TCP: Reliable Transport Service

1
TCP Reliable Transport Service

• Chapter 12

2
Introduction

• What was the most impressive aspect of the ENIAC
  computer?
• It was reliable in spite of the fact that its
  components (vacuum tubes) were unreliable
• Internet pioneers faced a similar challenge
• To provide a reliable transport service over
  unreliable (conceptually) technology

3
Introduction (continued)

• IP datagram service is unreliable in that
  datagrams may be
• Lost (usually dropped due to congestion)
• Duplicated
• Delivered out of order
• TCP provides reliable service nevertheless

4
The Service TCP Provides to Applications

• TCP is connection-oriented
• Each TCP connection has exactly two endpoints
• TCP is completely reliable
• TCP allows full duplex communication
• TCP provides a stream interface (unstructured)
• TCP uses a reliable connection startup
• And a graceful shutdown

5
TCP Connections

• Virtual
• An ordered pair of endpoints
• An endpoint is an ordered pair containing
• An IP address
• A TCP port number
• This information is kept by the OS
• netstat

6
Achieving Reliability

• A major challenge
• One computer reboots during a file transfer
• It is rebooted and a second connection is
  established
• Messages from the two sessions get mixed up
• Packets from the first session must be rejected

7
Positive Acknowledgement with Retransmission

• Retransmission is one of many techniques
• Send packet start timer
• Two possibilities
• Packet arrives
• Acknowledgement is sent back
• Stop timer and proceed
• Packet fails to arrive
• Timer goes off
• Resend packet and set another timer

8
Positive Acknowledgement

• http//www.calvin.edu/lave/figure-12.1.pdf
• Suppose an acknowledgement is lost.
• A duplicate packet will be sent.
• Sequence numbers are used to allow the
  destination software to reassemble a packet.

9
The Idea Behind Sliding Windows

• Stop-and-wait wastes network bandwidth
• Sliding window protocols allow the sender to
  transmit multiple packets before waiting for an
  acknowledgement.
• http//www.calvin.edu/lave/figure-12.3.pdf
• http//www.calvin.edu/lave/figure-12.4.pdf

10
The Transmission Control Protocol

• Very complex
• Specifies format of data and acks
• Specifies how TCP software distinguishes among
  multiple destinations on a host
• How hosts recover from lost or duplicated packets
• How hosts initiate and terminate a transfer

11
The Transmission Control Protocol

• Does not include
• The details of interaction with an application
• Eg, sockets
• Can be used with a variety of packet delivery
  systems

12
Ports, Connections, and Endpoints

• Like UDP, TCP resides in the transport layer
• Like UDP, TCP uses protocol port numbers to
  identify the ultimate destination within a host
• TCP ports are much more complex than UDP ports

13
Ports, Connections, and Endpoints (continued)

• TCP port numbers do not correspond to a single
  object. TCP is built on the connection
  abstraction, in which objects are identified by
  connections.
• A connection is a pair of endpoints.
• An endpoint is a pair, (host, port).
• Two connections might share one endpoint.

14
Passive and Active Opens

• An application program on one end performs a
  passive open by indicating to the OS that it will
  accept an incoming connection. The OS provides a
  TCP port number for its end of the connection.
• An application program on the other end uses an
  active open request a connection.

15
Passive and Active Opens (continued)

• The two TCP software modules communicate to
  establish and verify a connection.
• Once the connection is established, the
  application programs begin to pass data.
• TCP modules at each end guarantee reliable
  delivery.

16
Segments, Streams, and Sequence Numbers

• TCP views the data as a sequence of octets that
  it divides into segments. Usually, each segment
  travels in a single IP datagram.
• TCP uses a specialized sliding window mechanism
  to solve two problems efficient transmission and
  flow control.

17
Segments, Streams, and Sequence Numbers
(continued)

• TCP allows the receiver to restrict transmission
  until it has sufficient buffer space to
  accommodate more data.
• The sliding window mechanism operates the octet
  level. A sender keeps three pointers associated
  with each connection.
• http//www.calvin.edu/lave/figure-12.6.pdf

18
Sliding Window Protocol
19
Segments, Streams, and Sequence Numbers
(continued)

• TCP software at each end maintains two windows
  per connection for a total of four.
• One slides as data is sent, the other as data is
  received.

