1. TCP: Reliable Stream Transport
Chapter 13

2. Introduction
Application programs often need to send large volumes of data from one computer to another
An unreliable connectionless delivery system for large volume transfers is tedious and annoying
Why?
There is a need for reliable stream delivery which
isolates the application programs from the details of networking and
defines a uniform interface for stream transfer

3. Properties of a Reliable Delivery Service
Stream Orientation - data is a stream of bytes
Virtual Circuit Connection - a connection is agreed upon, data is transferred, disconnection
Buffered Transfer - sending application decides data size, the size may change at a lower layer, then data is available to destination application
Unstructured Stream - structure boundaries are not preserved
Full Duplex Connection - concurrent transfer in both directions

4. Providing Reliability
Reliable protocols usually require the receiver to send an acknowledgement (ACK) or a PAR
The sender keeps a copy of the packet that was sent and waits for an ACK
if an ACK is received, next packet is sent
if an ACK is not received, after a timer expires, the packet is retransmitted
Figure 13.1 is a timeline for normal case
Figure 13.2 is a timeline showing packet loss
What to do about duplicate packets or ACKs?
Remember the sequence # field in IP header?

5. Sliding Windows
In the case of Figure 13.1, the network is idle during machine delays
We can allow the sender to transmit multiple packets before waiting for an ACK
A protocol can place a small window over the sequence of packets to send and transmit all packets in that window as in Figure 13.3
The window slides forward as ACKs are sent for packets in the window

6. Sliding Windows
The performance of sliding window protocols depends on:
the window size
the speed at which the network accepts packets
See Figure 13.4 which timelines 3 packets
A window of size one is just like a positive acknowledgement with retransmission, or PAR
A steady state is reached when the sender can transmit packets as fast as the network can - the network is kept busy

7. Sliding Windows A sliding window protocol must:
remember which packets have been acknowledged
keep a separate timer for each unacknowledged packet
if a packet is lost, timer expires and sender retransmits
When the sender slides its window, it moves past all acknowledged packets
Similar software exists in the receiver
The window partitions the packets into:
those sent and ACKed, those being transmitted and those not yet transmitted

8. The Transmission Control Protocol
TCP is a transmission protocol, not software
TCP defines the concepts, it is not the implementation
What does TCP specify?
the format of data and acknowledgements
procedures to ensure that data arrives correctly
how software distinguishes among multiple destinations on a given machine
how to recover from errors
how to set up a connection to transfer data

9. The Transmission Control Protocol
What does the protocol not provide?
does not dictate the details of the interface between TCP and the application layer
TCP can be used with a variety of packet delivery systems, not just IP
TCP can use dial-up lines, local area networks, high speed fiber, long haul, etc.

10. Ports, Connections and Endpoints
TCP resides above the IP layer, along with UDP as shown in Figure 13.5
TCP
allows multiple application programs to communicate concurrently
demultiplexes incoming TCP traffic among applications
uses protocol port numbers to identify the ultimate destination within a machine

11. Ports, Connections and Endpoints
TCP ports identify virtual circuit connections, not a single object like the UDP port identifies
Connections are identified by a pair of endpoints
An endpoint is a pair of integers (host, port) where host is the IP address for a host and port is a TCP port on that host
Example:
A connection between a machine at MIT ( ) and a machine at Purdue ( ) is defined by ( , 1069) and ( , 25)

12. Ports, Connections and Endpoints
Multiple connections may be made to another protocol port on a given machine
Multiple connections may share a complete endpoint
Because TCP identifies a connection by a pair of endpoints, a given TCP port number can be shared by multiple connections on the same machine
this means that a program can provide concurrent service to multiple connections simultaneously without giving unique port numbers for each one ( )

13. Passive and Active Opens
TCP is connection-oriented and requires both endpoints to agree to participate
Thus, the application program on one end performs a passive open by telling its operating system that it is willing to accept an incoming connection
At that time, the O.S. assigns a TCP port number for this end
The application program at the other end issues an active open to its O.S. to request a connection
The two establish and verify a connection

