Overview Queueing Disciplines TCP Congestion Control Congestion Avoidance Mechanisms Quality of Service

Transcription

1 Cogestio Cotrol Overview Queueig Disciplies TCP Cogestio Cotrol Cogestio Avoidace Mechaisms Quality of Service Sprig 2018 CS 438 Staff - Uiversity of Illiois 1

2 Today s Topic: Vacatios Sa Fracisco Moterey Plaig a vacatio? Try a trip to sceic Moterey, Califoria! Moterey is a mere 3 hops from UIUC. What happeed? Sorry, FLIGHT OVERBOOKED. Please fly agai! Chicago UIUC Sprig 2018 CS 438 Staff - Uiversity of Illiois 2

3 Cogestio Cotrol readig: Peterso ad Davie, Ch. 6 Basics: Problem, termiology, approaches, metrics Solutios Router-based: queueig disciplies Host-based: TCP cogestio cotrol Cogestio avoidace DECbit RED gateways Quality of service Sprig 2018 CS 438 Staff - Uiversity of Illiois 3

4 Cogestio Cotrol Basics Problem Demad for etwork resources ca grow beyod the resources available Wat to provide fair amout to each user Examples Badwidth betwee Chicago ad Sa Fracisco Badwidth i a etwork lik Buffers i a queue Sprig 2018 CS 438 Staff - Uiversity of Illiois 4

5 Cogestio Collapse Defiitio Icrease i etwork load results i decrease of useful work doe May possible causes Spurious retrasmissios of packets still i flight Classical cogestio collapse Solutio: better timers ad TCP cogestio cotrol Udelivered packets Packets cosume resources ad are dropped elsewhere i etwork Solutio: cogestio cotrol for ALL traffic Sprig 2018 CS 438 Staff - Uiversity of Illiois 5

6 Dealig with Cogestio Rage of solutios Cogestio cotrol Cure cogestio whe it happes Cogestio avoidace Predict whe cogestio might occur ad avoid causig it Resource allocatio Prevet cogestio from occurrig Model of etwork Packet-switched iteretwork (or etwork) Coectioless flows (logical chaels/coectios) betwee hosts Sprig 2018 CS 438 Staff - Uiversity of Illiois 6

7 Cogestio Cotrol Goal Effective ad fair allocatio of resources amog a collectio of competig users Learig whe to say o ad to whom Resources Badwidth Buffers Problem Cotetio at routers causes packet loss Sprig 2018 CS 438 Staff - Uiversity of Illiois 7

8 Flow Cotrol vs. Cogestio Cotrol Flow cotrol Prevetig oe seder from overruig the capacity of a slow receiver Cogestio cotrol Prevetig a set of seders from overloadig the etwork! Sprig 2018 CS 438 Staff - Uiversity of Illiois 8

9 Cogestio is Natural Because Iteret traffic is bursty! If two packets arrive at the same time The ode ca oly trasmit oe ad either buffers or drops the other Sprig 2018 CS 438 Staff - Uiversity of Illiois 9

10 Cogestio is Natural Because Iteret traffic is bursty! If two packets arrive at the same time The ode ca oly trasmit oe ad either buffers or drops the other If may packets arrive i a short period of time The ode caot keep up with the arrivig traffic Causes delays, ad the buffer may evetually overflow Sprig 2018 CS 438 Staff - Uiversity of Illiois 10

11 Load ad Delay Typical behavior of queueig systems with bursty arrivals: Ideal: low delays ad high utilizatio Reality: must balace the two Maximizig power is a example Load Power = Delay Average Packet delay Power Load optimal load Load Sprig 2018 CS 438 Staff - Uiversity of Illiois 11

12 Basic Desig Choices Prevetio or Cure? Pre-allocate resources to avoid cogestio Sed data ad cotrol cogestio if ad whe it occurs Possible implemetatio poits Hosts at the edge of the etwork Trasport protocol Routers iside the etwork Queueig disciplies Uderlyig service model Best effort vs. quality of service (QoS) Sprig 2018 CS 438 Staff - Uiversity of Illiois 12