20
Variable Window Size and Flow Control

• TCP allows the window size to vary over time.
  Each ack contains a window advertisement that
  specifies how many octets of data the receiver is
  prepared to accept.
• This provides flow control. The window size may
  even be zero.

21
Variable Window Size and Flow Control (continued)

• Solves two independent problems.
• A hand-held PDA communicating with a fast PC may
  be overrun.
• A mechanism is needed to allow intermediate
  systems (routers) to control a source that sends
  more traffic than the system can tolerate.

22
Variable Window Size and Flow Control (continued)

• When intermediate machines become overloaded, the
  condition is called congestion, and mechanisms to
  solve the problem are called congestion control
  mechanisms.
• A poorly-designed protocol will make congestion
  worse.

23
TCP Segment Format

• The unit of transfer is called a segment.
• http//www.calvin.edu/lave/figure-12.7.pdf
• HLEN is in multiples of 32-bits. Length varies
  due to the presence of options.
• http//www.calvin.edu/lave/figure-12.8.pdf

24
Out of Band Data

• Interrupting a file transfer.
• Data may be identified as urgent.
• Implementation depends on the software.
• The URG bit will be set, and the URGENT POINTER
  refers to the end of the urgent data.
• TCP delivers such data to the server immediately,
  out of band. Eg, telnet.

25
TCP Options

• Each option begins with a 1-octet field
  specifying the type with a 1-octet length field.
• Padding is used to guarantee a multiple of 32
  bits in the header.

26
Maximum Segment Size Option

• Allows a receiver to specify a maximum segment
  size. This allows heterogeneous systems to
  communicate.
• Endpoints may establish the minimum MTU along the
  path between them.
• Small segments imply overhead. Large segments may
  have to be fragmented.

27
Maximum Segment Size Option (continued)

• It is hard to achieve the optimum segment size.
• Many implementations of TCP do not include a
  mechanism for doing so.
• Routes may change dynamically.
• The optimum size depends on lower-level protocol
  headers (IP options).

28
Window Scaling Options

• The WINDOW field in the TCP header is 16 bits
  long, the maximum size window is 64 Kbytes. A
  larger size is needed to obtain high throughput
  on a network such as a satellite channel, a long
  fat pipe.
• A type, a length, and a shift value, S.

29
Window Scaling Options (continued)

• A receiver shifts the WINDOW field S bits.
• The option can be negotiated when the connection
  is initially established in which case all
  successive advertisements are assumed to use the
  negotiated scale or the scaling factor can vary
  with each segment.

30
Timestamp Option

• Added to help TCP compute the delay on the
  underlying network and to handle the case where
  sequence numbers exceed 2 to the 32nd power.
  Protect Against Wrapped Sequence numbers, PAWS.
• Sender includes a timestamp, receiver returns it
  allowing the sender to compute the elapsed time.

31
TCP Checksum Option

• Like UDP, and for the same reason, TCP uses a
  pseudo-header to compute the checksum
• http//www.calvin.edu/lave/figure-12.9.pdf

32
Acknowledgements, Retransmission, and Timeouts

• Because TCP sends data in variable length
  segments, and because retransmitted segments can
  include more data than the original, acks cannot
  easily refer to segments. They refer to a
  position in the stream using the sequence
  numbers.
• A TCP ack specifies the sequence number of the
  next octet that the receiver expects to receive.

33
Acknowledgements, etc (continued)

• This acknowledgement scheme is called cumulative
  because it reports how much of the stream has
  accumulated.
• Advantages
• acks are easy to generate and unambiguous
• lost acks do not necessarily force
  retransmissions.

34
Acknowledgements, etc (continued)

• Disadvantages
• The sender does not receive information about all
  successful transmissions, but only about a single
  position in the stream.
• Example
• Sender sends 5 segments. The last 4 make it.
• Receiver keeps saying it wants the first one.
• Sender has no way of knowing that 80 arrived

35
Acknowledgements, etc (continued)

• When a timeout occurs, the sender must choose
  between two potentially inefficient schemes.
• Resend all 5 segments.
• Resend only the first and wait for the ack before
  deciding what to do next.

36
Timeout and Retransmission

• If the timer expires, the segment is resent.
• TCP is intended for use in an internet
  environment. It is impossible to know in advance
  how quickly acks will return.
• Transmission time varies dynamically.

37
Adaptive Retransmission

• TCP monitors the performance of each connection
  and deduces reasonable values for timeouts.
• TCP records the time at which each segment is
  sent and the time an ack arrives for the data in
  that segment.
• TCP adjusts based on an estimated round trip
  time, RTT.