14. Segments, Streams and Sequence Numbers
The data stream is viewed as a sequence of octets that are divided into segments for transmission
Each segment transmitted in a single IP datagram
Octets in the TCP data stream are numbered sequentially, and a sender keeps 3 pointers for each connection (see Figure 13.6):
the left of the sliding window:ack’d vs. unack’d
the right of the sliding window, marks the highest sequence # that can be sent before ACKs required
octets sent and octets not sent

15. Segments, Streams and Sequence Numbers
The receiver keeps similar window information
Because TCP connections are full-duplex, the TCP software maintains window information at each end for both sending and receiving
thus, four windows per connection

16. Variable Window Size and Flow Control
TCP allows the window size to vary over time
Each ACK indicates how many octets have been received
It also indicates a window advertisement which states how many additional octets the receiver is willing to receive (specifies the receiver’s current buffer size)
In response to an increased window advertisement, the sender increases the size of its sliding window
In response to a decreased window advertisement, the sender decreases the size of its sliding window

17. Variable Window Size and Flow Control
A sliding window
provides flow control
provides reliable transfer
To stop all transfer, the receiver can send a window advertisement of zero
A sender is allowed to transmit a segment with an urgent bit set
To avoid deadlock, the sender can probe periodically, by asking if buffer space is available

18. Variable Window Size and Flow Control
Two flow problems
internet flow control between source and ultimate destination (called end-to-end flow control)
flow control for intermediate systems (routers)
When intermediate systems become overloaded, we have a condition called congestion
We will look at ways to handle congestion later

19. TCP Segment Format
The unit of transfer between the TCP software on two machines is called a segment
Segments are exchanged to:
establish connections
transfer data
send ACKs
advertise window sizes
close connections
Using piggybacking, an ACK from one machine to another may travel in the same segment with data travelling in the opposite direction (in reality, this does not happen often)

20. TCP Segment Format A segment is made up of a header and data
The TCP header consists of:
source and destination ports which identify applications
a sequence number identifying this data’s position in the sender’s byte stream (stream flowing this direction)
an ACK number identifying which octet the source is expecting next (stream flowing opposite direction of this stream)
header length ( in 32 bit multiples) to cover options
6 code bits to indicate purpose and contents of segment (URG, ACK, PSH, RST, SYN, FIN)

21. TCP Segment Format how much data it is willing to accept, buffer size
window advertisements accompany each segment, and are piggybacking on ACKs
checksum
urgent pointer which specifies the end of the urgent data
what marks the beginning of the urgent data?

22. Out of Band Data
It may be necessary to send messages that are not a part of the regular data in the stream, out of band data
Example: sending a keyboard sequence that interrupts or aborts the program at the other end
This data is specified as urgent and is indicated by the URG bit and is in the data portion of this segment
its end is known by the number in the Urgent Pointer field

23. Maximum Segment Size Option
Both ends need to agree on a maximum segment size they will transfer
If the two endpoints are on the same network, this will be the network’s MTU
If not on the same network, they will go through a process of MTU discovery, or choose 536 (default size of IP datagram, 576, plus TCP and IP headers)
What happens when:
segment size is too small? network utilization is low
segment size is too big? fragmentation

24. TCP Checksum Computation
A pseudo header is prepended
Zeroes are added to make the segment a multiple of 16 bits
Takes the one’s complement of sum of all 16-bit entities, assuming the original checksum is zero
On the receiving side, IP passes source and destination IP addresses when it passes the segment

25. Acknowledgements and Retransmission
The receiver reconstructs the original stream sent by the sender by piecing together segments
Some segments may be lost, delayed, or arrive out of order
The receiver uses the sequence numbers to reconstruct the stream
The receiver acknowledges the longest contiguous prefix of the stream that has arrived correctly
It sends the sequence number of the segment it is expecting next