13 Flows Sequece of packets set betwee source/destiatio pair Similar to ed-to-ed abstractio of chael, but see at routers Maitai per-flow soft state at the routers Source 1 Source 2 Source 3 Router Router Router Destiatio 1 Destiatio 2 Sprig 2018 CS 438 Staff - Uiversity of Illiois 13

14 Router State Soft state: Hard state: Iformatio about flows Helps cotrol cogestio state used to support routig Not ecessary for correct routig Source 1 Router Source 2 Router Destiatio 1 Source 3 Router Destiatio 2 Sprig 2018 CS 438 Staff - Uiversity of Illiois 14

15 Cogestio Cotrol Router role Cotrols forwardig ad droppig policies Ca sed feedback to source Host role Moitors etwork coditios Adjusts accordigly Routig vs. cogestio Effective adaptive routig schemes ca sometimes help cogestio But ot always Sprig 2018 CS 438 Staff - Uiversity of Illiois 15

16 Cogestio Cotrol Taxoomy cogestio cotrol feedback-based implicit feedback, implemeted by hosts, cotrolled by widow abstractio, a.k.a. best effort reservatio-based, implemeted by routers, cotrolled by rate, a.k.a. quality of service/qos explicit feedback, implemeted by routers, but ot per flow why? Sprig 2018 CS 438 Staff - Uiversity of Illiois 16

17 Router-Cetric vs. Host- Cetric Flow Cotrol Router-cetric Host-cetric Each router takes resposibility for decidig Whe packets are forwarded Which packets are to be dropped Hosts observe etwork coditios ad adjust their behavior accordigly Iformig hosts of sedig limitatios Sprig 2018 CS 438 Staff - Uiversity of Illiois 17

18 Reservatio-Based vs. Feedback-Based Flow Cotrol Reservatio-based Ed host asks etwork for capacity at flow establishmet time Routers alog flow s route allocate appropriate resources If resources are ot available, flow is rejected Implies the use of router-cetric mechaisms Feedback-based Ed host begis sedig without askig for capacity Ed host adjusts sedig rate accordig to feedback Explicit vs. implicit feedback mechaisms May use router-cetric (explicit) or host-cetric (implicit) mechaisms Sprig 2018 CS 438 Staff - Uiversity of Illiois 18

19 Per-flow Cogestio Feedback Questio Why is explicit per-flow cogestio feedback from routers rarely used i practice? Sprig 2018 CS 438 Staff - Uiversity of Illiois 19

20 Per-flow Cogestio Feedback Problem Too may sources to track Millios of flows may fa i to oe router Ca t sed feedback to all of them Adds complexity to router Need to track more state Certaily ca t track state for all sources Wastes badwidth: etwork already cogested, ot the time to geerate more traffic Ca t force the sources (hosts) to use feedback Sprig 2018 CS 438 Staff - Uiversity of Illiois 20

21 Widow-based vs. Rate-based Flow Cotrol Remember Give a RTT ad widow size W, log term throughput rate is Rate = mi(lik speed, W/RTT) Sice rate ca be cotrolled by the widow size, is there really ay differece betwee cotrollig the widow size ad cotrollig the rate? Rate W Sprig 2018 CS 438 Staff - Uiversity of Illiois 21

22 Rate Cotrol Questio Why cosider rate cotrol? Problems Buffer space (widow size) is a itrisic physical quatity Ca provide rate cotrol with widow cotrol Oly eed estimate of RTT 0 1 RTT 2 RTT Aswer Wat rate cotrol whe graularity of averagig must be smaller tha RTT widow-cotrolled trasmissios time rate-cotrolled trasmissios Sprig 2018 CS 438 Staff - Uiversity of Illiois 22

23 Criticisms of Resource Allocatio Example Divide 10 Gbps badwidth out of UIUC Case 1: reserve whatever you wat Users lie of thought Result O average, I do t eed much badwidth, but whe my persoal Web crawler goes to work, I eed at least 100 Mbps, so I ll reserve that much. 100 users cosume all badwidth, all others get 0 Sprig 2018 CS 438 Staff - Uiversity of Illiois 23

