Network Protocol Design (ITC8061) - Part II -

Transcription

1 Netwrk Prtcl Design (ITC8061) - Part II -

2 utline v 3.1 Transprt-layer services v 3.2 Multiplexing and demultiplexing v 3.3 Cnnectinless transprt: UDP v 3.4 Principles f reliable data transfer v 3.5 Cnnectinriented transprt: TCP segment structure reliable data transfer flw cntrl cnnectin management v 3.6 Principles f cngestin cntrl v 3.7 TCP cngestin cntrl 1-2

3 TCP: Overview RFCs: 793, 1122, 1323, 2018, 2581 v pint-t-pint: ne sender, ne receiver v reliable, in-rder byte steam: n message bundaries v pipelined: TCP cngestin and flw cntrl set windw size v send & receive buffers scket dr applicatin writes data TCP send buffer v full duplex data: bi-directinal data flw in same cnnectin MSS: maximum segment size v cnnectin-riented: handshaking (exchange f cntrl msgs) init s sender, receiver state befre data exchange v flw cntrlled: sender will nt verwhelm receiver segment applicatin reads data TCP receive buffer scket dr 1-3

4 TCP segment structure URG: urgent data (generally nt used) ACK: ACK # valid PSH: push data nw (generally nt used) RST, SYN, FIN: cnnectin estab (setup, teardwn cmmands) Internet checksum (as in UDP) head len 32 bits surce prt # dest prt # sequence number acknwledgement number nt used U AP R S F checksum Receive windw applicatin data (variable length) Urg data pnter Optins (variable length) cunting by bytes f data (nt segments!) # bytes rcvr willing t accept 1-4

5 Cmpare with UDP 32 bits surce prt # dest prt # sequence number acknwledgement number L N/A Bits Receive windw checksum Urgent data ptr 32 bits surce prt # dest prt # length checksum Optins (variable length) Applicatin data (variable length) Applicatin data (variable length) v Mre Functinality v Mre Header Overhead (12 bytes + ptins [0 40]) 1-5

6 TCP seq. # s and ACKs Seq. # s: byte stream number f first byte in segment s data ACKs: seq # f next byte expected frm ther side cumulative ACK Q: hw receiver handles ut-f-rder segments A: TCP spec desn t say, - up t implementr User types C hst ACKs receipt f eched C Hst A Hst B Seq=42, ACK=79, data = C Seq=79, ACK=43, data = C Seq=43, ACK=80 simple telnet scenari hst ACKs receipt f C, eches back C time 1-6

7 TCP Rund Trip Time and Timeut Q: hw t set TCP timeut value? v lnger than RTT but RTT varies v t shrt: premature timeut unnecessary retransmissins v t lng: slw reactin t segment lss Q: hw t estimate RTT? v SampleRTT: measured time frm segment transmissin until ACK receipt ignre retransmissins v SampleRTT will vary, want estimated RTT smther average several recent measurements, nt just current SampleRTT 1-7

8 TCP Rund Trip Time and Timeut EstimatedRTT = (1 - a)* EstimatedRTT + a * SampleRTT Expnential weighted mving average influence f past sample decreases expnentially fast typical value: a =

9 Example RTT estimatin: 1-9

10 TCP Rund Trip Time and Timeut Setting the timeut v EstimtedRTT plus safety margin large variatin in EstimatedRTT -> larger safety margin v first estimate f hw much SampleRTT deviates frm EstimatedRTT: DevRTT = (1-b)*DevRTT + b* SampleRTT-EstimatedRTT (typically, b = 0.25) Then set timeut interval: TimeutInterval = EstimatedRTT + 4*DevRTT 1-10

11 utline v 3.1 Transprt-layer services v 3.2 Multiplexing and demultiplexing v 3.3 Cnnectinless transprt: UDP v 3.4 Principles f reliable data transfer v 3.5 Cnnectinriented transprt: TCP segment structure reliable data transfer flw cntrl cnnectin management v 3.6 Principles f cngestin cntrl v 3.7 TCP cngestin cntrl 1-11