26. Acknowledgements and Retransmission
When an acknowledgement is not received within a given timeout period, the segments may be retransmitted
only the first unacknowledged segment
worst case is to have to retransmit one at a time
all segments in the window
worst case is only the first one was needed

27. Timeout and Retransmission
Every time a segment is sent, TCP starts a timer and waits for an ACK
If the timer expires before an ACK is received, it assumes the segment was lost or corrupted and retransmits it
How long should the timeout be?
See Figure for roundtrip times of 100 successive IP datagrams on the Internet

28. Timeout and Retransmission
An adaptive retransmission algorithm for TCP monitors a connection and determines a reasonable timeout period (RTT - Round Trip Time) for that connection
TCP records the time that a segment was sent, and the time that an ACK was received for it
the elapsed time is a round trip sample
a new sample is obtained and the RTT is modified by
RTT = (alpha * oldRTT) + ((1-alpha) * newSample)
When performance changes, the value of RTT is modified accordingly

29. Timeout and Retransmission
When the value for alpha is close to:
zero: the weighted average responds to changes in delay very quickly
one: the weighted average does not change significantly when a temporary change is noticed

30. Accurate Measurement of Round Trip Samples
In measuring the round trip samples, we could have acknowledgement ambiguity if we had to retransmit
When we receive an ACK is it for the original datagram or for the retransmitted datagram?
Problems with assuming the ACK for the original, and with assuming the ACK is for the most recent transmission

31. Karn’s Algorithm
Karn’s algorithm for handling ambiguous acknowledgements is to not update the round trip estimate for retransmitted segments
If retransmission times are completely ignored, a large delay could not be noticed and the problem of spirally retransmissions could occur
Karn’s algorithm suggests a timer backoff
if the timer expires and retransmission is done, TCP increases the timeout up to a point that is larger than the delay along any path in the internet

32. Responding to High Variance in Delay
Research shows that previous algorithms do not respond well to a wide range of variation in delay
Queuing theory suggests that the variation in round trip time varies proportional to 1/(1-L) where L is the current network load
A 1989 specification of TCP requires estimating the average round trip time and the variance

33. Responding to Congestion
When congestion occurs, delays increase and routers queue datagrams
Routers have finite buffer space and when that limit is reached, datagrams will be discarded
Delay causes retransmissions and retransmissions cause more delay ... until congestion collapse
To avoid congestion, transmission rates must be reduced
Routers watch queue length and use ICMPsource quench

34. Responding to Congestion
TCP maintains a congestion window which is the smaller of the receiver’s advertised window and the current congestion window
multiplicative decrease - reduces the congestion window by half each time a segment is lost
if loss continues, TCP limits transmission to one datagram and doubles timeout values before retransmitting
provides significant and fast response to congestion
lets routers clear datagrams in their queues

35. Responding to Congestion
How does TCP recover when congestion ends?
slow start
start with setting the congestion window to the size of one segment
increase the congestion window by one segment each time an acknowledgement arrives
thus, one segment, two segments, four segments, eight …
once the congestion window is half of its original size before congestion occurred, it slows down and increases by one segment only, if all goes well

36. Congestion, Tail Drop and TCP
An early policy, called Tail Drop discarded a datagram if the input queue at a router was full
Since datagrams are typically multiplexed, with successive datagrams from a different source, the router might discard one segment from N connections rather than N segments from one connection
So, what’s wrong with that?
All N instances of TCP will enter slow start at the same time

37. Random Early Discard (RED)
A router uses Tmin and Tmax to mark positions in the queue
If the queue currently contains fewer than Tmin datagrams, add the new datagram to the queue
If the queue contains more than Tmax datagrams, discard the new datagram
If the queue contains between Tmin and Tmax datagrams, randomly discard the datagram with a probability p

38. Random Early Discard (RED)
A router slowly and randomly drops datagrams as congestion increases
How are Tmin and Tmax determined?
How is the value of p determined?