24 Criticisms of Resource Allocatio Example Divide 10 Gbps badwidth out of UIUC Case 2: fair/equitable reservatios 35,000 studets + 5,000 faculty ad staff Each user gets 250 kbps, almost 5x a modem! Sprig 2018 CS 438 Staff - Uiversity of Illiois 24

25 Resource Allocatio Back to the air travel aalogy Daily Chicago to Sa Fracisco flight, 198 seats Case 1: reserve whatever you wat 198 of us get seats. I m Gold...are you? Case 2: fair/equitable reservatios 2,000,000 possible customers seats per customer per flight Disclaimer: the passeger assumes all risks ad damages related to usuccessful reassembly i Chicago Sprig 2018 CS 438 Staff - Uiversity of Illiois 25

26 Widow Size For o-radom etwork with bottleeck capacity C: Source C Destiatio Rate = Throughput Delay C W RTT/2 W Power = throughput/delay RTT*C W Sprig 2018 CS 438 Staff - Uiversity of Illiois 27

27 Fairess Goals Allocate resources fairly Isolate ill-behaved users Still achieve statistical multiplexig Oe flow ca fill etire pipe if o coteders Work coservig à scheduler ever idles lik if it has a packet At what graularity? Flows, coectios, domais? Sprig 2018 CS 438 Staff - Uiversity of Illiois 28

28 What s Fair? Flow A Flow B Flow C Flow D Which is more fair: Globally Fair: Fa = Capacity/4, Fb = Fc = Fd = 3Capacity/4 or Locally Fair: Fa = Fb = Fc = Fd = Capacity/2 This is the socalled max-mi fair rate allocatio. The miimum rate is maximized. Sprig 2018 CS 438 Staff - Uiversity of Illiois 29

29 Max-Mi Fairess Flow A Flow B Flow C Flow D 1. No user receives more tha requested badwidth 2. No other scheme with 1 has higher mi badwidth 3. 2 remais true recursively o removig miimal user µ I = MIN(µ fair, r i ) Sprig 2018 CS 438 Staff - Uiversity of Illiois 30

30 Max-Mi Fairess: Example Capacity(C) = 10 3 Flows: r1 = 8, r2 = 6, r3 = 2 C/3 = 3.33 Ca service all of r3 Remove r3 from the accoutig: C = C r3 = 8; N = 2 C/2 = 4 Ca t service all of r1 or r2 So hold them to the remaiig fair share: f = Sprig 2018 CS 438 Staff - Uiversity of Illiois 31

31 Queueig Disciplies Goal Decide how packets are buffered while waitig to be trasmitted Provide protectio from ill-behaved flows Each router MUST implemet some queuig disciplie regardless of what the resource allocatio mechaism is Impact Directly impacts buffer space usage Idirectly impacts flow cotrol Sprig 2018 CS 438 Staff - Uiversity of Illiois 32

32 Queueig Disciplies Allocate badwidth Which packets get trasmitted Allocate buffer space Which packets get discarded Affect packet latecy Whe packets get trasmitted Sprig 2018 CS 438 Staff - Uiversity of Illiois 33

33 Schedulig Policies FIFO (First I First Out) a.k.a. FCFS (First Come First Serve) Service I order of arrival to the queue Maagemet Packets that arrive to a full buffer are discarded Aother optio: discard policy determies which packet to discard (ew arrival or somethig already queued) Sprig 2018 CS 438 Staff - Uiversity of Illiois 34

34 Schedulig Policies FIFO (First I First Out) Problem 1: sed more packets, get more service Selfish seders tryig to grab as much as they ca Malicious seders tryig to dey service to others Problem 2: ot all packets should be equal Sprig 2018 CS 438 Staff - Uiversity of Illiois 35