12 TCP reliable data transfer v TCP creates rdt service n tp f IP s unreliable service v Pipelined segments v Cumulative acks v TCP uses single retransmissin timer v Retransmissins are triggered by: timeut events duplicate acks v Initially cnsider simplified TCP sender: ignre duplicate acks ignre flw cntrl, cngestin cntrl 1-12

13 TCP sender events: data rcvd frm app: v Create segment with seq # v seq # is byte-stream number f first data byte in segment v start timer if nt already running (think f timer as fr ldest unacked segment) v expiratin interval: TimeOutInterval timeut: v retransmit segment that caused timeut v restart timer Ack rcvd: v If acknwledges previusly unacked segments update what is knwn t be acked start timer if there are utstanding segments 1-13

14 TCP: retransmissin scenaris Hst A Hst B Hst A Hst B Seq=92, 8 bytes data timeut SendBase = 100 Seq=92, 8 bytes data X lss ACK=100 Seq=92, 8 bytes data ACK=100 time lst ACK scenari Sendbase = 100 SendBase = 120 SendBase = 120 Seq=92 timeut Seq=92 timeut time Seq=100, 20 bytes data ACK=100 ACK=120 Seq=92, 8 bytes data ACK=120 premature timeut 1-14

15 TCP retransmissin scenaris (mre) Hst A Hst B Seq=92, 8 bytes data timeut Seq=100, 20 bytes data X lss ACK=100 SendBase = 120 ACK=120 time Cumulative ACK scenari 1-15

16 TCP ACK generatin [RFC 1122, RFC 2581] Event at Receiver Arrival f in-rder segment with expected seq #. All data up t expected seq # already ACKed Arrival f in-rder segment with expected seq #. One ther segment has ACK pending Arrival f ut-f-rder segment higher-than-expect seq. #. Gap detected Arrival f segment that partially r cmpletely fills gap TCP Receiver actin Delayed ACK. Wait up t 500ms fr next segment. If n next segment, send ACK Immediately send single cumulative ACK, ACKing bth in-rder segments Immediately send duplicate ACK, indicating seq. # f next expected byte Immediate send ACK, prvided that segment starts at lwer end f gap 1-16

17 Fast Retransmit v Time-ut perid ften relatively lng: lng delay befre resending lst packet v Detect lst segments via duplicate ACKs. Sender ften sends many segments backt-back If segment is lst, there will likely be many duplicate ACKs. v If sender receives 3 ACKs fr the same data, it suppses that segment after ACKed data was lst: fast retransmit: resend segment befre timer expires 1-17

18 utline v 3.1 Transprt-layer services v 3.2 Multiplexing and demultiplexing v 3.3 Cnnectinless transprt: UDP v 3.4 Principles f reliable data transfer v 3.5 Cnnectinriented transprt: TCP segment structure reliable data transfer flw cntrl cnnectin management v 3.6 Principles f cngestin cntrl v 3.7 TCP cngestin cntrl 1-18

19 TCP Flw Cntrl v receive side f TCP cnnectin has a receive buffer: app prcess may be slw at reading frm buffer flw cntrl sender wn t verflw receiver s buffer by transmitting t much, t fast speed-matching service: matching the send rate t the receiving app s drain rate 1-19

20 TCP Flw cntrl: hw it wrks (Suppse TCP receiver discards ut-f-rder segments) v spare rm in buffer = RcvWindw = RcvBuffer-[LastByteRcvd - LastByteRead] v Rcvr advertises spare rm by including value f RcvWindw in segments v Sender limits unacked data t RcvWindw guarantees receive buffer desn t verflw 1-20

21 utline v 3.1 Transprt-layer services v 3.2 Multiplexing and demultiplexing v 3.3 Cnnectinless transprt: UDP v 3.4 Principles f reliable data transfer v 3.5 Cnnectinriented transprt: TCP segment structure reliable data transfer flw cntrl cnnectin management v 3.6 Principles f cngestin cntrl v 3.7 TCP cngestin cntrl 1-21

