ECE 428 Transport-level Protocols (Layer 4) TCP: Transmission Control Protocol UDP: User Datagram Protocol

Transcription

1 ECE 428 Transport-level Protocols (Layer 4) TCP: Transmission Control Protocol UDP: User Datagram Protocol 1

2 Need for a Protocol above IP layer IP layer Delivers packets to a host from another host Delivery: best-effort basis Can reorder, lose, duplicate Is not sure if data has been delivered (no end-to-end ACK) Thus, need for an upper layer protocol TCP Deliver data to applications <= end-to-end semantics Maintain data flow between applications. Receiver s view» Reliable: Ordered, without loss, no duplicate. Drop data belonging to an earlier association between applications.» Flow control Sender s view: Confirmed delivery Congestion control: by the sender <= Reduce network congestion 2

3 Transport-level protocols TCP: Operates in a connection-oriented mode. Establish a connection between two applications Identify and discard old data segments from an earlier connection. Data transfer Flow and congestion control is done on a connection basis. Other desirable features: ordered, no loss, no duplicate. Disconnection Do it in a graceful manner so that segments are not transmitted when there is no one to receive it. User Datagram Protocol (UDP): Just send 3

4 TCP Header Source Port Destination Port Header Length Sequence Number Acknowledgment Number Reserved U A P R S F Window size Checksum Urgent Pointer H e a d e r Options Padding Data U: URG (Urgent) A: ACK P: PSH (Push) R: RST (Reset) S: SYN (Sync.) F: FIN (Finish) 4

5 TCP: Application Context Client Server Read/Write Port TCP Connection Read/Write Port TCP IP/DLC/ MAC/PHY Internet IP/DLC/ MAC/PHY Ports - Reserved for well-known services - Telnet/23, SMTP/25, FTP/20,21, HTTP/80, BGP/179, RIP/520, DNS/53, lp/515 - Free ports (allocated by the OS) 5

6 Source/destination Ports TCP: Header Port: A 16 bit local unique number on the host <= OS Port + Host IP => Unique end point of an application (Src Port + IP, Dst Port + IP): Unique connection ID Source and destination IP: NOT part of a TCP segment 32-bit seq. number SYN = 0 (DATA segment) Position of the first data byte of this segment in the sender s data stream SYN = 1 ISN to be used in the sender s byte stream. (in fact, ISN+1) Different each time a host requests a connection 6

7 TCP: Header 32-bit ACK number Valid if ACK = 1 Identifies the sequence number of the NEXT data byte that the sender of the ACK expects to receive. Header length in 4-byte units Lets the receiver know the beginning of the data area due to the variable length of the Option field. Reserved (6 bits) For future use. All 0 s. 7

8 TCP: Header URG: 1 => Urgent Pointer is valid ACK: 1 => ACK Seq# is valid PSH: 1 : The receiving TCP module passes the data to the application immediately 0 : The receiving TCP module may delay the data RST: 1 => Tells the receiver to abort the conn. SYN: This bit requests a connection FIN 1 : Sender has no more data to send, but is ready to receive. 8

9 Window Size TCP: Header The number of bytes the sender is willing to receive. Used in flow control and congestion control Checksum: For error detection Urgent Pointer: Valid if URG = 1 Urgent data Start byte is not specified, but it is considered to be the start of the seg. Final byte in receiver s buffer: Seq# + Urgent Ptr. The sender can send control information to the receiver to be processed on a priority basis. 9

10 Options MSS TCP: Header The Max Segment Size accepted by the sender Specified during connection set up Window Scale Allows the use of a larger advertised Window Size TimeStamp Intended to be used on high-speed connection Sequence number may wrap around during a connection. New segments are distinguished from old segments. Also used in Round-Trip Time (RTT) calculation 10

11 TCP Connection: General TCP connection A short- or long-term association between two apps. Comm params are exchanged before data segments: ISN Receive Window (RWND) Max Segment Size (MSS) Start of a connection is known to both the parties so that an old (terminated) connection has no impact. Bidirectional (Full-duplex) 11

12 TCP Conn.: Established in two ways Server Client Peer Peer Listen (Passive) Active Active Active Most common Possible mode The server must be running, and attached to a port known to the client. 12

13 TCP Connection: 3-way handshake Use the fields necessary to understand it Connection request (SYN) Sequence number Acknowledgement (ACK) Window size 13