35 Schedulig Policies FIFO Does ot discrimiate betwee traffic sources Cogestio cotrol left to the sources Tail drop droppig policy Fairess for latecy Miimizes per-packet delay Badwidth ot cosidered (ot good for cogestio) Sprig 2018 CS 438 Staff - Uiversity of Illiois 36

36 Schedulig Policies Priority Queuig Classes have differet priorities May deped o explicit markig or other header ifo e.g., IP source or destiatio, TCP Port umbers, etc. Service Trasmit packet from highest priority class with a o-empty queue Sprig 2018 CS 438 Staff - Uiversity of Illiois 37

37 Schedulig Policies Priority Queuig Isolatio for the high-priority traffic Almost like it has a dedicated lik Except for the (small) delay for packet trasmissio High-priority packet arrives durig trasmissio of low-priority Router completes sedig the low-priority traffic first Sprig 2018 CS 438 Staff - Uiversity of Illiois 38

38 Schedulig Policies Priority Queueig Versios Preemptive Postpoe low-priority processig if high-priority packet arrives No-preemptive Limitatio Ay packet that starts gettig processed fiishes before movig o May starve lower priority flows Sprig 2018 CS 438 Staff - Uiversity of Illiois 39

39 Schedulig Policies Roud Robi Each flow gets its ow queue Circulate through queues, process oe packet (if queue o-empty), the move to ext queue Sprig 2018 CS 438 Staff - Uiversity of Illiois 40

40 Schedulig Policies Fair Queueig (FQ) Explicitly segregates traffic based o flows Esures o flow captures more tha its share of the capacity Fairess for badwidth Delay ot cosidered Flow 1 Flow 2 Flow 3 Flow 4 Roud- Robi service Each flow is guarateed ¼ of capacity Sprig 2018 CS 438 Staff - Uiversity of Illiois 41

41 Fair Queueig with Variable Packet Legth How should we implemet FQ if packets are ot all the same legth? Bit-by-bit roud-robi Not feasible to implemet, must use packet schedulig Solutio: approximate 4 8? Sprig 2018 CS 438 Staff - Uiversity of Illiois 42

42 Fair Queueig with Variable Packet Legth Idea Let S i = amout of service flow i has received so far Always serve a flow with miimum value of S i Ca also use miimum (S i + ext packet legth) Upo servig a packet of legth P from flow i, update: S i = S i + P Never leave the lik idle if there is a packet to sed Work coservig A source will gets its fair share of the badwidth Uused badwidth will be evely divided betwee other sources Sprig 2018 CS 438 Staff - Uiversity of Illiois 43

43 Fair Queueig with Variable Packet Legth Problem A flow resumes sedig packets after beig quiet for a log time Effect Such a flow could be cosidered to have saved up credit Ca lock out all other flows util credits are level agai Solutio Eforce use it or lose it policy Compute S mi = mi(s i such that queue i is ot empty) If queue j is empty, set S j = S mi Sprig 2018 CS 438 Staff - Uiversity of Illiois 44

44 Fair Queueig with Variable Packet Legth Problem A flow resumes sedig packets after beig quite for a log time Effect Note: The text book computes Such a flow could be cosidered F = MAX(F i-1 to, have A i ) = P saved i up credit Ad the for multiple flows Ca lock out all other flows Calculate util credits F i for are each level packet agai Solutio that arrives o each flow Eforce use it or lose it policy Treat all F i as timestamps Next packet to trasmit is oe Compute S mi = mi(s i such that queue i is ot empty) If queue j is empty, set S with j = S lowest timestamp mi Sprig 2018 CS 438 Staff - Uiversity of Illiois 45

45 Extesio: Weighted Fair Queueig Exted fair queueig Notio of importace for each flow Suppose flow i has weight w i Example: w i could be the fractio of total service that flow i is targeted for Need oly chage basic update to S i = S i + P/w i Sprig 2018 CS 438 Staff - Uiversity of Illiois 46

