Information Network 1 Transport layer: TCP. Youki Kadobayashi Doudou Fall NAIST

Transcription

1 Iformatio Network 1 Trasport layer: TCP Youki Kadobayashi Doudou Fall NAIST

2 Trasport layer: a birds-eye view Hosts maitai state for each trasport-layer edpoit Routers do t maitai per-host state H R R R R H Trasport IP IP IP IP IP Applicatio Presetatio Sessio Trasport Network Data Lik Physical Applicatio Presetatio Sessio Trasport Network Data Lik Physical 2

3 Fuctios provided by the trasport layer Commuicatio betwee processes Idetify process Idetify iter-process chael P (P, Q) Q Iterface for upper layer Coectio-orieted (virtual circuit) Coectioless (datagram) Competitio ad arbitratio of etwork resource Flow cotrol Cogestio cotrol Next lecture 3

4 Trasport protocols i the Iteret protocol suite TCP (RFC793) UDP (RFC768) Trasmissio Cotrol Protocol User Datagram Protocol Coectio-orieted Coectioless Multiple fuctios IP & Process idetificatio for reliability Packets ca be lost SCTP (RFC4960) DCCP (RFC4340) Advaced topic 4

5 Idetifyig process ad coectio i TCP Uique idetificatio of process (IP, port) Uique idetificatio of TCP coectio (source IP, source port, destiatio IP, destiatio port) ( , 1040) ( , 22) 1040 coectio 22 process 80 coectio

6 Some well-kow protocols/ports Port Protocol Use 20,21 FTP File trasfer 22 SSH Remote logi, replacemet for telet 25 SMTP 80 HTTP World Wide Web 110 POP-3 Remote access 143 IMAP Remote access 443 HTTPS Secure Web (HTTP over SSL/TLS) 543 RTSP Media player cotrol 631 IPP Priter sharig 6

7 How SSH port became 22 (1) 7

8 How SSH port became 22 (1) 8

9 TCP service model (1) Coectio-orieted Virtual Circuit Looks like a circuit, but virtual: o physical wires betwee specific edpoits Adapts to speed Adapts to speed of itermediate etworks as well as the speed of edpoits 9

10 Establishig a TCP coectio 3-way hadshake SYN, SYN-ACK, ACK Esure full-duplex commuicatio SYN URG ACK PSH RST SYN FIN SYN-ACK 16bit source port 16bit destiatio port 32bit sequece umber 32bit ackowledgmet umber 4bit hle reserved flags 16bit TCP checksum 16bit widow size 16bit urget poiter ACK 10

11 Establishig TCP coectio: a cocrete example with Wireshark # tshark -i e0 - -f 'port 80' tcpdump: listeig o de0 Capturig o e > TCP > 80 [SYN] Seq=0 Wi=65535 Le=0 MSS=1460 WS=16 TSval= TSecr=0 SACK_PERM= > TCP 80 > [SYN, ACK] Seq=0 Ack=1 Wi=50137 Le=0 TSval= TSecr= MSS=1460 WS=8 SACK_PERM= > TCP > 80 [ACK] Seq=1 Ack=1 Wi= Le=0 TSval= TSecr= Reply with Sequece umber + 1 as a ACK implies ackowledgmet 11

12 Meaigs of Wireshark output time source IP -> dest IP TCP source port > dest port [ flags ] Seq= Ack= > TCP 80 > [SYN, ACK] Seq=0 Ack=1 Wi=50137 Le=0 TSval= TSecr= MSS=1460 WS=8 SACK_PERM=1 12

13 Virtual Circuit(2):TCP coectio release close FIN Ack of FIN FIN close Ack of FIN > TCP 80 > [FIN, ACK] Seq=659 Ack=2449 Wi= Le=0 TSval= TSecr= > TCP > 80 [ACK] Seq=2449 Ack=660 Wi= Le=0 TSval= TSecr= > TCP > 80 [FIN, ACK] Seq=2449 Ack=660 Wi= Le=0 TSval= TSecr= > TCP 80 > [ACK] Seq=660 Ack=2450 Wi= Le=0 TSval= TSecr=

14 TCP coectio reset RST Abortive release Noexistet port (e.g., dead process) telet to sh.aist.jp, port 80 will result i coectio reset: > TCP > 80 [SYN] Seq=0 Wi=65535 Le=0 MSS=1460 WS=16 TSval= TSecr=0 SACK_PERM= > TCP 80 > [RST, ACK] Seq=1 Ack=1 Wi=0 Le=0 If you kill Dropbox : > TCP > 80 [RST] Seq=310 Wi=0 Le=0 14

15 Virtual Circuit Summary: TCP state trasitio diagram State trasitio: trigger / respose 15

16 Troubleshootig TCP with state machie etstat $ etstat Active Iteret coectios Proto Recv-Q Sed-Q Local Address Foreig Address (state) tcp http SYN_SENT tcp ey-i-f101.1e100.http ESTABLISHED tcp http ESTABLISHED SYN set, awaitig SYN+ACK respose May other ope-source tools are available: lsof, trpt, tcptraceroute, tcptrace, tcpflow, etc. Try some of these tools i the provided VM image. 16