14 TCP Connection: 3-way handshake Active open Client Seg(Seq# = 8000,SYN) Server Passive open Seg(Seq#=15000, Ack = 8001, SYN+ACK, RWND = 5000) Seg(Seq#=8000, Ack = 15001, ACK, RWND = 10000) 14

15 TCP Connection: 3-way handshake SYN segment from client to server» SYN = 1» A random initial Seq# (ISN)» RWND is undefined (defined later )» Options SYN segment from server to client SYN = 1 A random initial Seq# (ISN) ACK = 1 (servers acks the received SYN segment) Ack Seq.#: The sequence # of first data byte to be received RWND: Receive window size ACK from client to server ACKs the second SYN segment RWND 15

16 TCP: Connection Management State Diagram Timeout/RST LISTEN/ (Create TCB) SYN/ SYN, ACK RST/ CLOSED LISTEN CLOSE/ SEND/ SYN CONNECT/ (Create TCB) SYN CLOSE or Time-out or RST/ (Delete TCB) SYN_RCVD SYN/ SYN, ACK SYN_SENT CLOSE/ FIN CLOSE/ FIN ACK/ ESTABLISHED FIN/ ACK SYN,ACK/ ACK CLOSE_WAIT FIN_WAIT1 ACK/ FIN/ ACK CLOSING FIN,ACK/ ACK ACK/ CLOSE/ FIN LAST_ACK ACK/ FIN_WAIT2 FIN/ ACK TIME_WAIT 2MSL Time-out/ (Delete TCB) 16

17 Client/Server Communication and State Transitions 17

18 Client Active open Active close 2MSL timer Closed SYN SENT Established FIN WAIT-1 FIN WAIT-2 Data TCP Operation SYN SYN+ACK Transfer ACK FIN ACK FIN Closed LISTEN SYN RCVD Established CLOSE WAIT LAST ACK Server Passive open Inform app. Passive close Client states TIME WAIT Closed ACK Closed Server states 18

19 TCP: Flow Control FC: Regulates the amount of data a source can send before receiving an ACK. Using a Sliding Window Protocol The bytes within the window are the bytes that can be in transit. The (sender s) window is opened/ closed. 19

20 TCP: Flow Control Window Size = min(rwnd, CWND) RWND: Receiver s window The receiver sends this info to the sender in a segment There is a field for this in segment header. CWND: Congestion window Used for congestion control Managed by the sender 20

21 TCP: Flow Control Silly Window Syndrome: TCP/IP header = 40 bytes (#of data bytes/total segment length) is very low. Can occur if the sender and/or the receiver is very slow. Syndrome created by sender (Nagle s solution) Sender sends the first segment even if it is a small one. Next, the sender waits until» An ACK is received, OR» A maximum-size segment is accumulated. Before sending the next segment and repeat the next... Syndrome created by receiver Clark s solution: Send an ACK, and close the window until another segment can be received or buffer is ½ empty. Delayed ACK: at most 500 ms; 21

22 Mechanisms for detecting TCP: Error Control Corrupted segments, lost segments, out-of-order segments, duplicated segments Mechanisms for error detection and correction Checksum (header + data) ACK Timeout (a retransmission timer for each segment) 22

23 ACK Types Positive ACK TCP: ACK ACK (flag) = 1 ACK Sequence# => The expected sequence number Selective ACK There is no provision for SACK in TCP header Some implementations use an Option field 23

24 ACK Generation Rules When an in-order data segment is received, delay the ACK until Another data segment is received, OR 500 ms has elapsed. When an out of sequence segment with a higher sequence # arrives Send an ACK with the expected seq# Ask for fast retransmission: Send 3 ACKs. When a missing segment arrives, send an ACK to announce the next seq# expected. If a duplicate segment arrives, immediately send an ACK. 24

25 Central to error control Retransmission occurs TCP: Retransmission When a retransmission timer expires Sender starts a Retrans. Time-Out (RTO) timer for each segment sent (except for ACK segments) Three duplicate ACKs are received A mechanism for fast retransmission Useful when the receiver notices one missing segment, but the subsequent segments are just fine.. Note: Out-of-order segments are simply buffered. Earlier implementations simply dropped those. 25

26 TCP: Congestion Control : Host H Internet (Net of routers) H Total Output rate No congestion Network capacity congestion H : H Network input Network output Total Input rate Too many packets are sent in Congestion 26

27 TCP: Causes of congestion Packets arriving on different input links want to go out on the same output link Queue builds up for the outgoing link. Router starts dropping packets. Slow routers Queues build up if computing tasks take too much time. Queuing buffers, updating tables, running routing protocols 27

