Chapter 6: The Transport Layer

1
Chapter 6 The Transport Layer

• CS455/555 Spring 2007

2
The Transport Service

• Provide efficient, reliable, cost-effective
  service to upper layers (application layer)
• Transport entity---the HW/SW that does the work
  to offer this service
• Connection-oriented and connectionless service
• Flow-control is an issue
• Transport layer adds additional
  reliability/flexibility to the services offered
  by the network layer
• True end-to-end protocol
• QoS parameters Specified by a user to the
  transport layer

3
The Transport Service (Cont.)

• We concentrate on the connection-oriented service
  only in this chapter
• Transport layer primitives LISTEN (executed by a
  server), CONNECT (by a client), SEND, RECEIVE,
  DISCONNECT (Fig. 6-2)
• TPDU Transport protocol data units (similar to
  packets at NWL and frames at DLL)
• Figure 6-4 Establish/disconnect a connection

4
Elements of Transport protocols

• TL is similar to DLL error-control, sequencing,
  flow-control, etc.
• How is TL different from DLL? Environments are
  different---single link vs. subnet(s) between the
  two entities.
• Subnet acts as a storage unit this implies
  that a message may float around and not reach the
  destination in a given time.
• No addressing needed at DLL TL needs addresses

5
Elements of Transport protocols Cont.)

• Addressing Transport service access points,
  TSAP (IPlocal port) or AAL-SAP
• Example of setting up a connection between a user
  process and a time-server on another machine
  (Page 494-496)
• Concept of name server to match the users
  (clients) with servers
• TSAP address is generally hierarchical indicating
  the machine and port

6
Establishing a Connection

• How can a transport entity establish a connection
  with another transport entity?
• Problems can arise since a network can store,
  lose, and duplicate packets.
• Example problem If a user has requested for a
  connection, sent a transaction, and requested for
  a disconnection. While the original 3 messages
  were executed correctly, later on duplicates of
  the three may once again at the destination. This
  is a problem.
• Delayed duplicates is a problem.
• Solution 1. Assign a connection ID for each
  connection. It is assigned by the requesting
  user. If a server receives an old ID, it
  recognizes it as a duplicate and ignores it.
• Problem Server has to keep a list of all of the
  earlier connection Ids. What happens if a machine
  crashes and this table is lost?

7
Elements of Transport protocols (Contd.)

• Restricting the life of a packet in a network
  three ways proposed in the text
• Tomlinsons solution to memory loss Each host
  should have a clock that does not lose memory
  when system crashes. The clock is a counter that
  is incremented at uniform intervals. The number
  of bits that hold the counter are larger than the
  bits used for a sequence number.
• Procedure When a connection is set up, the
  initiating host uses the low-order k bits of its
  clock as the initial sequence number (of k
  bits)----so each connection does not start
  numbering their TPDUs from 0 instead they start
  from the number the chosen k-bit sequence number.

8
Elements of Transport protocols (Contd.)
Initial Seq

• While the clock was used to determine the initial
  seq. , after that the seq is incremented
  sequentially. For example, if it started with seq
  1789, then subsequent TPDUs are given numbers
  1790, 1791, and so on.
• What happens if a host crashes during a session?
  Does it remember the last sequence already
  allocated to the last TPDU prior to failure? NO.

9
Elements of Transport protocols (Contd.)

• Why do we want to know the last seq assigned?
  Suppose when the system recovers, the clock is
  1801. But the last seq assigned was 1870. If for
  a new session we start with 1801, then the old
  TPDUs with numbers 1801-1870 may coexist with the
  new 1801-1870 in the network. This causes
  confusion. How do we solve this problem?

10
Elements of Transport protocols (Contd.)

• Note that the rate at which a clock may advance
  may be smaller than the rate of generating TPDUs.
  
• Each TPDU has both the connection and a
  sequence number.
• Example Assume that the clock advances at the
  rate of 1 tick a second. At t30, a host crashed,
  and the last TPDU sent on connection 5 was with
  seq 80.
• After recovering, at time t70, it reopens
  connection 5. The clock is now 70. So it starts
  the new connection with seq 70. Subsequenetly it
  sends TPDUs 70-85. The new 80 may clash with the
  old 80 at the destination which treats as them
  duplicates.
• One solution is for the recovered host to wait
  for T (time-to-live) units of time prior to
  starting a new session. This way all the old
  TPDUs may have exited the network by now. But
  this may be too long time to wait.

11
Elements of Transport protocols (Contd.)

• Two-way handshake to set up a connection may not
  work? Suppose A sends a connection set up request
  with an initial seq number. B accepts the request
  and sends a connection acceptance. But is lost on
  the way. Now a duplicate of the original request
  arrives later. B thinks it is a new request and
  sends Acceptance for the new connection also.
  When A receives it, it thinks it is for the 1st
  one. So there is a confusion. Solution Three-way
  handshake.