22 TCP Cnnectin Management Recall: TCP sender, receiver establish cnnectin befre exchanging data segments v initialize TCP variables: seq. #s buffers, flw cntrl inf (e.g. RcvWindw) v client: cnnectin initiatr Scket clientscket = new Scket("hstname","prt number"); v server: cntacted by client Scket cnnectinscket = welcmescket.accept(); Three way handshake: Step 1: client hst sends TCP SYN segment t server specifies initial seq # n data Step 2: server hst receives SYN, replies with SYNACK segment server allcates buffers specifies server initial seq. # Step 3: client receives SYNACK, replies with ACK segment, which may cntain data 1-22

23 TCP Cnnectin Management (cnt.) Clsing a cnnectin: client server client clses scket: clientscket.clse(); clse FIN Step 1: client end system sends TCP FIN cntrl segment t server Step 2: server receives FIN, replies with ACK. Clses cnnectin, sends FIN. timed wait clsed ACK FIN ACK clse 1-23

24 TCP Cnnectin Management (cnt.) Step 3: client receives FIN, replies with ACK. Enters timed wait - will respnd with ACK t received FINs Step 4: server, receives ACK. Cnnectin clsed. Nte: with small mdificatin, can handle simultaneus FINs. clsing timed wait clsed client FIN ACK FIN ACK server clsing clsed 1-24

25 TCP Cnnectin Management (cnt) TCP server lifecycle TCP client lifecycle 1-25

26 utline v 3.1 Transprt-layer services v 3.2 Multiplexing and demultiplexing v 3.3 Cnnectinless transprt: UDP v 3.4 Principles f reliable data transfer v 3.5 Cnnectinriented transprt: TCP segment structure reliable data transfer flw cntrl cnnectin management v 3.6 Principles f cngestin cntrl v 3.7 TCP cngestin cntrl 1-26

27 Principles f Cngestin Cntrl Cngestin: v infrmally: t many surces sending t much data t fast fr netwrk t handle v different frm flw cntrl! v manifestatins: lst packets (buffer verflw at ruters) lng delays (queueing in ruter buffers) v a tp-10 prblem! 1-27

28 Causes/csts f cngestin: scenari 1 v tw senders, tw receivers v ne ruter, infinite buffers v n retransmissin Hst B Hst A lin : riginal data unlimited shared utput link buffers l ut v large delays when cngested v maximum achievable thrughput 1-28

29 Causes/csts f cngestin: scenari 2 v ne ruter, finite buffers v sender retransmissin f lst packet Hst A l in : riginal data l ut l' in : riginal data, plus retransmitted data Hst B finite shared utput link buffers 1-29

30 Causes/csts f cngestin: scenari 2 v always: l = (gdput) in l ut v perfect retransmissin nly when lss: l > in l ut v retransmissin f delayed (nt lst) packet makes larger (than perfect case) fr same l ut l in csts f cngestin: mre wrk (retrans) fr given gdput unneeded retransmissins: link carries multiple cpies f pkt 1-30

31 Causes/csts f cngestin: scenari 3 v fur senders v multihp paths v timeut/retransmit Q: what happens as and increase? l in l in Hst A l in : riginal data l' in : riginal data, plus retransmitted data finite shared utput link buffers l ut Hst B 1-31

32 Causes/csts f cngestin: scenari 3 H s t A l u t H s t B Anther cst f cngestin: when packet drpped, any upstream transmissin capacity used fr that packet was wasted! 1-32

33 Appraches twards cngestin cntrl Tw brad appraches twards cngestin cntrl: End-end cngestin cntrl: v n explicit feedback frm netwrk v cngestin inferred frm end-system bserved lss, delay v apprach taken by TCP Netwrk-assisted cngestin cntrl: v ruters prvide feedback t end systems single bit indicating cngestin (SNA, DECbit, TCP/IP ECN, ATM) explicit rate sender shuld send at 1-33