28 General Principles of CC Static decisions - Decide when to accept new traffic - Decide when to discard packets (Congestion prevention policy) Dynamic decisions (in 3 parts) - Monitor the system to know when and where congestion occur. - Pass on this information to where action can be taken. - Adjust system operation to correct the problem. 28

29 Congestion Control Dynamic decision A variety of metrics can be used to monitor a system. Fraction of all packets discarded due to lack of buffer Average queue length Number of retransmitted packets Average packet delay Dissemination of congestion information A field can be reserved in packet header to carry this info. Hosts and routers can send probe packets to enquire. Flow adjustment Deny service to some users. Degrade service to some users. Have users schedule their demand in a more predictable manner. 29

30 Congestion Control Congestion Prevention Policies DLC level Don t discard out-of-sequence packets. Selective-Repeat is better than Go-back-N. May not use a separate packet to ACK (use piggyback). Network level Spread traffic over many paths. Use a good discard policy File transfer: Drop new packets Real-time: Drop old packets TCP level. Next 30

31 TCP: Congestion Control Achieved by putting one more condition for FC Actual Window Size = min (RWND, CWND) Main idea Slow start but quickly speed up to a threshold Congestion avoidance beyond threshold, increase linearly Congestion detection Go back to slow start. 31

32 TCP: Congestion Control Slow start Initially, CW = 1: Transmit 1 segment (MSS) If ACK received before TO CW = 2 (= CW x 2): Transmit 2 segments (MSS) If ACKs received before TO CW = 4 (= CW x 2): Transmit 4 segments (MSS) If ACKs received before TO CW = 8 (= CW x 2): Transmit 8 segments (MSS) : Continue until you hit a threshold: Congestion Threshold (CT) Normally, CT = 64 KBytes Congestion Avoidance: Additive Inc. Each time the whole window of segments is ACKed CW = CW + 1 CWmax = RWND Congestion Detection RTO timer goes off 3 copies of an ACK are received Update CT and CW RTO timer goes off CT = CW/2 and CW = 1 3 ACKs received CT = CW/2 and CW = CT 32

33 TCP: Congestion Control Example: SS-AIMD CW Time 33

34 Four kinds of timers TCP: Timers Retransmission Time-Out (RTO) timer Persistence timer Keepalive timer TIME-WAIT timer 34

35 Operation TCP: Timers (RTO)» For each segment transmitted (except ACK), start an RTO» If RTO goes off, retransmit the segment and restart RTO» If ACK is received before the RTO goes off, kill RTO RTT S (RTT Smoothed) After first measurement RTT S = RTT M After another measurement RTT S = (1 α )RTT S + α.rtt M RTT D (RTT Deviation) RTO After first measurement RTT D = RTT M /2 After another measurement RTT D = (1 β )RTT D + β. RTT S RTT M Original Initial value After a measurement RTO = RTT S + 4. RTT D 35

36 Problem Solution TCP: Timers (Persistence) A receiver can close the sender s window and reopen it with an ACK If the ACK is lost, there is deadlock. When a sending TCP receives a segment with RWND = 0, start a persistence timer. Persistence timer goes off: Send a probe segment (1 byte data) to alert the receiver. Persistence timer value» Initially: Equal to RTO» Subsequently: Doubled with each retransmission of the probe.» Saturates at 60 sec. 36

37 TCP: Timers (Keepalive and TIME-WAIT) Keepalive Timer To sustain mostly idle connections (as between BGP routers) Each time the server hears from a client Reset the timer: Length = 2 hours. If the server does not hear from the client for two hours» Send a probe segment. If there is no response after 10 probes (75 sec apart)» Assume that the client is down. TIME-WAIT Timer (2.MSL) Used during connection termination. 37

38 OS Support for TCP-based Network I/O 38

39 Server s calls OS Support for TCP-based Network I/O sockfd = socket(protocol options, ) status = bind(sockfd, *myaddress, ) status = listen(sockfd, backlog)» Convert the socket to a passive socket; -1 for error confd = accept(socketfd, *clientaddress, )» Returns a connected socket for a client; -1 for error status = read(confd, *buf, len) Client s calls sockfd = socket(protocol options, ) status = connect(sockfd, *serveraddress, )) status = write(sockfd, *buf, len) 39

40 OS Support for TCP-based Network I/O Interested in network programming? UNIX Network Programming The Socket Networking API Vol. 1, 3 rd Edition W. Richard Stevens, et al. Addison Wesley 40