17 SYN packet lost SYN packet is lost: Iside the etwork or The server rejects the packet Evetually o SYN-ACK arrives Seder sets a timer ad wait for the SYN-ACK Seder retrasmits the SYN-ACK if eeded How should the TCP seder set the timer? Seder has o idea how far away the receiver is Hard to guess a reasoable legth of time to wait Some TCP use a default of 3 or 6 secods 17

18 Hads-o: observe coectio setup/release I the provided virtual machie, 1. Observe coectio setup/release; e.g., by usig browser withi VM 2. Observe coectio reset by termiatig program; e.g., by skill Firefox 3. Observe coectio reset by coectig to oexistet port o the workig machie. e.g., by telet sh.aist.jp 80 18

19 Buffered trasfer: adapts to varyig speed of edpoits, as well as itermediate etworks Process Process Write() Read() block/ublock Write() Read() sed buffer recv buffer sed buffer recv buffer TCP coectio OS kerel OS kerel 19

20 TCP service model (2) Byte-stream service Upper layer ca t see boudaries betwee packets o boudary: structurig (framig) is eeded at upper layer O L L E H OK TCP beig viewed as byte-stream service O L L E H OK Full duplex idepedet two streams i sigle coectio Reliable masks packet reorderig, duplicatio, discard ad bit error 20

21 How does TCP implemet reliable stream service? ACK: Ackowledgmet Active ackowledgmet Explicitly ackowledge the receipt of packet Duplicate ACK Implicitly commuicate the packet loss iformatio Timeout ad Retrasmissio Wheever seder does t receive ACK after fixed time, a Timeout evet is triggered p Retrasmissio: assumig that trasmissio has failed Expoetial back-off: 3, 6, 12, 21

22 ACK: Ackowledgmet Seder Set ad ackowledged Set but uackowledged Nara Istitute of Sciece ad Techology User data arrives Packets i trasit Receiver Nara Isti 22

23 Piggybackig: Exploitig full-duplex chael Note: rough sketch Seder Receiver Set ad ackowledged Set but uackowledged Nara Istitute of Sciece ad Techology User data arrives Graduate S Packets i trasit Receiver Seder Nara Isti Graduate School of Iformatio Sciece User data arrives Set but uackowledged Set ad ackowledged 23

24 Duplicate ACK: implicit commuicatio of packet loss Seder Set ad ackowledged Set but uackowledged Nara Istitute of Sciece ad Techology User data arrives Packet loss Packets i trasit Receiver Nara Istitute o 24

25 From byte-stream to packets: TCP header Chop appropriate legth from byte-stream buffer & add TCP header IP Header TCP Header TCP segmet TCP data 16bit source port 16bit destiatio port 32bit sequece umber 32bit ackowledgmet umber 4bit hle reserved flags 16bit widow size 16bit TCP checksum 16bit urget poiter (optios) 20 octets (TCP data) 25

26 Nagle algorithm Q. If you added 20byte+20byte size header to 1byte data, is t that overhead big? Nagle algorithm (RFC896) There is oly oe small segmet which is uackowledged i the etwork. I case of short RTT: à acceptable overhead, as LAN badwidth is abudat à sed packets with small bufferig I case of log RTT à reduce overhead, due to WAN badwidth costraits 26

27 Hads-o: observe packet loss & retrasmissio Withi provided virtual machie, 1. Observe duplicate ACK; Emulate lossy etwork with: # tc qdisc add dev eth1 root etem loss 10% 1. Observe the effect of Nagle algorithm; Emulate satellite etwork with: # tc qdisc add dev eth1 root etem delay 500ms 27

28 Questios? 28

29 Summary thus far Trasport layer The trasport protocol o the iteret TCP TCP: service model, ad features Efficiecy: ACK, piggybackig, Nagle algorithm TCP coectio establishmet ad release Diagosis: tools + state machie kowledge 29

30 Trasport protocol challeges With a umber of ukow parameters: Number of active commuicatios Bottleeck badwidth Error rate how ca we accommodate as much commuicatios as possible, without collapsig etwork itself? 1 1 k Key ideas: probig, estimatio, self-policig, macro-level stability 30

31 Flow cotrol ad cogestio cotrol: Competitio ad Arbitratio of Network Resource Flow cotrol Absorb differece of trasmissio rate Recovery from sequece error Recovery from duplicatio, packets drop ad bit error Cogestio cotrol Sharig the badwidth while suppressig the cogestio Fair sharig of badwidth 31

32 Characteristics of TCP Flow Cotrol Ed to ed No global assigmet of resource Estimate available badwidth at idividual hosts Routers do ot explicitly allocate resource Implicit sigalig through packet drops Scalable Host-based autoomous method has better scalability, because it does t eed state maagemet iside etwork. Autoomous à Less state maagemet à Scalable 32

