:astheinputrateλ in approachesthemaximumthroughputr, packets will experience very long delays More realistic assumptions and models finite queue length(buffers) in routers effects of retransmission overhead full queues along multi-hops paths

23 Queuing Delay Conclusion:astheinputrateλ in approachesthemaximumthroughputr, packets will experience very long delays More realistic assumptions and models finite queue length(buffers) in routers effects of retransmission overhead full queues along multi-hops paths R λ out λ in

24 Queuing Delay Conclusion:astheinputrateλ in approachesthemaximumthroughputr, packets will experience very long delays More realistic assumptions and models finite queue length(buffers) in routers effects of retransmission overhead full queues along multi-hops paths R congestion λ out λ in

25 What to Do? Whattodowhenthenetworkiscongested? λ 3 =R/2 λ 1 =R/2 λ 2 =R/2 maxrate=r throughput = R

26 What to Do? Whattodowhenthenetworkiscongested? BACKOFF! λ 3 =R/2 λ 1 =R/2 λ 2 =R/2 maxrate=r throughput = R

27 What to Do? Whattodowhenthenetworkiscongested? BACKOFF! λ 3 =R/4 λ 1 =R/4 λ 2 =R/4 maxrate=r throughput = R

28 What to Do? Whattodowhenthenetworkiscongested? BACKOFF! λ 3 =R/4 λ 1 =R/4 λ 2 =R/4 maxrate=r throughput = R

29 What to Do? Whattodowhenthenetworkiscongested? BACKOFF! λ 3 =R/4 λ 1 =R/4 λ 2 =R/4 maxrate=r throughput = R

30 What to Do? Whattodowhenthenetworkiscongested? BACKOFF! λ 3 =R/4 λ 1 =R/4 λ 2 =R/4 maxrate=r throughput = R

31 Congestion Control (in TCP) Approach: Thesenderlimitsitsoutputrateaccordingtothestateofthenetwork Thesenderoutputratebecomes(partof)theinputrateforthenetwork(λ in )

32 Congestion Control (in TCP) Approach: Thesenderlimitsitsoutputrateaccordingtothestateofthenetwork Thesenderoutputratebecomes(partof)theinputrateforthenetwork(λ in ) Ingredients: 1.Howdoesthesendermeasurethestateofthenetwork? weneedeyestoseethetrafficahead

33 Congestion Control (in TCP) Approach: Thesenderlimitsitsoutputrateaccordingtothestateofthenetwork Thesenderoutputratebecomes(partof)theinputrateforthenetwork(λ in ) Ingredients: 1.Howdoesthesendermeasurethestateofthenetwork? weneedeyestoseethetrafficahead 2.howdoesthesendersetitsoutputrate? weneedacceleratorandbrakestospeeduporslowdown

34 Congestion Control (in TCP) Approach: Thesenderlimitsitsoutputrateaccordingtothestateofthenetwork Thesenderoutputratebecomes(partof)theinputrateforthenetwork(λ in ) Ingredients: 1.Howdoesthesendermeasurethestateofthenetwork? weneedeyestoseethetrafficahead 2.howdoesthesendersetitsoutputrate? weneedacceleratorandbrakestospeeduporslowdown 3. how should the sender control its output rate? weneedabrainandweneedtoknowhowtodrive!

35 Detecting Congestion (Eyes)

36 Detecting Congestion (Eyes) If all traffic is correctly acknowledged, with fresh acknowledgments, then the sender assumes(quite correctly) that there is no congestion

37 Detecting Congestion (Eyes) If all traffic is correctly acknowledged, with fresh acknowledgments, then the sender assumes(quite correctly) that there is no congestion Congestion means that some queues overflow in one or more routers between the sender and the receiver thevisibleeffectisthatsomesegmentsaredropped

38 Detecting Congestion (Eyes) If all traffic is correctly acknowledged, with fresh acknowledgments, then the sender assumes(quite correctly) that there is no congestion Congestion means that some queues overflow in one or more routers between the sender and the receiver thevisibleeffectisthatsomesegmentsaredropped Therefore the sender assumes that the network is congested when it(the sender) detects a segment loss duplicate acknowledgements(i.e., NACK) timeout(i.e.,noacksatall)

39 The sender maintains a congestion window W Congestion Window (Accellerator/Brakes)

40 Congestion Window (Accellerator/Brakes) The sender maintains a congestion window W The congestion window limits the amount of bytes that the sender pushes into the network before blocking waiting for acknowledgments

41 Congestion Window (Accellerator/Brakes) The sender maintains a congestion window W The congestion window limits the amount of bytes that the sender pushes into the network before blocking waiting for acknowledgments where LastByteSent LastByteAcked W W = min(congestionwindow, ReceiverWindow)

42 Congestion Window (Accellerator/Brakes) The sender maintains a congestion window W The congestion window limits the amount of bytes that the sender pushes into the network before blocking waiting for acknowledgments where LastByteSent LastByteAcked W W = min(congestionwindow, ReceiverWindow) The resulting maximum output rate is roughly λ = W 2L

43 Congestion Control (Brain, Algorithm)

44 Additive-increase and multiplicative-decrease Congestion Control (Brain, Algorithm)

45 Congestion Control (Brain, Algorithm) Additive-increase and multiplicative-decrease Slow start

46 Congestion Control (Brain, Algorithm) Additive-increase and multiplicative-decrease Slow start Reaction to timeout events