46 Fair Queuig Tradeoffs FQ ca cotrol cogestio by moitorig flows No-adaptive flows ca still be a problem why? Complex state Must keep queue per flow Hard i routers with may flows (e.g., backboe routers) Flow aggregatio is a possibility (e.g. do fairess per domai) Complex computatio Classificatio ito flows may be hard Must keep queues sorted by fiish times Chages wheever the flow cout chages Sprig 2018 CS 438 Staff - Uiversity of Illiois 47

47 Fair Queueig Questio What makes up a flow for fair queueig i the Iteret? Cosideratios Too may resources to have separate queues/variables for host-to-host flows Scale dow umber of flows Typically just based o iputs e.g., share outgoig STS-12 betwee icomig ISP s Sprig 2018 CS 438 Staff - Uiversity of Illiois 48

48 TCP Cogestio Cotrol Sprig 2018 CS 438 Staff - Uiversity of Illiois 49

49 Host Solutios Host has very little iformatio Assumes best-effort etwork Acts idepedetly of other hosts Host actios Reduce trasmissio rate below cogestio threshold Cotiuously moitor etwork for sigs of cogestio Sprig 2018 CS 438 Staff - Uiversity of Illiois 50

50 Detectig Cogestio How ca a TCP seder determie that the etwork is uder stress? Network could tell it (ICMP Source Quech) Risky, because durig times of overload the sigal itself could be dropped (ad add to cogestio)! Packet delays go up (kee of load-delay curve) Tricky: oisy sigal (delay ofte varies cosiderably) Packet loss Fail-safe sigal that TCP already has to detect Complicatio: o-cogestive loss (checksum errors) Sprig 2018 CS 438 Staff - Uiversity of Illiois 51

51 TCP Cogestio Cotrol Idea Assumes best-effort etwork FIFO or FQ Each source determies etwork capacity for itself Implicit feedback ACKs pace trasmissio (self-clockig) Challege Determiig iitial available capacity Adjustig to chages i capacity i a timely maer Sprig 2018 CS 438 Staff - Uiversity of Illiois 52

52 TCP Cogestio Cotrol Basic idea Add otio of cogestio widow Effective widow is smaller of Advertised widow (flow cotrol) Cogestio widow (cogestio cotrol) Chages i cogestio widow size Slow icreases to absorb ew badwidth Quick decreases to elimiate cogestio Sprig 2018 CS 438 Staff - Uiversity of Illiois 53

53 TCP Cogestio Cotrol Specific strategy Self-clockig Sed data oly whe outstadig data ACK d Equivalet to sed widow limitatio metioed seder receiver Sprig 2018 CS 438 Staff - Uiversity of Illiois 54

54 TCP Cogestio Cotrol Specific strategy Self-clockig Growth Sed data oly whe outstadig data ACK d Equivalet to sed widow limitatio metioed Add oe maximum segmet size (MSS) per cogestio widow of data ACK d It s really doe this way, at least i Liux: see tcp_cog_avoid i tcp_iput.c. Actually, every ack for ew data is treated as a MSS ACK d Kow as additive icrease Sprig 2018 CS 438 Staff - Uiversity of Illiois 55

55 TCP Cogestio Cotrol Specific strategy (cotiued) Decrease Cut widow i half whe timeout occurs I practice, set widow = widow /2 Kow as multiplicative decrease Additive icrease, multiplicative decrease (AIMD) Sprig 2018 CS 438 Staff - Uiversity of Illiois 56

56 Additive Icrease/ Multiplicative Decrease Objective Adjust to chages i available capacity Basic idea Cosequeces of over-sized widow much worse tha havig a uder-sized widow Over-sized widow: packets dropped ad retrasmitted Uder-sized widow: somewhat lower throughput Sprig 2018 CS 438 Staff - Uiversity of Illiois 57

57 Additive Icrease/ Multiplicative Decrease Tools React to observace of cogestio Probe chael to detect more resources Observatio O otice of cogestio Decreasig too slowly will ot be reactive eough O probe of etwork Icreasig too quickly will overshoot limits Sprig 2018 CS 438 Staff - Uiversity of Illiois 58

