Transport: elements of transport protocols

Elements of transport protocols Addressing

• NSAP (network service access point) — IP address

• TSAP (transport service access point) — port

• some services are always on the same ports (e.g. mail server on TCP 25, ssh on TCP 22)

• portmapper process

  • finds TSAP port for a specified service name, using its database

  • a new service registers itself with the portmapper

  • kind of like directory assistance phone operator

• initial connection protocol

  • each machine that offers services has a process server (inetd on UNIX)

  • this acts a proxy for less used servers

  • listens to set of ports at same time, waiting for a conn request

Connection establishment

• using sequence numbers for packets

  • label packets with sequence numbers that will not be reused within T seconds

  • packet hop limit used to remove old packets

  • time-of-day clock used to choose sequence number

• three-way handshake

  • one peer checks with other if connection request is current

  • e.g. host 1 chooses sequence number x and sends it in CONNECTION REQUEST to host 2, host 2 replies with ACK containing its own sequence number y, host 1 acknowledges host 2’s sequence number in first data segment it sends

Connection release

• asymmetric release — one host gtfo, other host is left hanging. can lead to data loss.

• symmetric release — participants agree to end connection

  • two-army problem — one white army vs two blue armies. one white beats one blue, but two blue beat white if attack at same time. can they coordinate attack using messengers? No.

• one way to semi-solve this is by timers and only allowing a certain number of timeouts, but y’know…it’s not great

Flow control

• slow down sender if receiver can’t handle the data rate

• received packets are buffered at the receiver

• buffer management is done separately

• available buffer space is receiver window size

Crash recovery

• recovery from layer N crash can only be done by layer N+1 (higher abstraction)

• but only if layer N+1 has info that lets it reconstruct previous state

• e.g. if there’s a problem on the transport layer, the transport layer’s lazy ass leaves it to the application layer to fix its problems

Congestion control

• needed to slow down sender if network can’t handle the data rate

• bandwidth allocation

  • max-min fairness

    • maximises minimum bandwidth, then uses excess where possible

    • bandwidth given to one flow can’t be increased without decreasing bandwidth to another flow

  • amount of bandwidth needed for a connection varies, so need to converge to ideal point

  • sharing bandwidth between users

  • dynamic adjustments with AIMD — convergence:

• problem is, TCP data loss means congestion, and wireless loses data commonly. so wireless must mask data loss