38
RTT

• RTT (a Old_RTT) ((1 a)
  New_Round_Trip_Sample)
• Choosing a close to 1 makes the weighted average
  immune to changes that last a short time.
• Choosing a close to 0 makes the average respond
  quickly to changes in delay.

39
RTT (continued)

• http//www.calvin.edu/lave/figure-12.10.pdf
• TCP computes the timeout value as b RTT
• Originally b was set to 2. Better techniques have
  been developed recently.

40
Accurate Measurement of Round Trip Samples

• Sounds easy, but .
• Acknowledgement ambiguity. If a segment must be
  resent, and then an ack arrives, is it for the
  original segment or the retransmitted segment?
• No way to tell so what should we assume?

41
Accurate Measurement of Round Trip Samples
(continued)

• If you measure from the original segment, RTT
  will grow without bound in cases where the
  internet loses datagrams.
• If you measure from the most recent
  retransmission can also fail. Suppose the
  end-to-end delay suddenly increases. The original
  ack may be associated with the retransmission
  lowering the next timeout. RTT can end up being
  half what it should be so that each segment is
  sent twice.

42
Karns Algorithm

• Ignore round trip estimates for retransmitted
  segments.
• A simplistic implementation can lead to failure.
  Consider a segment sent after a sharp increase in
  delay. The timeout is too low, but the next batch
  of round trip estimates will be ignored.

43
Karns Algorithm (continued)

• Use a timer backoff strategy. If the timer
  expires, increase the timeout (multiply by a
  constant, say 2).
• Works well even in networks with high packet loss.

44
Responding to High Variance in Delay

• Previously mentioned computations do not adapt
  well to a wide range in variation in delay.
• Queueing theory suggests that the round trip time
  increases proportional to 1/(1-L), where L is the
  current fractional load, and the variation, s,
  is inversely proportional to (1-L) squared.

45
Responding to High Variance in Delay (continued)

• If an internet is running at 50 capacity, we
  expect the round trip delay to vary by a factor
  of 4 from the mean round trip time. At 80 the
  factor is 25. Using the original TCP technique
  for estimating round trip time with b 2, means
  that the round trip estimation can adapt to loads
  of at most 30.

46
Responding to High Variance in Delay (continued)

• The 1989 specification for TCP requires
  estimating the average round trip time and the
  variance and to use the estimated variance as the
  constant, b.

47
Responding to High Variance in Delay (continued)

• DIFF SAMPLE Old_RTT
• Smoothed_RTT Old_RTT d DIFF
• DEV Old_DEV r(DIFF - Old_DEV)
• Timeout Smothed_RTT e DEV
• Where DEV is the estimated mean deviation, d and
  r are fractions, and e is a factor

48
Responding to High Variance in Delay (continued)

• d controls how quickly the new sample affects the
  weighted average
• r controls how quickly the new sample affects the
  mean deviation
• e controls how much the deviation affects the
  round trip timeout

49
Responding to High Variance in Delay (continued)

• TCP chooses d and r to be inverses of powers of 2
• Research suggests d 1/8, r ¼, and e 3 will
  work well

50
Responding to High Variance in Delay (continued)

• http//www.calvin.edu/lave/figure-12.11.pdf
• http//www.calvin.edu/lave/figure-12.12.pdf
• Comments on page 210

51
Response to Congestion

• Severe delay caused by an overload of datagrams
  at one or more switching points, eg routers
• The router begins to enqueue datagrams until it
  can forward them. When it reaches capacity, it
  starts to drop datagrams.
• Endpoints do not know where or why congestion has
  occurred.

52
Response to Congestion (continued)

• Most transmission protocols use timeout and
  retransmission so they respond to increased delay
  by retransmitting datagrams, increasing the
  congestion.
• This leads to congestion collapse.
• Therefore, TCP must reduce transmission rates
  when congestion occurs.

53
Response to Congestion (continued)

• Routers watch queue lengths and use techniques
  such as ICMP source quench to inform hosts that
  congestion has occurred, but transport protocols
  can help avoid congestion by reducing
  transmission rates automatically.
• Two techniques are used slow-start and
  multiplicative decrease.

54
Multiplicative Decrease

• In addition to the limit imposed by the sliding
  window protocol, TCP uses a second limit, the
  congestion window limit or congestion window.
• Allowed_window is the minimum of the advertised
  window and the congestion window.

