TCP "Real" Reliable Transport

Transcription

1 TCP "Real" Reliable Transport CS 356 University of Texas at Austin Dr. David A. Bryan VERY SIGNIFICANT content used or adapted from Computer Networking: A Top- Down Approach, 6e, Kurose and Ross, Addisson- Wesley, or the supporting instructor slides Kurose and Ross or Addison- Wesley, All rights reserved. Remaining content by David A. Bryan unless otherwise noted.

2 Remember from last time We saw UDP, which provided no reliability Then we looked at a simplified model of how we would make this reliable (modeled using state machines) Used Ack/No Ack with sequence numbers Then moved to Ack/Ack previous sequence numbers Timers when messages not received Goal was to understand how to build a real reliable transport TCP So how are the primitives (pieces) above used in TCP?

3 TCP: Overview RFCs: 793,1122,1323, 2018, 2581 point-to-point: one sender, one receiver reliable, in-order byte steam: app sees no "message boundaries" connection-oriented: handshake first to establish the connection pipelined: We don't wait for everything to be acknowledged send multiple packets full duplex data: bi-directional data flow in same connection

4 Pipelining We talked about "stop and wait" send one, wait for ack With pipelining, we send more, some number "in flight" at any time: sender receiver sender receiver The number in flight (3 here) is called the window size

5 Half- Duplex vs. Full- Duplex Half Duplex What we saw before one side (A) sends, the other (B) acks If (B) wants to send it waits and the sends after (A) TCP Full Duplex Both sides send at same time data sent in the acks (A) sends data to (B) (B) acks it and sends back at same time. A B A B send data rcv ack rcv data send ack data ack data ack rcv data send ack send data rcv ack send data rcv ack w/data send ack w/data rcv ack w/data data ack w/data ack w/data ack w/ data rcv data send ack w/data rcv ack w/data send ack w/data

6 UDP: segment header 32 bits source port # dest port # length application data (payload) checksum UDP segment format UDP was missing fields for a number of the things we were adding it had no reliability, so simply needed source/dest, length, and the checksum All we could do was discard bad packets No sequence numbers, for example

7 TCP segment structure head len 32 bits source port # dest port # sequence number acknowledgement number not used UAP R S F checksum receive window Urg data pointer options (variable length) application data (variable length) LOTs more in a TCP header As before, lots wrapping the actual data (what we saw in our code) We'll look at some of these. First, notice there is a sequence number and an acknowledgement number

8 Reality: Sequence Numbers in TCP We saw a simple example where we alternated between sequence numbers 0/1, but that made our example a stop and wait approach TCP is not stop- and- wait, so with pipelining we send many (with incrementing seq. number), and get many acks (with corresponding seq. number) Again, some number (window size) of packets "in flight" at a time to track Extra wrinkle TCP is full- duplex so sending back data at same time!

9 The Fields Used ACK flag head len 32 bits source port # dest port # sequence number acknowledgement number not used UAP R S F checksum receive window Urg data pointer options (variable length) counting by bytes of data (not segments!) # bytes rcvr willing to accept application data (variable length)

10 Sequence Numbers in TCP TCP presents the data to the application like it was a stream of bytes not a set of packets App doesn't care if 5000 bytes of data was 10 packets of 500 bytes or 500 packets of 10 bytes TCP is also in order the application doesn't get the data out of order. If 1,2,3,5 is received, app gets 1,2,3 then waits. 5 isn't delivered until 4 is retransmitted. To make this easier, TCP sequence (and ack) numbers are actually the # of data bytes sent, not # of packets. Don't start at zero why? Exchanged at startup

11 TCP seq. numbers, ACKs sequence numbers: byte stream number of first byte in segment's data acknowledgements: seq # of next byte expected from other side cumulative ACK Q: how receiver handles outof-order segments A: TCP spec doesn't say, - up to implementor outgoing segment from sender source port # dest port # sequence number acknowledgement number rwnd checksum sent ACKed urg pointer sender sequence number space sent, notyet ACKed sent not yet ( inflight ) incoming segment to sender source port # dest port # sequence number acknowledgement number A rwnd checksum urg pointer

12 TCP seq. numbers, ACKs Host A Host B User types C host ACKs receipt of echoed C Seq=42, ACK=79, data = c Seq=79, ACK=43, data = C Seq=43, ACK=80 host ACKs receipt of c, echoes back C Example: Capitalize server (just sends back what you sent capitalized)

