Reliability and Congestion Control in TCP

1
Reliability and Congestion Control in TCP

• Recall TCP/IP (TCP over the Internet Protocol)
  and UDP/IP.
• The Internet Protocol is a packet delivery system
  with the following characteristics
• Unreliable
• Data delivery not guaranteed.
• Packet can be lost, duplicated, delayed, or
  delivered out of order.
• Does not inform sender or receiver of any
  delivery problems.

2
IP Data Transfer Mechanism

• Connectionless
• Each packet treated independently from all others
  (even between same machines).
• A sequence of packets between the same two
  machines may take different routes, and some may
  arrive while others do not.
• Best-Effort Service
• Will try to deliver packets.
• Fails primarily due to insufficient resources or
  network failure.

3
So What Does IP Provide?

• Provides routing function.
• Determines path over which data will be sent.
• Each IP packet consists of two parts
• Header Contains information related to the
  protocol such as the IP address of sender and
  receiver.
• Data Payload. Information being exchanged.
• Each host has routing table.
• Consists of IP address (well, actually network
  prefix) and next hop.
• Destination IP address looked up in table. Packet
  sent out to next host in path.

4
IP Routing

• Each host in path from sender to receiver simply
  determines the next hop the packet will take.
• Example routing table for some host on Internet
  directly connected to networks 221.0.0.0,
  201.0.0.0, 122.0.0.0.

To reach host on network
Send to network
247.0.0.0 221.0.0.0 242.0.0.0 201.0.0.0 221.0.0.
0 122.0.0.0
All packets received heading for network
247.0.0.0 send to network 221.0.0.0 (which will
determine next hop for packet).
5
TCP/IP

• TCP uses IP delivery and routing services.
• Builds reliability on top of IP services.
• Recall TCP guarantees
• All packets delivered in same order sent.
• No duplicate data.
• All sent data will eventually arrive (barring
  catastrophic network failure).
• How reliability provided?

6
Basic Approach Positive Acknowledgements with
Retransmissions

• Sender keeps record of each packet sent.
• Sets timer when packet sent.
• Waits for acknowledgement before sending next
  packet.
• Retransmits packet if acknowledgement does not
  arrive before timer expires.

7
Timing Diagram
Receive Packet 2 Send Ack 2
8
Timing Diagram
Sender
Receiver
Network
Packet Lost
Send Packet 1 Start Timer
Packet Should Arrive and Ack Sent
Ack should arrive. Timer Expires
Retransmit Packet 1 Start Timer
Receive Packet 1 Send Ack 1
9
Duplicates

• Each packet has sequence number
• Receiving TCP remembers sequence numbers.
• Discards duplicates.
• Includes Sequence number in acknowledgements.

10
Sliding Windows

• Simple scheme (send one packet, wait for
  acknowledgement, send next ..) is very
  inefficient.
• Wastes significant bandwidth while waiting.
• Would like to send out more packets before having
  to block and wait for response.
• Developed idea of sliding window, where all
  packets within window can be unacknowledged.
• As acknowledgements come in, window is adjusted.

11
Sliding Window
12
0 1 2 3 4 5 6
7
13
Sliding Transmission Window in TCP

• TCP uses a sliding window mechanism to control
  transmission.
• Increases efficiency
• Send all packets within window without waiting
  for an acknowledgement.
• As acknowledgments for segments come in, the
  window is moved.
• Example Consider a transmission window size of
  five (next slide).

14
TCP Retains Three Pointers
0 1 2 3 4 5 6
7
A B C
15
Every packet to the left of A has been sent and
acknowledged. Every packet between A and B has
been transmitted. Packets between B and C are
eligible to send but have not yet been
sent. Packets to the right of C are not eligible
for transmission. All packets between A and C
are sent without waiting for an acknowledgement.
16
Moving the Transmission Window

• Assume packets 0 2 and 5 have been
  acknowledged.
• Slide window to first unacknowledged packet (3 in
  this example).
• Note TCP sets a separate timer for each
  outstanding packet in window.

17
TCP and Sliding Window
0 1 2 3 4 5 6
7
A B C
18
Flow Control

• Do not want to overwhelm capabilities of the
  receiver.
• Transmission window size is variable and set by
  receiver.
• Changed in response to buffer availability.
• Do not want to overwhelm the network.
• Accomplished by limiting number of outstanding
  (unacknowledged) packets to the size of the
  window.
• In terms of performance the larger the window the
  better.
• More efficient to keep the network busy while
  waiting for acknowledgements.

19
Network Congestion

• The routers along a network path have finite
  capacity.
• Limit on their ability to deal with network
  packets.
• If routers get more data than they can handle
  will start queuing packets.
• When buffer space is exhausted will start
  discarding packets.
• Condition termed congestion.
• Results in long delays in transferring data.

20
TCP and Congestion

• Recall TCP uses a timer to determine when to
  retransmit a packet.
• If network is congested then packets and
  acknowledgements will be lost.
• This will trigger more retransmissions.
• Retransmissions will add to network congestion,
  resulting in more packet and acknowledgement
  loss, resulting
• Condition called congestion collapse.

21
Avoiding Congestion Collapse

• TCP defines two mechanisms to avoid congestion
  collapse
• Slow start.
• Multiplicative Decrease.
• Multiplicative Decrease
• TCP maintains a second limit on the size of the
  transmission window called congestion window
  limit.
• The number of packets placed onto the network is
  determined by the minimum of the transmission
  window and congestion window.
• Congestion window is reduced by 50 for each lost
  packet.

22
First Loss
0 1 2 3 4 5 6
7
Transmission and Congestion Window
New Congestion Window
23
Second Loss
0 1 2 3 4 5 6
7
New Congestion Window
Old Congestion Window
Transmission Window
24
Time Between Retransmissions

• In addition to aggressively reducing window size
  also modifies the amount of before the next
  retransmission.
• If one packet lost, TCP doubles the time between
  retransmissions.
• Quite easily reduce window to one packet, and
  wait a long time between retransmissions.

25
Slow Start

• TCP does not immediately double the size of the
  congestion window when packets start flowing
  again.
• Increases the congestion window by one packet per
  acknowledgement.
• Assume congestion window is one.
• Send out packet and wait for acknowledgement.
• When received, increase congestion window to two
  and send next two packets.
• When both acknowledgements are received increase
  congestion window to four.
• When four are received, increase to eight.