47 Additive-Increase/Multiplicative-Decrease

48 Additive-Increase/Multiplicative-Decrease HowWisreduced:ateverylossevent,TCPhalvesthecongestionwindow

49 Additive-Increase/Multiplicative-Decrease HowWisreduced:ateverylossevent,TCPhalvesthecongestionwindow e.g.,supposethewindowsizewiscurrently20kb,andalossisdetected TCPreducesWto10Kb

50 Additive-Increase/Multiplicative-Decrease HowWisreduced:ateverylossevent,TCPhalvesthecongestionwindow e.g.,supposethewindowsizewiscurrently20kb,andalossisdetected TCPreducesWto10Kb How W is increased: at every(good) acknowledgment, TCP increments W by 1MSS/W,soastoincreaseWbyMSSeveryround-triptime2L.Thisprocessis called congestion avoidance

51 Additive-Increase/Multiplicative-Decrease HowWisreduced:ateverylossevent,TCPhalvesthecongestionwindow e.g.,supposethewindowsizewiscurrently20kb,andalossisdetected TCPreducesWto10Kb How W is increased: at every(good) acknowledgment, TCP increments W by 1MSS/W,soastoincreaseWbyMSSeveryround-triptime2L.Thisprocessis called congestion avoidance e.g.,supposew =14600andMSS =1460,thenthesenderincreasesWto16060 after 10 acknowledgments acknowledgments

52 Additive-Increase/Multiplicative-Decrease WindowsizeWovertime W Time

53 WhatistheinitialvalueofW? Slow Start

54 Slow Start WhatistheinitialvalueofW? TheinitialvalueofWisMSS,thatis1segment,whichisquitelowformodern networks

55 Slow Start WhatistheinitialvalueofW? TheinitialvalueofWisMSS,thatis1segment,whichisquitelowformodern networks To get quickly to a good throughput level, TCP increases its sending rate exponentially for its first growth phase, up to a slow-start threshold(ssthresh)

56 Slow Start WhatistheinitialvalueofW? TheinitialvalueofWisMSS,thatis1segment,whichisquitelowformodern networks To get quickly to a good throughput level, TCP increases its sending rate exponentially for its first growth phase, up to a slow-start threshold(ssthresh) After the threshold, TCP proceeds with its linear push

57 Slow Start WhatistheinitialvalueofW? TheinitialvalueofWisMSS,thatis1segment,whichisquitelowformodern networks To get quickly to a good throughput level, TCP increases its sending rate exponentially for its first growth phase, up to a slow-start threshold(ssthresh) After the threshold, TCP proceeds with its linear push Thisprocessiscalled slowstart becauseofthesmallinitialvalueofw

58 Asweknow,threeduplicateACKsareinterpretedasaNACK Timeouts vs. NACKs

59 Timeouts vs. NACKs Asweknow,threeduplicateACKsareinterpretedasaNACK BothtimeoutsandNACKssignalaloss,buttheysaydifferentthingsaboutthe status of the network

60 Timeouts vs. NACKs Asweknow,threeduplicateACKsareinterpretedasaNACK BothtimeoutsandNACKssignalaloss,buttheysaydifferentthingsaboutthe status of the network A timeout indicates congestion

61 Timeouts vs. NACKs Asweknow,threeduplicateACKsareinterpretedasaNACK BothtimeoutsandNACKssignalaloss,buttheysaydifferentthingsaboutthe status of the network A timeout indicates congestion Three(duplicate) ACKs suggest that the network is still able to deliver segments along that path

62 Timeouts vs. NACKs Asweknow,threeduplicateACKsareinterpretedasaNACK BothtimeoutsandNACKssignalaloss,buttheysaydifferentthingsaboutthe status of the network A timeout indicates congestion Three(duplicate) ACKs suggest that the network is still able to deliver segments along that path So,TCPreactsdifferentlytoatimeoutandtoatripleduplicateACKs

63 AssumingthecurrentwindowsizeisW =W Timeouts vs. NACKs

64 Timeouts vs. NACKs AssumingthecurrentwindowsizeisW =W Timeout gobacktow =MSS setssthresh =W/2 runslowstartuptow =ssthresh then proceed with congestion avoidance

65 Timeouts vs. NACKs AssumingthecurrentwindowsizeisW =W Timeout gobacktow =MSS setssthresh =W/2 runslowstartuptow =ssthresh then proceed with congestion avoidance NACK(i.e., triple duplicate-ack) setssthresh =W/2 cutwinhalf:w =W/2 run congestion avoidance, ramping up W linearly Thisiscalledfastrecovery

66 Sender Behavior W Time

67 Sender Behavior W MSS Time

68 Sender Behavior W NACK MSS Time

69 Sender Behavior W NACK MSS Time

70 Sender Behavior timeout W NACK MSS Time

71 Sender Behavior timeout W NACK MSS Time

72 Sender Behavior timeout W NACK MSS Time

73 Sender Behavior timeout W NACK MSS Time

74 Sender Behavior NACK timeout W NACK MSS Time

75 Sender Behavior NACK W NACK timeout NACK MSS Time

76 Sender Behavior NACK W NACK timeout NACK MSS Time

77 Sender Behavior NACK W NACK timeout NACK MSS Time

78 Sender Behavior SS=slow start CA=congestion avoidance SS CA SS CA CA CA NACK W NACK timeout NACK MSS Time