13 Timer Remember that sequence numbers let us tell if a packet was corrupt (we don't ack the bad packet) and here if we are sending fast enough, can also tell we have a missing packet How? If sending slow, I never got an ack, we also needed to add a timer to know when to retransmit that packet (since we assume it wasn't received) Problem was picking the timer Takes longer for a packet to get to China than Dallas So again, reality TCP has an elegant way to calculate the timer

14 TCP round trip time, timeout Q: how to set TCP timeout value? longer than RTT but RTT varies (China vs. UT, slower connection, more congestion ) too short: premature timeout, unnecessary retransmissions too long: slow reaction to loss Q: how to estimate RTT? SampleRTT:measure the time from segment transmission until ACK receipt ignore retransmissions SampleRTTwill vary, want estimated RTT smoother average several recent measurements, not just current SampleRTT average: EstimatedRTT

15 TCP round trip time, timeout EstimatedRTT = ((1- α)* EstimatedRTT) + (α * SampleRTT) exponential weighted moving average influence of past sample decreases exponentially fast typical value: α = (⅛) 350 RTT: gaia.cs.umass.edu to fantasia.eurecom.fr 300 RTT (milliseconds) samplertt EstimatedRTT time (seconds)

16 TCP round trip time, timeout EstimatedRTT = (⅞ * EstimatedRTT) + (⅛ * SampleRTT) exponential weighted moving average influence of past sample decreases exponentially fast typical value: α = RTT: gaia.cs.umass.edu to fantasia.eurecom.fr 300 RTT (milliseconds) samplertt EstimatedRTT time (seconds)

17 TCP round trip time, timeout timeout interval: Use EstimatedRTT plus safety margin large variation in EstimatedRTT è larger safety margin estimate SampleRTT deviation from EstimatedRTT: DevRTT = (1-β) * DevRTT + β * SampleRTT-EstimatedRTT (typically, β = 0.25) TimeoutInterval = EstimatedRTT + 4*DevRTT estimated RTT safety margin

18 TCP sender events: if data rcvd from app: create segment with seq # seq # is byte- stream number of first data byte in segment start timer if not already running think of timer as for oldest unacked segment expiration interval: TimeOutInterval if timeout: retransmit segment that caused timeout restart timer if ack rcvd: if ack acknowledges previously unacked segments update what is known to be ACKed start timer if there are still unacked segments

19 TCP ACK generation [RFC 1122, RFC 2581] event at receiver arrival of in-order segment with expected seq #. All data up to expected seq # already ACKed arrival of in-order segment with expected seq #. One other segment has ACK pending arrival of out-of-order segment higher-than-expect seq. #. Gap detected arrival of segment that partially or completely fills gap TCP receiver action delayed ACK. Wait up to 500ms for next segment. If no next segment, send ACK immediately send single cumulative ACK, ACKing both in-order segments immediately send duplicate ACK, indicating seq. # of next expected byte immediate send ACK, provided that segment starts at lower end of gap

20 TCP: retransmission scenarios Host A Host B Host A Host B Seq=92, 8 bytes of data SendBase=92 Seq=92, 8 bytes of data timeout X ACK=100 timeout Seq=100, 20 bytes of data ACK=100 ACK=120 Seq=92, 8 bytes of data ACK=100 lost ACK scenario SendBase=100 SendBase=120 SendBase=120 Seq=92, 8 bytes of data ACK=120 premature timeout

21 TCP: retransmission scenarios Host A Host B Seq=92, 8 bytes of data Seq=100, 20 bytes of data X ACK=100 ACK=120 Seq=120, 15 bytes of data (don't need to resend the 100 data!) Cumulative ACK

22 TCP fast retransmit time- out period often relatively long: long delay before resending lost packet so, detect lost segments via duplicate ACKs. sender often sends many segments back- to- back if segment is lost, there will likely be many duplicate ACKs. TCP fast retransmit if sender receives 3 ACKs for same data ( triple duplicate ACKs ), resend unacked segment with smallest seq # likely that unacked segment lost, so don't wait for timeout

23 TCP fast retransmit Host A Host B Seq=92, 8 bytes of data Seq=100, 20 bytes of data X ACK=100 timeout ACK=100 ACK=100 ACK=100 Seq=100, 20 bytes of data fast retransmit after sender receipt of triple duplicate ACK

24 TCP Reliability Combination of timers, acknowledgement, retransmissions ensures reliable information Key points Full- Duplex data and acks sent together Sequence numbers are count of data, not packets Ack the last received in order not necessarily last received Timeout adapts and based on performance of connection In order delivery (1,2,3,5 means application gets 1,2,3 until 4 is delivered )

