No connection establishment Do not perform Flow control Error control Retransmission Suitable for small request/response scenario E.g.

Transcription

1 UDP & TCP 2018/3/26

2 UDP Header

3 Characteristics of UDP No connection estabishment Do not perform Fow contro Error contro Retransmission Suitabe for sma request/response scenario E.g., DNS

4 Remote Procedure Ca (RPC)

5 The Rea-Time Transport Protoco (a) The position of RTP in the protoco stack. (b) Packet nesting.

6 RTP Header P: padding to a mutipe of 4 X: extension header CC: contributing sources Payoad type: encoding agorithm Sync source id: stream ID

7 Transmission Contro Protoco (TCP) TCP Segment Header TCP Connection Estabishment & Reease TCP Connection Management Modeing TCP Transmission Poicy TCP Congestion Contro TCP Timer Management Transactiona TCP

8 TCP Basics 16-bits port We-known ports < 1024 Connections: Fu dupex and point-to-point Byte stream, not messages Urgent data E.g., when user hit DEL or CTRL-C, send a data Receiver is interrupted (signaed)

9 Some TCP Services Port Protoco Use 21 FTP Fie transfer 23 Tenet Remote ogin 25 SMTP E-mai 69 TFTP Trivia Fie Transfer Protoco 79 Finger Lookup info about a user 80 HTTP Word Wide Web 110 POP-3 Remote e-mai access 119 NNTP USENET news

10 The TCP Service Mode (a) Four 512-byte segments sent as separate IP datagrams. (b) The 2048 bytes of data deivered to the appication in a singe READ CALL.

11 TCP Header

12 TCP Options Specify maximum TCP payoad (defaut 536 bytes) MSS Negotiate a window scae scae factor for enarging window size (max 14-bit to the eft) NAK: i.e., seective repeat instead of go-backn

13 TCP Connection Estabishment 6-31 (a) TCP connection estabishment in the norma case. (b) Ca coision.

14 Cosing TCP Connection Cosing a connection: cient coses socket: cientsocket.cose(); Step 1: cient end system sends TCP FIN contro segment to server Step 2: server receives FIN, repies with ACK. Coses connection, sends FIN. cient cose timed wait cosed server cose 14

15 Cosing TCP Connection (cont.) Step 3: cient receives FIN, repies with ACK. Enters timed wait - wi respond with ACK to received FINs Step 4: server, receives ACK. Connection cosed. Note: with sma modification, can hande simutaneous FINs. cient cosing timed wait cosed server cosing cosed 15

16 TCP Connection States

17 TCP Connection Management Modeing Heavy soid ine: norma path for a cient Heavy dashed ine: norma path for a server Light ines: unusua events. Transition is abeed by the event causing it and the action resuting from it.

18 TCP Fow Contro (Receiver Window Management)

19 Nage's Agorithm Probem: send sma packets, get ack's, and waste bandwidth Soution: Data come to the sender one byte at a time Send first byte and buffer a the rest When ack is received, send a in the buffer Exceptions Mouse movement in X window apps

20 Siy Window Syndrome Cark's soution: deay window update unti some amount of space is avaiabe

21 TCP Congestion Contro End-end contro (no network assistance) Sender imits transmission: LastByteSent-LastByteAcked == CongWin Roughy, rate = CongWin RTT Bytes/sec CongWin is dynamic, function of perceived network congestion How does sender perceive congestion? oss event = timeout or 3 dupicate acks TCP sender reduces rate (CongWin) after oss event three mechanisms: AIMD sow start conservative after timeout events 21

22 TCP AIMD mutipicative decrease: cut CongWin in haf after oss event additive increase: increase CongWin by 1 MSS every RTT in the absence of oss events: probing Long-ived TCP connection MSS (Max Segment Size) 22

23 TCP Sow Start When connection begins, CongWin = 1 MSS Exampe: MSS = 500 bytes & RTT = 200 msec When connection begins, increase rate exponentiay fast unti first oss event initia rate = 20 kbps avaiabe bandwidth may be >> MSS/RTT desirabe to quicky ramp up to respectabe rate 23

24 TCP Sow Start (more) When connection begins, increase rate exponentiay unti first oss event: doube CongWin every RTT done by incrementing CongWin for every ACK received Summary: initia rate is sow but ramps up exponentiay fast RTT Host A Host B time 24

25 Refinement After 3 dup ACKs: CongWin is cut in haf window then grows ineary But after timeout event: CongWin instead set to 1 MSS; window then grows exponentiay to a threshod, then grows ineary Phiosophy: 3 dup ACKs indicates network capabe of deivering some segments timeout before 3 dup ACKs is more aarming 25

26 Refinement (more) Q: When shoud the exponentia increase switch to inear? A: When CongWin gets to 1/2 of its vaue before timeout. Impementation: Variabe Threshod At oss event, Threshod is set to 1/2 of CongWin just before oss event 26

27 Summary: TCP Congestion Contro When CongWin is beow Threshod, sender in sowstart phase, window grows exponentiay. When CongWin is above Threshod, sender is in congestion-avoidance phase, window grows ineary. When a tripe dupicate ACK occurs, Threshod set to CongWin/2 and CongWin set to Threshod. When timeout occurs, Threshod set to CongWin/2 and CongWin is set to 1 MSS. 27

