TCP Congestion Control

1

• TCP Congestion Control

2
TCP Congestion Control
Strategy Send packets into the network without
reservations and then react to observable events
that occur.
3
End-to-End Congestion Control

• Distributed algorithm there is no central
  observer monitoring whether packets are being
  dropped by routers. Sources should get feedback
  from endpoints.
• A source sends a packet out. When it receives an
  ACK, it concludes the packet arrived at the
  destination (and has thus left the network). At
  this point, it should be safe to insert a new
  packet.

4
Additive Increase/Multiplicative Decrease

• TCP is connection oriented for each open
  connection, define a state variable
  CongestionWindow.
• The source uses this variable to determine how
  much data it can have in transit at any given
  time.
• The maximum number of unACKed bytes is the
  minimum of AdvertisedWindow and CongestionWindow.

• A TCP source can send no faster than the slowest
  component (network or destination host) can
  accommodate.

5
Additive Increase/Multiplicative Decrease

• AdvertisedWindow is set by the receiving side of
  a connection.
• CongestionWindow should be decreased when
  congestion is higher, and increased when
  congestion is lower.
• Use timeouts to sense congestion for every
  timeout observed, halve the CongestionWindow
  (multiplicative decrease). The variable is never
  allowed to fall below on MSS.
• For every ACKed packet, TCP adds one MSS to
  CongestionWindow

6
The Evolution of CongestionWindow

• Note that the algorithm states that the source is
  willing to decrease the window at a much faster
  rate than it is willing to increase it.

Question Why should the algorithm be more
conservative with increases than with decreases?
7
A Fairy Tale
TCP

• Once upon a time
• There was a little transport protocol called TCP.
  She was always eager to make the best use of the
  available bandwidth. When a connection was
  started, she tried to send as many packets as the
  AdvertisedWindow allowed. Things didnt go very
  well for TCP after that

BIG BAD PACKET LOSS WOLF
8
Factoid

• TCP is used over connections with bandwidth
  ranging from 9600 bps to 2.4 Gbps.

9
Slow Start

• Set CongestionWindow to 1. Send as many packets
  as indicated in CongestionWindow.
• If all packets sent are ACKed, then double
  CongestionWindow.
• Send as many packets as indicated in
  CongestionWindow.
• If all packets sent are ACKed, then double
  CongestionWindow.
• Send as many packets as indicated in
  CongestionWindow.
• If all packets sent are ACKed, then double
  CongestionWindow.

16
CongestionWindow
8
4
2
1
time
10
Packets in Transit
Additive Increase
Slow Start
11
Ok but where does it stop?

• Eventually, one or more packets are not ACKed
  because somewhere along the way they got dropped.
• If the connection times our waiting for an ACK,
  halve the CongestionWindow.
• The last value of used before there was packet
  loss is the target congestion window. Since
  this is useful information, we save it in a
  variable called CongestionThreshold.
• Reset CongestionWindow to one packet and start
  over. For every ACK received, increment
  CongestionWindow by one until CongestionThreshold.
  
• After that, increment CongestionWindow slowly, by
  one packet for every elapsed RTT.

12
TCP Congestion Control in action
Timeout CW halved, CT set
Packet loss no new packets sent
Linear ramp up
Exponential ramp up
Exponential ramp up til CT
Linear ramp up
Packet loss no new packets sent
13
Fast Retransmit

• Implementation detail TCP timers were too
  coarse-grained.
• Consequence Long periods of time in which the
  connection would go dead waiting for a timer to
  expire.
• Heuristic Sometimes it may be beneficial to
  trigger a retransmit before a timer expires.
• Fast retransmit Every time a packet arrives at
  the receiver, ACK it, even if this SeqNo has
  already been ACKed. Clearly, this will lead to
  duplicate ACKs. The sender uses this information
  as a hint that some packet has been lost. TCP
  waits for three duplicate ACKs before
  retransmitting a packet.

14
Fast Retransmit in action
15
Fast Recovery

• Fast retransmit may signal congestion. When that
  happens, instead of dropping the CongestionWindow
  down to one packet and run slow start, use the
  ACKs in transit to trigger the sending of new
  packets.