33 TCP: key characteristics Very simple algorithm Macroscopic self-stabilizatio TCP does't assume the presece of greedy odes. Elimiatio of global cotrol system Reject the idea of itermediate policig system Modest performace across almost all data-liks As opposed to optimal performace i specific coditio Stepwise improvemets over 30 years 33

34 Flow Cotrol o TCP Cotrol available badwidth Slidig widow Usig sequece umber, ot packet umber Widow size Cotrol trasmissio iterval of packets ACK clockig Miscellaeous Error detectio with TCP checksum Packet drop detectio with duplicate ACK, timeout 34

35 Slidig widow Set ad ackowledged Set but uackowledged Seder Nara Istitute of Sciece ad Techology Widow size User data arrives Sequece umber Packets i flight Receiver Nara Isti 35

36 Cogestio Cotrol o TCP Fair-share model: fair distributio of badwidth Ed to ed Icrease ad decrease of widow size additive icrease multiplicative decrease AIMD is kow to be self-stabilizig (R. Jai, et al., Aalysis of the icrease ad decrease algorithms for cogestio avoidace i computer etworks, 1989) Switch icreasig strategy of widow size Detect cogestio through packet drops 36

37 Icreasig Widow size TCP switches icreasig strategy of cogestio widow (cwd) by slow start threshold (ssthresh) (Outlie of the algorithm) O receivig a ACK: if (cwd < ssthresh) { /* slow start: expoetial icrease */ cwd += 1; } else { /* cogestio avoidace: additive icrease */ cwd += 1 / cwd; } 37

38 Icreasig Widow size Slow start expoetial icrease Cogestio avoidace additive icrease Source: TCP/IP Illustrated, Vol.1 Effectiveess: see V. Jacobso, Cogestio Avoidace ad Cotrol, SIGCOMM

39 Decreasig Widow size (Outlie of algorithm ) O detectig packet drop: ssthresh = cwd / 2; if (timeout) { cwd = 1; } Source: TCP/IP Illustrated, Vol.1 39

40 Hads-o: observe slow start & cogestio avoidace Start Wireshark Access speed test website: Filter the TCP stream with: ip.src== Plot the Time-Sequece Graph Select oe of the filtered packets Go to Statistics -> TCP StreamGraph -> Time Sequece Graph (Steves) Go to Statistics -> TCP StreamGraph -> Time Sequece Graph (tcptrace) 40

41 Questios? 41

42 Packet drop Detectio with 2 methods: Duplicate ACK Retrasmissio Time Out (RTO) How do I judge the time out? RTO ~ RTT Estimator 42

43 RTO Calculatio Source: TCP/IP Illustrated, Vol.1 Err = M A A <- A + gerr D <- D + h( Err - D) RTO = A + 4D A: smoothed RTT D: smoothed mea deviatio g: gai for the average (1/8) h: gai for the deviatio (1/4) 43

44 RTO calculatio Source: A Quick Tour Aroud TCP, 44

45 Hads-o: observe the effect of loss/delay Start Wireshark Start iperf Istall iperf3 o the VM: $sudo apt-get istall iperf3 ru iperf3: $iperf3 -c p t 2 Cofigure tc Loss: $sudo tc qdisc add dev eth1 root etem loss 10% To cofirm: $sudo tc qdisc show To remove: $sudo tc qdisc del dev eth1 root etem loss 10% Delay: $sudo tc qdisc add dev eth1 root etem delay 500ms Observe the effect of loss/delay Select oe of the filtered packets Go to Statistics -> TCP StreamGraph -> Time Sequece Graph (Steves) Go to Statistics -> TCP StreamGraph -> Time Sequece Graph (tcptrace) 45

46 User Datagram Protocol (UDP) RFC 768 Ureliable Data Protocol (J) Coectioless trasport protocol Avoid overhead ad delays of ordered, reliable delivery Sed messages to ad receive them from a socket Lightweight delivery service IP ad port umbers to support (de)multiplexig Optioal error checkig o the packet cotets 46

47 Icetives o usig UDP Fier cotrol over what data is set ad whe UDP packages the data ad sed the packet as soo as a applicatio process writes ito the socket No delay for coectio establishmet There are o prelimiaries i UDP No coectio state No allocatio of buffers, parameters, sequece umbers, etc. Small packet header overhead: oly eight -bytes log 47

48 UDP applicatios Multimedia streamig Real-time Trasport Protocol (RTP): iteret radio, telephoy, music-o-demad, video-o-demad Retrasmissio of lost/corrupted packets is ot worthwhile Simple query protocols like Domai Name System Overhead of coectio establishmet is overkill Easier to have applicatio retrasmit if eeded 48

49 Summary Trasport layer The trasport protocol o the iteret TCP TCP: service model, ad features Efficiecy: ACK, piggybackig, Nagle algorithm TCP coectio establishmet ad release Flow cotrol ad cogestio cotrol i TCP 49