25 TCP Flow and Congestion Control In addition, TCP provides control for two things: Flow control don't overload the receiver Congestion control don't overload the network

26 TCP flow control application may remove data from TCP socket buffers. slower than TCP receiver is delivering (sender is sending) application process TCP socket receiver buffers TCP code application OS flow control receiver controls sender, so sender won't overflow receiver's buffer by transmitting too much, too fast from sender IP code receiver protocol stack

27 TCP flow control receiver advertises free buffer space by including rwnd value in TCP header of receiver-to-sender segments RcvBuffer size set via socket options (typical default is 4096 bytes) many operating systems autoadjust RcvBuffer sender limits amount of unacked ( in-flight ) data to receiver's rwnd value guarantees receive buffer will not overflow RcvBuffer rwnd to application process buffered data free buffer space TCP segment payloads receiver-side buffering

28 TCP segment structure 32 bits head len source port # dest port # sequence number acknowledgement number not used UAP R S F checksum receive window Urg data pointer options (variable length) rwnd, # bytes rcvr willing to accept application data (variable length)

29 Connection Management before exchanging data, sender/receiver handshake : agree to establish connection (each knowing the other willing to establish connection) agree on connection parameters (sequence numbers, rwnd) application application connection state: ESTAB connection variables: seq # client-to-server server-to-client rcvbuffer size at server,client network connection state: ESTAB connection Variables: seq # client-to-server server-to-client rcvbuffer size at server,client network Socket clientsocket = newsocket("hostname","port number"); Socket connectionsocket = welcomesocket.accept();

30 TCP segment structure 32 bits head len source port # dest port # sequence number acknowledgement number not used UAP R S F checksum receive window Urg data pointer options (variable length) SYN, FIN and RST bits are used for connection management application data (variable length)

31 TCP: opening a connection client, server establishes connection with message with SYN bit send TCP segment with SYN bit = 1 Inlcude sequence number client starts with respond to received SYN with ACK on receiving SYN, ACK can be combined with own SYN Sequence number + 1 for Ack Number Send own starting sequence number If the server isn't listening on that port, replies with RST bit set (doesn't accept connection)

32 TCP 3- way handshake client state LISTEN SYNSENT ESTAB choose init seq num, x send TCP SYN msg received SYNACK(x) indicates server is live;; send ACK for SYNACK;; this segment may contain client-to-server data SYNbit=1, Seq=x SYNbit=1, Seq=y ACKbit=1;; ACKnum=x+1 ACKbit=1, ACKnum=y+1 choose init seq num, y send TCP SYNACK msg, acking SYN received ACK(y) indicates client is live server state LISTEN SYN RCVD ESTAB

33 TCP: closing a connection client, server each close their side of connection send TCP segment with FIN bit = 1 respond to received FIN with ACK on receiving FIN, ACK can be combined with own FIN simultaneous FIN exchanges can be handled

34 TCP: closing a connection client state server state ESTAB ESTAB clientsocket.close() FIN_WAIT_1 FIN_WAIT_2 can no longer send but can receive data wait for server close FINbit=1, seq=x ACKbit=1;; ACKnum=x+1 can still send data CLOSE_WAIT TIMED_WAIT timed wait for 2*max segment lifetime FINbit=1, seq=y ACKbit=1;; ACKnum=y+1 can no longer send data LAST_ACK CLOSED CLOSED Transport Layer 3-40

35 Congestion Control congestion: informally: too many sources sending too much data too fast for network to handle different from flow control! manifestations: lost packets (buffer overflow at routers) long delays (queueing in router buffers) Let's look at some scenarios

36 Causes/costs of congestion: scenario 1 two senders, two receivers one router, infinite buffers output link capacity: R no retransmission Host B original data: λ in Host A unlimited shared output link buffers throughput: λ out R R/2 λ out delay λ in R/2 λ in R/2 maximum per-connection throughput: R/2 v large delays as arrival rate, λ in, approaches capacity Try to send more (λ in é ), but λ out can't increase, so delay!

