1. TCP Tutorial - Part III -
Internet Computing KUT (
Youn-Hee Han
It is licensed under a Creative Commons Attribution 2.5 License

2. TCP Flow Control
Computer Network

3. Sliding Window Revisited
Sending application
Receiving application
TCP
TCP
LastByteWritten
LastByteRead
LastByteAcked
LastByteSent
NextByteExpected
LastByteRcvd
(a)
(b)
Sending side
LastByteAcked ≤ LastByteSent
LastByteSent ≤ LastByteWritten
buffer bytes between LastByteAcked and LastByteWritten
Receiving side
LastByteRead < NextByteExpected
NextByteExpected ≤ LastByteRcvd + 1
buffer bytes between LastByteRead and LastByteRcvd
Computer Network

4. TCP Flow Control Send buffer size: SendBuffer
Receive buffer size: RcvBuffer
Receiving side
LastByteRcvd - LastByteRead ≤ RcvBuffer
AdvertisedWindow = RcvBuffer - (LastByteRcvd - LastByteRead)
Sending side
LastByteSent - LastByteAcked ≤ AdvertisedWindow
EffectiveWindow = AdvertisedWindow - (LastByteSent - LastByteAcked)
LastByteWritten - LastByteAcked ≤ SendBuffer
block sender if tries to write y bytes but (LastByteWritten - LastByteAcked) + y > SenderBuffer
Always send ACK in response to arriving data segment
Computer Network

5. TCP Flow Control TCP sender has a ‘SendBuffer’
LastByteSent - LastByteAcked ≤ AdvertisedWindow
block sender if tries to send the next segment but (LastByteSent - LastByteAcked) + thee next sgement size > AdvertisedWindow
LastByteWritten – LastByteAcked ≤ SendBuffer
block sender if tries to write y bytes but (LastByteWritten - LastByteAcked) + y > SenderBuffer
EffectiveWindow
LastByteWritten
AdvertisedWindow (from receiver)
Sent but not acked
Not yet sent
Sequence numbers
SendBuffer
LastByteAcked
LastByteSent
Computer Network

6. TCP Flow Control flow control TCP receiver has a “RcvBuffer”:
sender won’t overflow
receiver’s buffer by
transmitting too much,
too fast
flow control
TCP receiver has a “RcvBuffer”:
LastByteRead
RcvBuffer (possible window)
ACKed but not
delivered to user
recv’d
but not
ACKed
Sequence numbers
missing data
AdvertisedWindow
NextByteExpected
LastByteRcvd
Application process may be slow at reading from buffer
speed-matching service: matching the send rate to the receiving app’s drain rate
Receiver advertises spare room by including value of “window sizes” in segments
Computer Network

7. Window Advertisement Issues
Persist when AdvertisedWindow = 0
send a segment with 1 byte of data every so often which triggers a response that contains the current advertised window (eventually nonzero)
Flow Control can limit Throughput
Let rtt be the round-trip time, i.e., the time from sending a segment until an acknowledgement (ACK) is received
Let t = wnd/b be the time to transmit a full “window” of data, where b is link bandwidth
Throughput is wnd/rtt
Sender
Receiver t
rtt
wnd bytes
Computer Network

8. TCP Congestion Control
Computer Network

9. TCP Congestion Control
informally: “too many sources sending too much data too fast for network to handle”
different from flow control!
manifestations:
long delays (queueing in router buffers)
lost packets (buffer overflow at routers)
Computer Network

10. TCP Congestion Control
Flow Control vs. Congestion Control
Src
Dest
Limits amount of data that destination must buffer
Src
Dest
Attempts to reduce buffer overflow inside the network
Computer Network

11. Congestion Collapse 10 Mbps 1.5 Mbps
If both sources send full speed, the router is completely overwhelmed
congestion collapse: senders lose data from congestion and they resend, causing more congestion (can be self-reinforcing)
has been observed many times
Computer Network

12. Basic Rule Basic Rule End-to-end congestion control
IP layer provides no explicit feedback regarding network congestion
There is no (or little) congestion  TCP sender increases its send rate
There is congestion along the path  TCP sender reduces its send rate
Computer Network