58 Additive Icrease/ Multiplicative Decrease New TCP state variable CogestioWidow Similar to AdvertisedWidow for flow cotrol Limits how much data source ca have i trasit MaxWi = MIN(CogestioWidow, AdvertisedWidow) EffWi = MaxWi - (LastByteSet - LastByteAcked) TCP ca sed o faster the the slowest compoet, etwork or destiatio Idea Icrease CogestioWidow whe cogestio goes dow Decrease CogestioWidow whe cogestio goes up Sprig 2018 CS 438 Staff - Uiversity of Illiois 59

59 Additive Icrease/ Multiplicative Decrease Questio How does the source determie whether or ot the etwork is cogested? Aswer Timeout sigals packet loss Packet loss is rarely due to trasmissio error (o wired lies) Lost packet implies cogestio! Sprig 2018 CS 438 Staff - Uiversity of Illiois 60

60 Additive Icrease/ Multiplicative Decrease Algorithm Icremet CogestioWidow by oe packet per RTT Liear icrease Divide CogestioWidow by two wheever a timeout occurs I practice Multiplicative decrease icremet a little for each ACK Ic = MSS * MSS/CogestioWidow CogestioWidow += Ic Source Destiatio Sprig 2018 CS 438 Staff - Uiversity of Illiois 61

61 AIMD Sawtooth Trace KB Packet loss is see as sig of cogestio ad results i a multiplicative rate decrease Factor of 2 TCP periodically probes for available badwidth by icreasig its rate Loss halved Time (secods) 10.0 Sprig 2018 CS 438 Staff - Uiversity of Illiois 62

62 Additive Icrease/Decrease Both icrease/ decrease by the same amout User 2 s Allocatio T 0 T 1 Overload Additive icrease improves fairess Uderutilizatio Additive decrease reduces fairess User 1 s Allocatio Sprig 2018 CS 438 Staff - Uiversity of Illiois 64

63 Muliplicative Icrease/Decrease Both icrease/ decrease by the same amout User 2 s Allocatio T 0 T 1 Uderutilizatio Overload Additive icrease improves fairess Additive decrease reduces fairess User 1 s Allocatio Sprig 2018 CS 438 Staff - Uiversity of Illiois 65

64 Why is AIMD Fair? Additive icrease gives slope of 1, as throughout icreases Multiplicative decrease decreases throughput proportioally User 2 s Allocatio R equal badwidth share User 1 s Allocatio loss: decrease widow by factor of 2 cogestio avoidace: additive icrease loss: decrease widow by factor of 2 cogestio avoidace: additive icrease R Sprig 2018 CS 438 Staff - Uiversity of Illiois 66

65 AIMD Sharig Dyamics A D x 1 x 2 No cogestio à rate icreases by oe packet/rtt every RTT Cogestio à decrease rate by factor 2 B E Sprig 2018 CS 438 Staff - Uiversity of Illiois 67

66 AIMD Sharig Dyamics A D x 1 x 2 B E Rates equalize à fair share Sprig 2018 CS 438 Staff - Uiversity of Illiois 68

67 TCP Start Up Behavior KB How should TCP start sedig data? AIMD is good for chaels operatig at capacity AIMD ca take a log time to ramp up to full capacity from scratch It could take a log time to get started! Time (secods) 10.0 Sprig 2018 CS 438 Staff - Uiversity of Illiois 69

68 TCP Start Up Behavior How should TCP start sedig data? AIMD is good for chaels operatig at capacity AIMD ca take a log time to ramp up to full capacity from scratch Use Slow Start to icrease widow rapidly from a cold start Sprig 2018 CS 438 Staff - Uiversity of Illiois 70

69 TCP Start Up Behavior: Slow Start Iitializatio of the cogestio widow Cogestio widow should start small Avoid cogestio due to ew coectios Start at 1 MSS, Iitially, CWND is 1 MSS Iitial sedig rate is MSS/RTT Reset to 1 MSS with each timeout timeouts are coarse-graied, ~1/2 sec Sprig 2018 CS 438 Staff - Uiversity of Illiois 71

