Content

  1. Service Model

  2. Segment Structure

    1. Header

  3. Connection Management

    1. Connection Establishment - Three-Way Handshake
    2. Connection Termination - Four-Way Handshake
  4. Reliable Data Transfer
  5. Retransmission Timeout Mechanism
    1. Adaptive Timeout Mechanism
    2. Retransmission Ambiguity
  6. Flow Control
  7. Congestion Control
    1. Congestion Window
    2. Bandwidth Probing
    3. TCP Congestion Control Algorithm

Service Model

  • Endpoint-to-endpoint
  • Full duplex
    • Maintain independent sequencing info for each direction
  • Connection-oriented
    • Logical connection only in the end systems (not intermediate devices)
    • Connection setup & release
      • Handshaking to establish shared state info
      • Keeping connection state before close
  • Reliable in-order byte stream
    • Reliability
      • Checksum, sequence number, acknowledgment, retransmission, timeout
    • Stream-oriented
      • Pack application data (from 1 or more send() calls) into segments within Maximum Segment Size (MSS)
  • Multiplexing/demultiplexing
    • Each connection specified by (my IP,my port,peer IP,peer port)

Segment Structure

  • Sequence number
    • Each byte has its own sequence #
    • Indicated the first byte in this segment
    • Initial sequence number (ISN) is generated by one end during connection setup and informed to the peer during 3-way handshaking
  • Acknowledgement number field
    • Valid when ACK bit is set in header
    • Next expected byte to be received
    • Cumulative ACK (same as GBN)
  • Fixed 20-byte header + up to 40-bytes optional part
  • Each side determines MSS and inform its peer during connection setup
    • Limited by lower-link layer MTU (max transmission unit)
  • Header length
    • In unit of 4-bytes
  • Flag field
    • URG: urgent data
    • ACK: segment carries acknowledgement
    • PSH: urge to pass to upper layer immediately
    • RST: reset connection
    • SYN: used in connection setup
    • FIN: used in connection teardow
  • Receive window
    • For flow control
    • Size of remaining buffer at the receiver's side in bytes
    • Becomes the size of send window
  • Internet checksum
    • Same as UDP
  • Options
    • MSS: largest segment a node wants to receive, announced during connection establishment

Connection Management

Connection Establishment - Three-Way Handshake

  • If no process listening on the server port, a reset segment (RST bit set) will be sent
  • Receiver of SYN must sent back an ACK, since a seq # is consumed

Connection Termination - Four-Way Handshake

  • Treat bidirectional channel as 2 unidirectional channels; release connection independently
  • Timed wait state
    • To receive retransmitted FIN segment if ACK lost
  • Each FIN consumes 1 seq #

Reliable Data Transfer

  • TCP window
    • Not traditional sliding window
    • Used by remote peer to limit the number of unACKed bytes can be sent
  • Out-of-order segments handled by your own implementation
  • Cumulative acknowledgement
  • Piggybacked acknowledgement
    • Data + acknowledgement sent together to save bandwidth
    • Delayed ACK

Retransmission Timeout Mechanism

  • TCP may or may not immediately transfer data upon receiving data from application, triggered when:
    • Amount of data exceeds MSS
    • URGENT data
    • TCP thinks it's time to send
  • Single transmission timer (same as GBN)
    • Cannot depend on RTT - too unstable!

Adaptive Timeout Mechanism

SampleRTT = segment transmission ~ ACK receipt
EstimatedRTT_i+1 = (1-α) EstimatedRTT_i + α SampleRTT_i+1 (typical smoothing factor α = 0.125)
Difference = SampleRTT_i+1 - EstimatedRTT_i
DevRTT_i+1 = (1-β) DevRTT_i + β |Difference| (recommended β = 0.25)
TimeoutInterval = EstimatedRTT + 4 x DevRTT (safety margin to adjust to variations)

Retransmission Ambiguity

When ACK is sent back, is it referring to the original transmission, or the retransmission?

  • Don't update timeout for retransmitted segments
  • TCP doesn't derive the timeout interval on retransmitted segment by EstimatedRTT and DevRTT
  • Instead, double the timeout interval on each timeout (exponential backoff)
    • Data loss mostly due to congestion delay
  • For the next transmission, reuse the timeout value until the ACK for this segment is received

Flow Control

Regulates the rate at which the sender can send data while the receiver can consume without overflow.

RecvWin = RecvBuffer – (LastByteRcvd – LastByteRead)

  • Deadlock may occur if ACK = 4096 WIN = 2048 is lost!
    • Periodically send 1 byte to test the receiver when RecvWin = 0

Congestion Control

To avoid sending data overrunning the network resources.

  • Congestion effects
    • Longer delays
    • Packet loss
  • Principles of congestion control
    • No explicit feedback from network
    • Perceive from observed losses (timeout & duplicate ACKs)

Congestion Window

Reflects the estimated network capacity by the sender.

  • Max. unACKed data at a time: min{RcvWin,CongWin}
    • (LastByteSent – LastByteAcked) <= min{RcvWin,CongWin}
  • Allowed to send no faster than the slowest component

Bandwidth Probing

TCP Congestion Control Algorithm

  1. Slow start
    • Initially, set CongWin = 1 MSS, ssthresh = 64 KiB
    • Double CongWin every RTT (in practice, increase CongWin by 1 MSS for every ACK received)
    • When timeout
      • Set ssthresh = CongWin / 2
      • Set CongWin = 1 MSS
      • Increase exponentially until ssthresh
      • Increase linearly from then on -> congestion avoidance
  2. Congestion avoidance - additive increase, multiplicative decrease
    • Increment CongWin by 1 MSS per RTT (in practice, increase a little for each ACK received)
      • CongWin += MSS * MSS / CongWin
    • When triple duplicate ACKs: likely to be data loss
      • Drop transmission rate by half -> fast recovery
    • When timeout
      • Set ssthresh = CongWin / 2
      • Set CongWin = 1 MSS -> slow start
  3. Fast recovery (+ fast retransmit)
    • Retransmit what appears to be missing without waiting for timeout
    • Drop transmission rate by half
    • Set CongWin = ssthresh + 3 * MSS (3 duplicate packets)
    • When another duplicate ACK received
      • Increment CongWin by 1 MSS
      • Transmit new segment if allowed by the new CongWin
    • When non-duplicate ACK received
      • Set CongWin = ssthresh -> congestion avoidance

results matching ""

    No results matching ""