13. Basic Rule Window Size Determination What does flow control do?
avoids overrunning the receiver
What does congestion control do?
avoid overrunning router buffers; avoid saturating the network
Congestion Window (=cwin)
It imposes a constraint on the rate at which a TCP sender can send traffic into the network
cwnd
It is what the network can handle
flow control window (wnd)
It is what the receiver can handle
Computer Network

14. Basic Rule Window Size Determination What mechanism do we use?
both use windows
[flow control] wnd (=AdvertisedWindow)
[congestion control] cwnd
actual window used is the MIN of wnd and cwnd
Dynamic Window Size (version 1.0) : Windows Size = min{cwin, wnd}
Computer Network

15. TCP Congestion Principals
underlying principle: packet conservation
at equilibrium, inject packet into network only when one is removed
basis for stability of physical systems
TCP components:
how to get there: slow start
how to stay there: congestion avoidance
Computer Network

16. TCP Congestion Principals
Additive-Increase, Multiplicative-Decrease (AIMD)
Approach: increase transmission rate until loss occurs
additive increase ( congestion avoidance phase)
increase cwin by 1 MSS every RTT until loss detected
continuously probing for usable bandwidth
multiplicative decrease
cut cwin in half after loss (4020105)
It is not allowed to drop below one MSS
Saw tooth
behavior: probing
for bandwidth
Computer Network

17. Slow Start Initialize cwnd = 1
Upon receipt of every ACK, cwnd = cwnd + 1
TCP sender maintains three new variables:
cwnd – congestion window
It is determined by inferred information about the level of congestion in the network.
ssthresh – slow-start threshold
It can be thought of as an estimate of the level below which congestion is not expected
Flight Size
No. of unacknowledged segments
Implications
Will overshoot window and cause packet loss
but remember, packet loss is part of the plan
Computer Network

18. Slow Start Slow Start When connection begins, cwin = 1 MSS
Example: MSS = 500 bytes & RTT = 200 msec
initial rate = 20 kbps
available bandwidth may be >> MSS/RTT
desirable to quickly ramp up to respectable rate
Host A
Host B
one segment
RTT
two segments
Slow Start algorithm
Delayed & Cumulative Ack
initialize: cwin = 1
for (each segment ACKed)
cwin++
until (loss event OR cwin >ssthresh)
four segments
time
Summary: initial rate is slow but ramps up exponentially fast
Exponential increase (per RTT) in window size
Computer Network

19. Slow Start Example (redrawn from [Jacobson88a] Fig 2) one RTT 0R 1 1R
one pkt time 1R 1 2 3 2R 2 3 4 6 5 7 3R 4 5 6 7 8 10 12 14 9 11 13 15
Computer Network

20. When “Slow Start” ends? Implementation:
Q: When should the exponential increase switch to linear?
(Want to end when the pipe is full)
A: do end when cwnd > ssthresh and move to “Congestion Avoidance phase”
SSTHRESH
Implementation:
Variable ssthresh
At loss event, ssthresh is set to 1/2 of Flight Size just before loss event
ssthresh = max {Flight Size/2, 2*MSS}
cwin=1
ssthresh is typically set to very large value on connection setup
ssthresh represents the rough estimation of the network capacity
Computer Network

21. Slow Start and Congestion Avoidance
State Transition
cwin instead set to 1 MSS;
window then grows exponentially
to ssthresh, then grows linearly
Congestion Avoidance
upon receiving ACK
Increase cwnd by 1/cwnd
This is additive increase (over 1 RTT it adds up to increasing by 1 segment)
why not multiplicative increase?
growing too fast in equilibrium => oscillations
Slow Start
Congestion Avoidance
SSTHRESH
SSTHRESH
Computer Network

22. Slow Start and Congestion Avoidance
Slow Start: aggressively grow congestion window until congestion is detected
Congestion Avoidance
Initially:
cwnd = 1*MSS (Maximum Segment Size)
ssthresh is very large.
If no loss:
cwnd = 2*cwnd (after each new ACK) (This gives exponential growth of cwnd)
If loss (timeout):
ssthresh = max{flight size/2, 2*MSS} (multiplicative decrease)
cwnd = 1*MSS
If no loss:
increase cwnd at most 1*MSS per RTT (additive increase)
If loss:
ssthresh = max{flight size/2, 2*MSS} (multiplicative decrease)
cwnd = 1*MSS.
Computer Network