70 TCP Start Up Behavior: Slow Start Growth of the cogestio widow Liear growth could be pretty wasteful Might be much less tha the actual badwidth Liear icrease takes a log time to accelerate Start slow but the grow fast Seder starts at a slow rate Icrease the rate expoetially Util the first loss evet Sprig 2018 CS 438 Staff - Uiversity of Illiois 72

71 Slow Start Objective Determie iitial available capacity Idea Begi with CogestioWidow = 1 packet Double CogestioWidow each RTT Icremet by 1 packet for each ACK Cotiue icreasig util loss Source Destiatio Sprig 2018 CS 438 Staff - Uiversity of Illiois 73

72 Slow Start Example 0R 1R 2R 3R 1 oe pkt time Sprig 2018 CS 438 Staff - Uiversity of Illiois 74

73 Aother Slow Start Example CWD size: Src D A D D A A D D D D A A A A Dest Sprig 2018 CS 438 Staff - Uiversity of Illiois 75

74 Slow Start Used Whe first startig coectio Whe coectio times out Why is it called slow-start? Because TCP origially had o cogestio cotrol mechaism The source would just start by sedig a whole widow s worth of data Sprig 2018 CS 438 Staff - Uiversity of Illiois 76

75 TCP Cogestio Cotrol Maitai threshold widow size Threshold value Iitially set to maximum widow size Set to 1/2 of curret widow o timeout Use multiplicative icrease I practice Whe cogestio widow smaller tha threshold Double widow for each widow ACK d Icrease cogestio widow by oe MSS for each ACK of ew data (or N bytes for N bytes) Sprig 2018 CS 438 Staff - Uiversity of Illiois 77

76 Slow Start How log should the expoetial icrease from slow start cotiue? Use CogestioThreshold as target widow size Estimates etwork capacity Whe CogestioWidow reaches CogestioThreshold switch to additive icrease Liear probig Expoetial slow start Sprig 2018 CS 438 Staff - Uiversity of Illiois 78

77 Slow Start Iitial values CogestioThreshold = 8 CogestioWidow = 1 Loss after trasmissio 7 CogestioWidow curretly 12 Set Cogestiothreshold = CogestioWidow/2 Set CogestioWidow = 1 Sprig 2018 CS 438 Staff - Uiversity of Illiois 79

78 Slow Start Example trace of CogestioWidow KB Problem Have to wait for timeout CW flattes out due to loss Liear icrease Slow start util CW = CT Timeout: CT = CT/2 = 11 CW = 1 Ca lose half CogestioWidow of data Sprig 2018 CS 438 Staff - Uiversity of Illiois 80

79 Fast Retrasmit ad Fast Recovery Problem Coarse-grai TCP timeouts lead to idle periods Solutio Fast retrasmit: use duplicate ACKs to trigger retrasmissio Packet 1 Packet 2 Packet 3 Packet 4 Packet 5 Packet 6 Retrasmit packet 3 Seder Receiver Sprig 2018 CS 438 Staff - Uiversity of Illiois 81 ACK 1 ACK 2 ACK 2 ACK 2 ACK 2 ACK 6

80 Fast Retrasmit ad Fast Recovery Sed ACK for each segmet received Whe duplicate ACK s received Resed lost segmet immediately Do ot wait for timeout I practice, retrasmit o 3rd duplicate Fast recovery Whe fast retrasmissio occurs, skip slow start Cogestio widow becomes 1/2 previous Start additive icrease immediately Sprig 2018 CS 438 Staff - Uiversity of Illiois 82

81 Fast Retrasmit ad Fast Recovery KB Results Fast Recovery Bypass slow start phase Icrease immediately to oe half last successful CogestioWidow (ssthresh) Sprig 2018 CS 438 Staff - Uiversity of Illiois 83

82 TCP Cogestio Widow Trace Cogestio Widow timeouts fast retrasmissio slow start period additive icrease Time threshold cogestio widow Sprig 2018 CS 438 Staff - Uiversity of Illiois 84