28 TCP Sender Congestion Contro Event State TCP Sender Action Commentary ACK receipt for previousy unacked data Sow Start (SS) CongWin = CongWin + MSS, If (CongWin > Threshod) set state to Congestion Avoidance Resuting in a doubing of CongWin every RTT ACK receipt for previousy unacked data Congestion Avoidance (CA) CongWin = CongWin+MSS * (MSS/CongWin) Additive increase, resuting in increase of CongWin by 1 MSS every RTT Loss event detected by tripe dupicate ACK SS or CA Threshod = CongWin/2, CongWin = Threshod, Set state to Congestion Avoidance Fast recovery, impementing mutipicative decrease. CongWin wi not drop beow 1 MSS. Timeout SS or CA Threshod = CongWin/2, CongWin = 1 MSS, Set state to Sow Start Enter sow start Dupicate ACK SS or CA Increment dupicate ACK count for segment being acked CongWin and Threshod not changed 28

29 TCP Timer Management (a) Probabiity density of ACK arriva times in the data ink ayer. (b) Probabiity density of ACK arriva times for TCP.

30 Agorithm for Retransmit Timer Maintain variabe RTT (round-trip-time) RTT = α RTT + (1-α)M Typicay, α = 7/8 Timeout is β RTT Jacobson propose making proportiona to standard deviation of ack arriva time PDF RTT-M ( 1-α ) Deviation D = αd + Timeout = RTT + 4 X D

31 Linear Contro Many different possibiities for reaction to congestion and probing Examine simpe inear contros Window(t + 1) = a + b Window(t) Different a i /b i for increase and a d /b d for decrease Supports various reaction to signas Increase/decrease additivey Increased/decrease mutipicativey Which of the four combinations is optima? 31

32 Outine TCP advantages in Congestion contro 32

33 Objectives Simpe router behavior Distributedness Efficiency: X = Σx i (t) Fairness: (Σx i ) 2 /n(σx i2 ) What are the important properties of this function? Convergence: contro system must be stabe 33

34 Basic Contro Mode Reduce speed when congestion is perceived How is congestion signaed? Either mark or drop packets How much to reduce? Increase speed otherwise Probe for avaiabe bandwidth how? 34

35 Phase Pots Simpe way to visuaize behavior of competing connections over time User 2 s Aocation x 2 User 1 s Aocation x 1 35

36 Phase Pots What are desirabe properties? Fairness Line What if fows are not equa? User 2 s Aocation x 2 Overoad Optima point Underutiization Efficiency Line User 1 s Aocation x 1 36

37 Additive Increase/Decrease Both X 1 and X 2 increase/ decrease by the same amount over time Additive increase improves fairness and additive decrease reduces fairness User 2 s Aocation x 2 T 0 T 1 Fairness Line Efficiency Line User 1 s Aocation x 1 37

38 Muipicative Increase/Decrease Both X 1 and X 2 increase by the same factor over time Extension from origin constant fairness User 2 s Aocation x 2 T 0 T 1 Fairness Line Efficiency Line User 1 s Aocation x 1 38

39 Convergence to Efficiency Fairness Line x H User 2 s Aocation x 2 Efficiency Line User 1 s Aocation x 1 39

40 Distributed Convergence to Efficiency a<0 & b>1 a=0 a>0 & b>1 b=1 Fairness Line x H User 2 s Aocation x 2 a>0 & b<1 a<0 & b<1 Efficiency Line User 1 s Aocation x 1 40

41 Convergence to Fairness Fairness Line x H User 2 s Aocation x 2 x H Efficiency Line User 1 s Aocation x 1 41

42 Convergence to Efficiency & Fairness Intersection of vaid regions For decrease: a=0 & b < 1 Fairness Line x H User 2 s Aocation x 2 x H Efficiency Line User 1 s Aocation x 1 42

43 What is the Right Choice? Constraints imit us to AIMD Can have mutipicative term in increase (MAIMD) AIMD moves towards optima point User 2 s Aocation x 2 x 0 x 2 x 1 Fairness Line Efficiency Line *Additive Increase Mutipicative Decrease User 1 s Aocation x 1 43

44 Important Lessons Transport service UDP à mosty just IP service TCP à congestion controed, reiabe, byte stream Types of ARQ protocos Stop-and-wait à sow, simpe Go-back-n à can keep ink utiized (except w/ osses) Seective repeat à efficient oss recovery Siding window fow contro TCP fow contro Siding window à mapping to packet headers 32bit sequence numbers (bytes) 44

45 Other TCP Timers Persistence timer: to probe receiver's window size Keepaive timer: ide for a whie, check if peer is aive TIMED WAIT state timer: 2 max packet ifetime

46 Transactiona TCP (T/TCP) (a) RPC using norma TCP. (b) RPC using T/TCP.

47 Network Performance Measurement The basic oop for improving network performance. Measure reevant network parameters, performance. Try to understand what is going on. Change one parameter.

48 System Design for Better Performance CPU speed is more important than network speed. Reduce packet count to reduce software overhead. Minimize context switches. Minimize copying. You can buy more bandwidth but not ower deay. Avoiding congestion is better than recovering from it. Avoid timeouts.