Chapter 3 Transport Layer

Transcription

1 Chapte 3 Tanspot Laye Reti di Elaboatoi Coso di Lauea in Infomatica Univesità degli Studi di Roma La Sapienza Canale A-L Pof.ssa Chiaa Petioli Pate di queste slide sono state pese dal mateiale associato al libo Compute Netwoking: A Top Down Appoach, 5th edition. All mateial copyight J.F Kuose and K.W. Ross, All Rights Reseved Thanks also to Antonio Capone, Politecnico di Milano, Giuseppe Bianchi and Fancesco LoPesti, Un. di Roma To Vegata Tanspot Laye 3-1

2 TCP: contollo di congestione Il TCP ha dei meccanismi di contollo della congestione m il flusso dei dati in ingesso in ete è anche egolato dalla situazione di taffico in ete m se il taffico in ete pota a situazioni di congestione il TCP iduce velocemente il taffico in ingesso m in ete non vi è nessun meccanismo pe notificae esplicitamente le situazioni di congestione m il TCP ceca di scopie i poblemi di congestione sulla base degli eventi di pedita dei pacchetti Tanspot Laye 3-2

3 TCP Congestion Contol end-end contol (no netwok assistance) sende limits tansmission: LastByteSent-LastByteAcked Roughly, CongWin CongWin ate = Bytes/sec RTT CongWin is dynamic, function of peceived netwok congestion How does sende peceive congestion? loss event = timeout o 3 duplicate acks TCP sende educes ate (CongWin) afte loss event thee mechanisms: m AIMD m slow stat m consevative afte timeout events Tanspot Laye 3-3

4 Stating a TCP tansmission A new offeed flow may suddenly oveload netwok nodes m eceive window is used to avoid ecv buffe oveflow m But it may be a lage value (16-64 KB) Idea: slow stat m Stat with small value of cwnd m And incease it as soon as packets get though Aival of ACKs = no packet losts = no congestion Initial cwnd size: m Just 1 MSS! m Recent (1998) poposals fo moe aggessive stats (up to 4 MSS) have been found to be dangeous Tanspot Laye 3-4

5 Detecting congestion and estating Segment gets lost m Detected via RTO expiation m Indiectly notifies that one of the netwok nodes along the path has lost segment Because of full queue Restat fom cwnd=1 (slow stat) But intoduce a supplementay contol: slow stat theshold sstesh = max(min(cwnd,window)/2,2mss) m The idea is that we now KNOW that thee is congestion in the netwok, and we need to incease ou ate in a moe caeful manne m Ssthesh defines the congestion avoidance egion Tanspot Laye 3-5

6 Congestion avoidance If cwnd < ssthesh m Slow stat egion: Incease ate exponentially If cwnd >= ssthesh m Congestion avoidance egion : Incease ate linealy Coisponde ad un segmento pe finesta m At ate 1 MSS pe RTT Pactical implementation: cwnd += MSS*MSS/cwnd Good appoximation fo 1 MSS pe RTT Altenative (exact) implementations: count!! Which initial ssthesh? ssthesh initially set to 65535: uneachable! In essence, congestion avoidance is flow contol imposed by sende while advetised window is flow contol imposed by eceive Tanspot Laye 3-6

7 Congestion window cwnd (in MSS) Simplified example (oveall) Timeout: cwnd = 1 ssthesh=8 Timeout: cwnd = 1 ssthesh=6 1 Numbe of tansmissions Tanspot Laye 3-7

8 The Fast Retansmit Algoithm Idea: use duplicate ACKs! Receive esponds with an ACK evey time it eceives an outof-ode segment ACK value = last coectly eceived segment FAST RETRANSMIT algoithm: if 3 duplicate acks ae eceived fo the same segment, assume that the next segment has been lost. Retansmit it ight away. Helps if single packet lost. Not vey effective with multiple losses And then? A congestion contol issue RTO : FR Seq=50 Seq=100 Seq=150 Seq=100 Tanspot Laye 3-8