55
Multiplicative Decrease (continued)

• TCP assumes that most datagram loss comes from
  congestion (wireless!) and uses the following
  strategy
• Upon loss of a segment reduce the congestion
  window by half (min is 1).
• For remaining segments backoff the transmission
  timer exponentially.

56
Slow-start

• When starting traffic on a new connection or
  increasing traffic after congestion, increase the
  congestion window by one segment each time an ack
  arrives.
• It takes only log base 2 of N round trips to get
  a congestion window of N so slow-start is
  something of a misnomer.
• These techniques have improved the performance of
  TCP by factors of 2 to 10.

57
Slow-start (continued)

• One additional restriction.
• Once the congestion window reaches one half of
  its size before congestion, TCP enters a
  congestion avoidance phase, and slows down the
  rate of increase. It increases the congestion by
  1 only if all segments in the window have been
  acknowledged.
• Additive Increase Multiple Decrease
• AIMD

58
Fast Recovery and Other Techniques

• The early version of TCP, sometimes referred to
  as Tahoe used the scheme described above.
• In 1990 the Reno version of TCP appeared
  introducing several changes.
• Fast recovery or fast retransmit has higher
  throughput where only occasional loss occurs.

59
Fast Recovery

• The loss of a single segment
• Receiver sends an ACK for the point in the stream
  where the missing segment begins.
• From the senders point of view, a series of ACKs
  arrive with the same sequence number.
• Fast retransmit uses a series of three duplicate
  ACKs to trigger a retransmit before the timer
  goes off.

60
Fast Recovery (continued)

• To maintain higher throughput, fast retransmit
  continues to send data from the window while
  waiting for an ACK for the retransmitted segment.
• The congestion window is artificially inflated
  (increased by 1 for each duplicate ACK). TCP
  keeps many segments in flight.

61
Fast Recovery (continued)

• The NewReno version handles the case where two
  segments are lost in a single window.
• An ACK arrives
• Sequence number at end of window
• Only missing segment
• In the middle
• Retransmit that second segment immediately

62
Fast Recovery (continued)

• TCP reduces transmission when congestion occurs.
  UDP does not.
• TFRC (TCP Friendly Rate Control) has been
  proposed.
• Emulates TCP for UDP slowing the rate at which
  UDP is sent.

63
Explicit Feedback Mechanisms

• TCP infers congestion. TCP has implicit
  information.
• Selective Acknowledgement (SACK)
• Receiver specifies which data has been received
  and which is still missing. The sender then knows
  exactly which segment to resend.
• Proposed. More work to do. Page 215.
• First sequence number in a block and the one
  immediately beyond the block.

64
Explicit Congestion Notification

• ECN. Routers notify TCP when congestion occurs.
• Routers use a pair of bits in the IP header to
  record congestion.
• Receiver uses a pair of bits in the TCP header to
  inform the sender.

65
Congestion, Tail Drop, and TCP

• The impact of the lack of communication between
  layers.
• A choice of policy or implementation at one layer
  can have a dramatic effect on the performance of
  higher layers. (middle/upper management)
• If a router delays some datagrams more than
  others, TCP will back off.

66
Congestion, Tail Drop, and TCP (continued)

• A router becomes overrun and drops datagrams.
  Early routers used a tail-drop policy to manage
  queue overflow.
• If the input queue is filled when a datagram
  arrives, discard the datagram.
• A router will discard one segment from N
  connections rather N from one connection.
• Global synchronization. All connections back off.

67
Random Early Detecion

• RED. Avoids tail-drop whenever possible.
• Random early drop (discard).
• Let N be the number of datagrams in the queue,
  and let a datagram arrive.
• N lt min. Add to queue.
• N gt max. Discard.
• Min lt N lt max. Randomly discard.

68
RED (continued)

• The choice of min, max, and p is key.
• Min must be large enough to achieve high
  utilization. Max must be more than min by the
  increase in queue size during one TCP round trip.
• P is computed based on the current queue size and
  the thresholds. (Linear)

69
RED (continued)

• Network traffic is bursty so RED cannot use a
  linear scheme.
• Later datagrams would be dropped with higher
  probability with a negative impact on TCP
  throughput.
• Borrow a page from TCPs book use a weighted
  average queue size using the old average and the
  current queue size.

