Retransmission Protocols

• Send Windows
• Go-Back-N
• Selective Repeat Protocol
• bandwidth-delay product
• TCP retransmission
• Kermit
• Xmodem

Send Window

• send window is how many outstanding bytes or packets sender can have
• if window = 0, sender must stop
• if the number of outstanding packets (bytes) = window, sender must stop
• window = packets buffered by network + packets buffered by receiver
• networks with high latency require large windows
• receivers with small amounts of memory require small windows
• a large window means less coupling between sender and receiver
• flow control: receiver can slow down sender

Go-Back-N

• like ABP, but window size W > 1 (if W = 1, GBN = ABP)
• sender sends up to W packets, each with sequence number
• receiver sends ack (with sequence number) for every in-order packet
• receiver discards out-of-order packets
• when sender gets ack i, it sends packet i+W
• if packet i not acked within timeout T, sender retransmits i, i+1, ..., i+W-1
• if it takes time t to send a packet, T >= W * t
• since there can be W outstanding packets, sequence numbers must have range at least 0... W

Selective Repeat Protocol

• like GBN, but only packets that are unacked at timeout are resent
• more efficient of network usage: doesn't need to retransmit entire window on timeout
• requires more receiver resources: receiver must buffer out-of-order packets and deliver them in order once the missing packets are received
• sequence numbers must have range at least 0... 2W-1, since we need to distinguish retransmissions due to lost acks from new packets (packets in the "last" window from packets in the "current" window)

Bandwidth-delay product

• window = packets buffered by network + packets buffered by receiver
• receiver knows how much it will buffer
• how many bytes/packets are buffered by the network?
  • bottleneck router is the slowest router in the path
  • bottleneck router can forward B bytes (packets)/second
  • total latency, from sending packets to receiving acks, is d seconds
  • bandwidth-delay product is B * d bytes/packets
  • B * d data is buffered by the network

TCP retransmission

• similar to selective repeat, but
• acks are cumulative and network not FIFO, so
• we don't exactly know which packet is dropped
• instead, receiver sends duplicate acks when receiving new out-of-order packets
• after 3 duplicate acks, TCP retransmits next packet
• packet losses that don't get 3 duplicate acks:
  • packet loss at end of data stream
  • packet loss when window is small
• timeout in TCP is adaptive, based on measured network latency

Correctness of GBN, SRP, TCP

• will sender always retransmit a dropped packet?
• can receiver always distinguish retransmission from new packet?
• does receiver have enough information to deliver packets in order?

Kermit

• file transfer over point-to-point links
• 1 byte each for mark, length, sequence, type, 0-94 bytes of data, 1 byte checksum
• printable ASCII characters for all control fields
• type field identifies packet: ack/nack, connection establishment/teardown, etc.
• length identifies data length
• sequence is modulo 64
• ABP on half-duplex connections
• optional SRP on full-duplex connections

XModem

• 1 byte each for mark, sequence, complement sequence, 128 bytes of data, 1 byte checksum
• redundant sequence, checksummed data
• ABP on half-duplex link
• receiver sends nack after 15 seconds of idle
• sender sends right after nack
• if packet received correctly, receiver sends ack
• if packet has errors, receiver sends nack
• acks and naks have no sequence numbers
• sender needs no timer:
  • unlikely total packet loss
  • recover after 15s