37 Causes/costs of congestion: scenario 2 one router, finite buffers sender retransmission of timed-out (dropped) packet application-layer input = application-layer output: λ in = λ out transport-layer input includes retransmissions : λ in λ in λ in : original data λ'in : original data, plus retransmitted data λ out Host A Host B finite shared output link buffers

38 Result : scenario 2 idealization: perfect knowledge (magically know buffer state) sender sends only when router buffers available R/2 λ out λ in R/2 λ in : original data λ'in : original data, plus retransmitted data λ out A Host B finite shared output link buffers

39 Result : scenario 2 Realistic: duplicates v v packets can be lost, dropped at router due to full buffers sender times out prematurely, sending two copies, both of which are delivered R/2 λ out λ in R/2 when sending at R/2, some packets are retransmissions including duplicated that are delivered! timeout copy λ in (λ' in ) λ out A Host B

40 Result : scenario 2 Realistic: duplicates packets can be lost, dropped at router due to full buffers sender times out prematurely, sending two copies, both of which are delivered λ out R/2 λ in R/2 when sending at R/2, some packets are retransmissions including duplicated that are delivered! costs of congestion: more work (retrans) for given "goodput" unneeded retransmissions: link carries multiple copies of pkt decreasing "goodput" You don't really want to send more (λ in é ) after you hit R/2. How to detect?

41 Approaches towards congestion control two broad approaches towards congestion control: network-assisted congestion control: routers provide feedback to end systems Much more required of network Some (less popular) tech tried this end-end congestion control: no explicit feedback from network congestion inferred from end-system observed loss, delay end clients try to play nice! approach taken by TCP

42 TCP Approach to Congestion Control Very interesting (and famous) algorithm to do this We have to understand a few things for this: MSS Maximum Segment Size the biggest amount of data we can send on the connection Packet size - headers 1460 bytes is common Congestion Window (cwnd) Remember: Window size is how many packets can we have in flight at same time. If window size is 4 we can have four out at once. THIS IS A SECOND WINDOW! cwnd and rwnd! One stops overflow of receiver, one for network congestion

43 TCP congestion control:aimd additive increase multiplicative decrease approach: sender increases transmission rate (window size), probing for usable bandwidth, until loss occurs additive increase: increase cwnd by 1 MSS every RTT until loss detected multiplicative decrease: cut cwnd in half after loss AIMD saw tooth behavior: probing for bandwidth cwnd: TCP sender congestion window size additively increase window size. until loss occurs (then cut window in half) time

44 TCP Congestion Control: details sender sequence number space cwnd last byte ACKed last byte sent sender limits transmission: sent, notyet ACKed ( inflight ) LastByteSent- LastByteAcked < cwnd cwnd is dynamic, function of perceived network congestion Also has to be rwnd! TCP sending rate: roughly: send cwnd bytes, wait RTT for ACKS, then send more bytes rate ~ cwnd RTT bytes/sec

45 TCP Slow Start Bad name, actually a way to ramp up faster when connection begins, increase rate exponentially until first loss event: initially cwnd = 1 MSS double cwnd every RTT done by incrementing cwnd for every ACK received initial rate is slow but ramps up exponentially fast Backs off (to 1 again) when you hit congestion Now remember ½ that level as a threshold Host A RTT Host B time

46 TCP: detecting, reacting to loss Different algorithms TCP Tahoe, and newer TCP Reno Some even newer For loss indicated by timeout (TCP Reno or Tahoe): cwnd set to 1 MSS; window then grows exponentially (as in slow start) to threshold (½ of what we got to last time), then grows linearly For loss indicated by 3 duplicate ACKS: See above for TCP Tahoe (same thing) For Reno: dup ACKs indicate network capable of delivering some segments cwnd is cut in half (not to 1) then grows linearly

47 TCP: switching from slow start to CA Q: when should the exponential increase switch to linear? A: when cwnd gets to 1/2 of its value before timeout. Implementation: variable ssthresh on loss event, ssthresh is set to 1/2 of cwnd just before loss event

48 TCP Fairness fairness goal: if K TCP sessions share same bottleneck link of bandwidth R, each should have average rate of R/K TCP connection 1 TCP connection 2 bottleneck router capacity R

49 Why is TCP fair? two competing sessions: additive increase gives slope of 1, as throughout increases multiplicative decrease decreases throughput proportionally R equal bandwidth share loss: decrease window by factor of 2 congestion avoidance: additive increase loss: decrease window by factor of 2 congestion avoidance: additive increase Connection 1 throughput R

50 Fairness (or being less fair ) Fairness and UDP multimedia apps often do not use TCP do not want rate throttled by congestion control instead use UDP: send audio/video at constant rate, tolerate packet loss Fairness, parallel TCP connections application can open multiple parallel connections between two hosts web browsers do this e.g., link of rate R with 9 existing connections: new app asks for 1 TCP, gets rate R/10 new app asks for 11 TCPs, gets R/2