23. Slow Start and Congestion Avoidance
Initally:
- cwnd = 1*MSS
- ssthresh = very high (64kbytes)
If a new ACK comes:
- if cwnd < ssthresh  update cwnd according to slow start
if cwnd  ssthresh  update cwnd according to congestion avoidance
If timeout (i.e. loss) :
- ssthresh = max{flight size/2, 2*MSS} ;
cwnd
(initial) ssthresh
Loss, e.g. timeout
ssthresh
time
slow start – in green
congestion avoidance – in blue
Computer Network

24. Slow Start and Congestion Avoidance
Computer Network

25. Jacobson88a Figure 3: No CC of Slow-Start
Computer Network

26. Jacobson88a Figure 4: with CC of Slow-Start
Computer Network

27. Problems So Far have way to fill pipe (slow start)
have way to run at equilibrium (congestion avoidance)
but tough transition
no good initial ssthresh
large ssthresh causes packet loss, every time
need approaches to quickly recover from packet loss
Computer Network

28. Fallback Mechanism: Timeout
retransmission timer (RTO)
if no ACK after RTO fires, reset cwnd=1 and resend lowest unACK’ed segment
but they’re very crude
force slow-start again
coarse-grain TCP timeouts lead to idle periods
are often slow—a long time with no traffic
…but is there more information?
Computer Network

29. Could We Do Better?
Receipt of dupACKs tells the sender that the receiver is still getting new segments, i.e. there is still data flowing between sender and receiver
Why does sender go back to slow start always?
Computer Network

30. Fast Retransmit & Fast Recovery
allows and uses duplicate ACKs to trigger retransmission
Reaction to 3 dup ACKs for Segment N
i.e. 4 identical ACKs without the arrival of any other intervening packets,
Retransmits the segment N without waiting for timeout
And move to Fast Recovery phase
Fast Retransmit
Philosophy:
3 dup ACKs indicates network capable of delivering some segments
timeout indicates a “more alarming” congestion scenario
Computer Network

31. Fast Retransmit & Fast Recovery
a duplicate ACK implies the receiver got a packet out of order
an earlier packet might have been lost (or delayed)
TCP sender waits until it sees three duplicate ACKs before retransmitting a packet
Computer Network

32. Fast Retransmit & Fast Recovery
SSTHRESH is cut in half of cwnd
ssthresh = max {flight size/2, 2*MSS}
Set cwnd as follows:
cwnd = ssthresh + 3
window then grows linearly
cwnd++
Additive increase for every subsequent dup ACK
Transmit a segment if permitted by cwnd
When a new ACK received
Exit Fast Recovery phase
Reset cwnd to ssthresh (resume congestion avoidance)
When Timeout occurs  perform Slow-Start
Fast Recovery
This artificially "inflates" the congestion window by the number of segments (three) that have been sent and which the receiver has buffered.
It will acknowledge receiving all segments sent till moving to Fast Recovery phase (assuming that no more segments were lost).
in order to reflect the additional segment that has been sent and which the receiver has buffered
Computer Network

33. Fast Retransmit & Fast Recovery
cwnd
(initial) ssthresh
fast-retransmit
fast-retransmit
timeout
new ACK
new ACK
Time
Slow Start
Congestion Avoidance
“inflating” cwnd with dupACKs
roughly a 20% improvement in the throughput
“deflating” cwnd with a new ACK
Computer Network

34. TCP Congestion Control - Summary
When cwin is below SSTHRESH,
sender in slow-start phase, window grows exponentially.
When cwin is above SSTHRESH,
sender is in congestion-avoidance phase, window grows linearly.
When a triple duplicate ACK occurs,
Immediately retransmit the packet
 named Fast Retransmit
ssthresh set to max {0.5 * flight size, 2*MSS} and cwin set to ssthresh+3.
 named Fast Recovery
When timeout occurs,
ssthresh set to cwin/2 and cwin is set to 1 MSS.
“Fast” – because it doesn’t wait for time out when not getting an ACK for a segment
Computer Network