A Protocol for Packet Network Intercommunication

One-line summary: Description, assumptions, and justification for proposed design features of IP and TCP, given the philosophy that the internetting protocol's goal should be to make connection of heterogeneous networks as easy as possible.

Overview/Main Points

• gateway and gateway routing based on source and dest addresses
• TCP port demultiplexing
• fragmentation and heterogeneous MTU's
• end-to-end reliability
• window based end-to-end flow control
• end-to-end message reordering based on sequence numbers
• building "logical" connections using connectionless network by exchanging initial parameters and maintaining state at each end

• gateways should insulate networks from link-level addressing differences, MTU differences, performance differences
• end-to-end reliability needed for hassle-free internetting, to hide varying reliability in different nw's
• routing: next hop gateway should be "derivable from" source and/or dest addresses; in the section on TCP, they describe the predecessor to the masking scheme currently used
• allow for nw's with different MTU's rather than enforcing a global minimum MTU, because:
  • want to isolate internal packet size parameters across network boundaries
  • encryption, associative addressing, etc might cause a "legal" packet to grow unpredictably larger than MTU
  • difficult to increase "global minimum" MTU (inertia, installed software base, etc.) to exploit new technology
  • Reassembly should not be done at gateways: serious buffering and state problems, plus requires all fragments of a packet to pass through same gateway (reduces flexibility in routing)

• Goal: demultiplexing of messages to different applications on a host. Proposal: use 16-bit port number per "stream".
• Interesting application: authors mention that if communicating parties agree in advance on a "port hopping sequence", could use for privacy by obscuring which messages belong to which stream.
• Streams to different ports shoudl not be multiplexed into a single TCP-level message, since this introduces artificial inter-process dependencies at receiving host (packet parsing, retransmission, etc.)
• Routing: original proposal was that a "network address" be an 8-bit network number plus 16-bit TCP port number. Gateways demultiplex TCP port numbers to applications, but authors allow for possibility that machines in local network understand the internetwork addressing scheme.
• Message reordering: use sequence number unique to pair of communicating ports.
• Retransmission after timeout: "It is our expectation that ... [it] will not be called upon very often in practice"; authors felt that IP would mostly be reliable except under severe congestion.
• Important. Careful discussion of what a "connection" really is, in light of evidence (ARPAnet) that connectionless protocols can be constructed; derivation of the current TCP meaning of "connection" as shared state using a connectionless network as medium. IMHO, a major leap of insight.

Relevance

Flaws

• IP and TCP not as well separated as they became after the OSI 7-layer model was "imposed" on TCP/IP, though the paper does take care to keep routing and "connections" logically separate.
• Sometimes ideas and concepts are presented, other times implementation suggestions or data layout.
• It's mentioned that packet accounting will be needed since data will be traversing networks of different ownership, but this point is never returned to in detail. Current evidence seems to suggest it has been de-emphasized :)