39. Establishing a TCP Connection
To establish a connection, TCP uses a three-way handshake as shown in Figure (consider right side the server and left side the client - see handout)
the first segment has the SYN bit set, port number of the server that the client wants to connect to, and an initial sequence number (ISN)
the second segment has both the SYN and ACK bits set, server sends its own ISN, the ACK contains the client’s ISN plus one
the third segment is an ACK with the server’s ISN plus one

40. Establishing a TCP Connection
Usually the TCP software on one machine waits passively for a handshake, the other initiates it
However, if both request a connection at the same time, this can be done, and is called a simultaneous open

41. Initial Sequence Numbers
Each machine in a connection chooses an ISN at random
These numbers are shared with the other machine in the connection establishment handshake
Segments that follow in the data transfer period will be numbered sequentially from the initial numbers

42. Closing a TCP Connection
When one side of the connection has no more data to send, it will close the connection in one direction
At the end of its data transfer, this side will send a segment with the FIN bit set
The other side will acknowledge the FIN segment and may continue sending until it has no more data, when it sends its own FIN segment
See Figure third step is ACK

43. TCP Connection Reset
Normally, connections are released as in the previous section
But, sometimes a connection is broken by accident
To completely close the connection, it can be reset
This is done to release resources like buffers
One side sends a segment with the RST bit set

44. TCP State Machine See Figure 13.15 Circles represent states
Arrows represent transitions between states
Labels show what causes the transition and what it sends in response
Notice the syn, ack , rst and fin indicators that are in the segment header

45. Forcing Data Delivery
We don’t always want to wait until a buffer is full in order to send it
If an application is gathering keystrokes, we want them to appear on the screen as they are entered, not as a block of keystrokes
TCP provides a push operation that forces delivery of octets without waiting for the buffer to fill - it uses the psh code

46. TCP Reserved Port Numbers
Well-known ports were originally those under 256
But some over 1024 have been assigned for such
TCP and UDP have some commonly numbered ports
See some of the currently assigned TCP port numbers in Figure 13.16

47. Silly Window Syndrome
The problem that is presented for this discussion is one that arises when an application reads incoming data one octet at a time
When a connection is established, the receiving TCP allocates a buffer of K octets
If the sender generates data quickly, the sending TCP transmits segments to fill the entire window
Then the receiver indicates that it has no buffer space available
When the receiver reads one octet, it can advertise that it has one octet of space available
Thus, very small segments are generated with large overhead (41 bytes/segment) and much processing for little data

48. Avoiding Silly Window Syndrome
Receive Side
The receiver maintains the actual available buffer space, but will not advertise an increase until the window can be advanced by:
half of the receiver’s buffer
or the maximum size of a segment
Delayed Acknowledgements
The receiver delays sending an ACK when the window is not sufficiently large enough to advertise (!> 500ms)
advantage: delayed ACKs decrease traffic, increase throughput
disadvantage: the sender may retransmit

49. Avoiding Silly Window Syndrome
Send Side
New data is placed in the buffer, but the sender does not send until a maximum size segment is filled
If still waiting to send when an ACK arrives, send all data that has accumulated in the buffer
Apply the rule even if the user has requested a push
This is called the Nagle algorithm and requires little computation
In general, the receiver avoids advertising a small window and the sender delays transmission

50. Summary
TCP provides
reliable stream delivery
full-duplex connections between two machines allowing for exchange of large volumes of data
TCP uses sliding windows for efficient use of the network
TCP provides flow control and allows systems of varying speeds to communicate
Segments are used to transfer data or control information

51. Summary
TCP implements flow control by letting the receiver advertise the amount of data it is willing to accept, yet supports out of band messages
Current TCP uses:
exponential backoff for retransmission timers
congestion avoidance algorithms
heuristics to avoid transferring small packets

52. For Next Time
Several chapters on routing
Read Chapter 14