9 What happens AFTER RTO? (without fast etansmit)seq=100 Seq=50 Cuent cwnd = 6 RTO Seq=150 Seq=350 set cwnd = 1 and tx seq=100 ack=400! And then, estat nomally with cwnd=2 and send seq=400,450 Tanspot Laye 3-9

10 Idea del fast etansmit Dovebbe potae ad un Diveso modo di gestie L evento da pate del Contollo di congetsione? Cuent cwnd = 6 TCP RENO (with fast etansmit) RTO Seq=50 Seq=100 Seq=150 set cwnd = 1 and tx seq=100 Seq=350 Seq=100 ack=400! And then, estat nomally with cwnd=2 and send seq=400,450 Same as befoe, but shote time to ecove packet loss! Tanspot Laye 3-10

11 Motivations fo fast ecovey FAST RECOVERY: The phase following fast etansmit (3 duplicate acks eceived) Seq=50 Seq=100 TAHOE appoach: slow stat, to potect netwok afte congestion Seq=150 Howeve, since subsequent acks have been eceived, no had congestion situation should be pesent in the netwok: slow stat is a too consevative estat! 3d dupack Seq=350 Seq=100 Tanspot Laye 3-11

12 FAST RECOVERY RULES: Retansmit lost segment Set cwnd = cwnd/2 Restat with congestion avoidance (linea) stat fast ecovey phase: Set counte fo duplicate packets ndup=3 Use inflated window: w = cwnd+ndup Upon new dup_acks, incease ndup, not cwnd (and send new data) Upon ecovey ack, deflate window setting ndup=0 Fast ecovey ules cwnd = 6 Fast Retansmit & ecovey: cwnd=3, ndup=3 cwnd=3, ndup=4 cwnd=3, ndup=5 Recovey ack=400 cwnd=3 Seq=50 Seq=100 Seq=150 Seq=350 Seq=100 Seq=400 Seq=450 Seq=500 Tanspot Laye 3-12

13 Idle peiods Afte a long idle peiod (exceeding one RTO), eset the congestion window to one. Congestion Window CWND 1 Timeout SSThesh Time Receive Window Idle Inteval Tanspot Laye 3-13

14 Futhe TCP issues Timeout = packet loss occuence in an intenal netwok oute TCP (both Tahoe & Reno) does not AVOID packet loss Simply REACTS to packet loss Timeout: CONCLUSION: a TCP able to AVOID packet loss should be much bette.. Timeout: Towad next Timeout cwnd Numbe of tansmissions Tanspot Laye 3-14

15 TCP Fainess Fainess goal: if K TCP sessions shae same bottleneck link of bandwidth R, each should have aveage ate of R/K TCP connection 1 TCP connection 2 bottleneck oute capacity R Tanspot Laye 3-15

16 Why is TCP fai? Two competing sessions: Additive incease gives slope of 1, as thoughout inceases multiplicative decease deceases thoughput popotionally R equal bandwidth shae Connection 2 thoughput Connection 1 thoughput loss: decease window by facto of 2 congestion avoidance: additive incease loss: decease window by facto of 2 congestion avoidance: additive incease R Tanspot Laye 3-16

17 Fainess with UDP taffic A seious poblem fo TCP m in heavy netwok load, TCP educes tansmission ate. Non congestion-contolled taffic does not. m Result: in link oveload, TCP thoughput vanishes! This is why we still live in a Wold Wide Wait time (Webcams ae destoying TCP taffic) Tanspot Laye 3-17

18 Mixing TCP & UDP taffic Link 45 Mbps UDP TCP UDP TCP1 TCP2 Tanspot Laye 3-18

19 Fainess (moe) Fainess and UDP Multimedia apps often do not use TCP m do not want ate thottled by congestion contol Instead use UDP: m pump audio/video at constant ate, toleate packet loss Reseach aea: TCP fiendly Fainess and paallel TCP connections nothing pevents app fom opening paallel connections between 2 hosts. Web bowses do this Example: link of ate R suppoting 9 cnctions; m new app asks fo 1 TCP, gets ate R/10 m new app asks fo 11 TCPs, gets R/2! Tanspot Laye 3-19