70
RED (continued)

• Avg (1 gamma) Old_avg gamma
  Current_queue_size
• If gamma is small, the average will track long
  term trends, but remain immune to short bursts.
  (gamma .02)
• RED computations use powers of 2 and integer
  arithmetic.
• Queue is measured in octets.

71
RED (continued)

• Small datagrams (remote login traffic or requests
  to servers or ACKs) have lower probability of
  being dropped.
• The IETF now recommends that routers implement
  RED.
• Handles congestion, avoids sychronization, and
  allows short bursts well.

72
Establishing a TCP Handshake

• To establish a connection, TCP uses a three-way
  handshake.
• http//www.calvin.edu/lave/figure-12.13.pdf
• SYN, SYN-ACK, ACK
• Avoids getting connections mixed up

73
Initial Sequence Numbers

• A initiates a connection sends its initial
  sequence number, x, in the sequence field
• B records the sequence number and replies by
  sending its sequence number as well as an ACK
  that it expects x1.
• Use the number of the next octet expected.
• Sequence numbers are supposed to be random!

74
Closing a TCP Connection

• Connections are closed in one direction. Only
  when both connections have been closed does the
  TCP software delete its record of the connection.
• http//www.calvin.edu/lave/figure-12.14.pdf

75
Closing a TCP Connection (continued)

• Subtle. After receiving a FIN, TCP sends an ACK
  and informs the application. This may take time,
  even human interaction. Finally, the second FIN
  is sent, and the original site replies with the
  third message, an ACK.

76
TCP Connection Reset

• Sometimes abnormal conditions arise that force an
  application to break a connection.
• TCP provides a reset facility.
• One side sends a segment with the RST bit in the
  CODE field set.
• The other side aborts the connection and informs
  its application program. Transfer ceases
  immediately in both directions.

77
TCP State Machine

• http//www.calvin.edu/lave/figure-12.15.pdf

78
Forcing Data Delivery

• Example, a remote session, say ssh.
• TCP provides a push mechanism that forces
  delivery without waiting for the buffer to fill.
  The PSH bit in the CODE field is set so the other
  end will pass on the data immediately.

79
Reserved TCP Port Numbers

• http//www.calvin.edu/lave/figure-12.16.pdf
• Although TCP and UDP are independent, they use
  the same numbers where appropriate, eg 53.
• Ports 0-1023 are reserved (well known ports).

80
TCP Performance

• TCP tackles a complex task, yet the code is
  neither cumbersome nor inefficient.
• At Cray Research, Inc. researchers have
  demonstrated TCP throughput approaching a gigabit
  per second.

81
Silly Window Syndrome and Small Packets

• When the receiving application reads one octet of
  data, that much space is available in its buffer.
  It might send an advertisement showing a window
  of one octet. The sender then sends another
  octet.
• The silly window syndrome (SWS) plagued early TCP
  implementations.

82
Avoiding Silly Window Syndrome

• TCP specifications now include heuristics to
  prevent SWS.
• Sending machine avoids transmitting a small
  amount of data in each segment.
• Receiver avoids sending small increments in
  window advertisements.

83
Receive-Side Silly Window Avoidance

• Suppose a receiving application extracts data
  slowly.
• Wait until the available space reaches one half
  of the total buffer size or a maximum sized
  segment.

84
Delayed Acknowledgements

• Second approach to SWS
• Delay sending an ACK when the window is not
  sufficiently large to advertise.
• Increase throughput.
• Include data with ACKs (piggyback)
• Problem? Dont delay too long or the data will be
  resent. 500 millisecond limit. ACK every other
  segment.

85
Send-Side Silly Window Avoidance

• Surprising and elegant.
• Allow the sending application to make multiple
  calls to write, and collect the data before
  sending. This is called clumping.
• Algorithm is adaptive, something like slow start.
  It is called self-clocking because it does not
  compute delays.

86
Send-Side SWA (continued)

• When an application generates additional data to
  be sent over a connection for which previous data
  has been sent but not acknowledged, place the
  data in a buffer but do not send until you have a
  maximum sized segment. If still waiting to send
  when an ACK arrives, send all the data. Apply the
  rule even when the user requests a push operation.

87
Send-Side SWA (continued)

• If the application is reasonably fast compared to
  the network, successive segments will contain
  many octets. If the application is slow, small
  segments will be sent without long delay.
• This is called the Nagel algorithm and requires
  little computational overhead.