34 utline v 3.1 Transprt-layer services v 3.2 Multiplexing and demultiplexing v 3.3 Cnnectinless transprt: UDP v 3.4 Principles f reliable data transfer v 3.5 Cnnectinriented transprt: TCP segment structure reliable data transfer flw cntrl cnnectin management v 3.6 Principles f cngestin cntrl v 3.7 TCP cngestin cntrl 1-34

35 TCP Cngestin Cntrl v end-end cntrl (n netwrk assistance) v sender limits transmissin: LastByteSent-LastByteAcked CngWin v Rughly, rate = CngWin RTT Bytes/sec v CngWin is dynamic, functin f perceived netwrk cngestin Hw des sender perceive cngestin? v lss event = timeut r 3 duplicate acks v TCP sender reduces rate (CngWin) after lss event three mechanisms: AIMD slw start cnservative after timeut events 1-35

36 TCP AIMD multiplicative decrease: cut CngWin in half after lss event additive increase: increase CngWin by 1 MSS every RTT in the absence f lss events: prbing 24 Kbytes cngestin windw 16 Kbytes 8 Kbytes time Lng-lived TCP cnnectin 1-36

37 TCP Slw Start v When cnnectin begins, CngWin = 1 MSS Example: MSS = 500 bytes & RTT = 200 msec initial rate = 20 kbps v available bandwidth may be >> MSS/RTT desirable t quickly ramp up t respectable rate When cnnectin begins, increase rate expnentially fast until first lss event 1-37

38 TCP Slw Start (mre) v When cnnectin begins, increase rate expnentially until first lss event: duble CngWin every RTT dne by incrementing CngWin fr every ACK received v Summary: initial rate is slw but ramps up expnentially fast RTT Hst A Hst B ne segment tw segments fur segments time 1-38

39 Refinement v v After 3 dup ACKs: CngWin is cut in half windw then grws linearly But after timeut event: CngWin instead set t 1 MSS; windw then grws expnentially t a threshld, then grws linearly Philsphy: 3 dup ACKs indicates netwrk capable f delivering sme segments timeut befre 3 dup ACKs is mre alarming 1-39

40 Refinement (mre) Q: When shuld the expnential increase switch t linear? A: When CngWin gets t 1/2 f its value befre timeut. TCP Tahe threshld TCP Ren Implementatin: v Variable Threshld v At lss event, Threshld is set t 1/2 f CngWin just befre lss event 1-40

41 Summary: TCP Cngestin Cntrl v When CngWin is belw Threshld, sender in slw-start phase, windw grws expnentially. v When CngWin is abve Threshld, sender is in cngestin-avidance phase, windw grws linearly. v When a triple duplicate ACK ccurs, Threshld set t CngWin/2 and CngWin set t Threshld. v When timeut ccurs, Threshld set t CngWin/2 and CngWin is set t 1 MSS. 1-41

42 TCP Fairness Fairness gal: if K TCP sessins share same bttleneck link f bandwidth R, each shuld have average rate f R/K TCP cnnectin 1 TCP cnnectin 2 bttleneck ruter capacity R 1-42

43 Why is TCP fair? Tw cmpeting sessins: v Additive increase gives slpe f 1, as thrughut increases v multiplicative decrease decreases thrughput prprtinally R equal bandwidth share Cnnectin 2 thrughput lss: decrease windw by factr f 2 cngestin avidance: additive increase lss: decrease windw by factr f 2 cngestin avidance: additive increase Cnnectin 1 thrughput R 1-43

44 Fairness (mre) Fairness and UDP v Multimedia apps ften d nt use TCP d nt want rate thrttled by cngestin cntrl v Instead use UDP: pump audi/vide at cnstant rate, tlerate packet lss v Research area: TCP friendly Fairness and parallel TCP cnnectins v nthing prevents app frm pening parallel cnnectins between 2 hsts. v Web brwsers d this v Example: link f rate R supprting 9 cnnectins; new app asks fr 1 TCP, gets rate R/10 new app asks fr 11 TCPs, gets R/2! 1-44