12
Three-way handshake protocol

• Each side can start with a different sequenec
  number
• Host 1 chooses a seq. and sends a CR to host 2.
• Host 2 sends an ACK with its own seq as well as
  the seq that it received in host 1s request. On
  receiving the ACK, host 1 starts sending data
  along with a piggy backed ACK for host 2s
  message.
• Duplicate request Suppose a duplicate (delayed)
  CR reaches host 2. It sends an ACK. But when Host
  1receives the duplicate ACK, it rejects it and
  instead sends a REJECT message. Host 2 then
  abandons the connection.
• Duplicate CR and delayed DATAACK from Host 1 On
  receiving the duplicate CR, host 2 will send an
  ACK for the old CR. When Host 1 receives it, it
  will send a REJECT.
• Claim No combination of CR, CA, or other TPDUs
  can cause the protocol to fail.

13
Releasing a Connection

• Asymmetric release (any one releasing breaks a
  connection) vs. symmetric release (treats the
  connection as two separate unidirectional
  connections)
• Asymmetric When a host disconnects and stops
  sending, it should still be ready to receive
  TPDUs until the host gets the disconnect message
• Symmetric Both should agree to disconnect? No
  such algorithm exists to reach consensus in a
  faulty environment.
• Example The two-army problem

14
Releasing a connection (Cont.)

• Three-way handshake to solve the problem in most
  cases
• Host 1 sends DR Host 2 sends DR in return host
  1 sends an ACK and releases the connection when
  host 2 gets ACK, it too releases the connection.
• In case the ACK is lost, host2 timesout and
  releases the connection.
• In case the original DR from host 1 to host 2 is
  lost, host 1 times-out and retransmits.
• Sometimes host 1 may have to several DRs and
  finally just disconnect after several attempts.

15
Flow-control and buffering

• Flow-control is between the two hosts
• At the data link layer, the flow-control is only
  between a DL at one node and all its neighbors.
  Generally, there are only a few of them.
• At the transport layers, a host may be running
  several programs needing several transport
  connections. Thus, buffer allocation is an issue.
• A network layer ACK does not imply transport
  layer ACK.
• The sender should buffer a TPDU until it is
  assured that it has reached the destination host.
• Variable length TPDUs (unlike fixed length data
  frames)

16
Flow-control and buffering

• Source buffering vs. destination buffering
• Low-BW bursty output Source buffering
• File transfer and other high-bandwidth traffic
  Destination buffering
• Buffers could be allocated per connection or
  collectively for all connections running between
  two hosts
• Necessary to deal with dynamic buffer
  allocation---unlike DLL
• Separate buffer space from ACK (this unlike DLL)
  See Fig 6-16

17
Flow-control and buffering

• Each host should send control TPDUs with ACK and
  buffer allocation periodically to prevent
  deadlocks (in case earlier allocations are lost)
• Another constraint is link capacity---we will
  discuss it later.
• Upward multiplexing and downward multiplexing

18
Crash Recovery

• Lost datagrams, disconnected VCs
• Host (e.g. server) crashes---cannot be made
  transparent to higher layers
• Truly End-to-end ACK???

19
A SIMPLE TRANSPORT PROTOCOL

• TL Primitives
• connumLISTEN(local)
• connum CONNECT(local,remote)
• statusSEND(connum, buffer, bytes)
• statusRECEIVE(connum, buffer, bytes)
• statusDISCONNECT(connum)
• Interaction with the network layer to_net, and
  from_net calls
• Packet types CALL REUQUEST, CALL ACCEPTED, CLEAR
  REQUEST, CLEAR CONFIRMATION, DATA, CREDIT

20
INTERNET TRANSPORT PROTOCOLS

• TCP---Connection-oriented
• UDP---connectionless
• The chapters focus is on TCP
• TCP---Transmission Control Protocol
• Reliable end-to-end byte stream over an
  unreliable internetwork

21
TCP Service Model

• End points Sockets
• Socket number host IP 16-bit port number local
  to host
• Connections are identified by the socket pairs at
  both ends
• All TCP connections are full-duplex and
  point-to-point
• Byte stream and not message stream
• If an application wants TCP to send data
  immediately (I.e., not wait for other data to
  come), it can use a PUSH flag
• URGENT data flag

22
A SIMPLE TRANSPORT PROTOCOL

• Each transport connection is always in one of
  seven states
• IDLE
• WAITING
• QUEUED
• ESTABLISHED
• SENDING
• RECEIVING
• DISCONNECTING
• State transitions occur when a primitive is
  executed, a packet arrives, or the timer expires
• State transition table Fig 6-21 (Page 523)
• See Figure 6-21