Rational for

                        Variable Length

                           Addresses


                         Peter S. Ford
                 Los Alamos National Laboratory
                         peter@lanl.gov
                        +1 505 662 3761

                          Toronto IETF
                          26 July 1994

========

Outline of Talk

  Get some scope of the problem domain

      What we know

      What we don't know

  Look at the design space.

  Why Variable Length Addresses (VLA) meet these requirements w/impact
  on IPng evaluated.

========

What We Know

  Datagram Network, source and destination addresses

  Packet Forwarding Functionality
      longest prefix match
      some kind of source routing appears likely

  Host functionality
      address used for lookup of TCB, speed is important (hash functions, 
      etc.)
 
  Bits are not free (bandwidth and memory)

  Addresses are used in a variety of environments:

      serverless vs serverfull autoconfig
      private vs global nets
      local vs global traffic
      source routing through mesh of providers

  Need to support very flexible address assignment, history indicates that
  any inconvenience in this area results in private address spaces (they opt
  out).

========

What is Encoded in Addresses

      Unicast addresses 
          end systems
              global (routing structure coded in address)
              local (fairly flat--ap and go)

      Routing domains (cluster addresses)
              used for domain by domain routing (provider selection)

      Multicast addresses

  Design of address length has been trivialized to:

            "one size fits all"

  (and what is the size?...)

========

What We Don't Know

    number of systems?
        hosts
        processes
        disposable devices (phones, consumer electronics, set top box, 
          etc.)
        providers
        ??? what is NL addressable

    shape of addresses
        # of nesting/levels for providers
        # of nesting/levels of address admin
        local structure of routing
        flat vrs hierarchical vs ??? routing
        delegated addressing plan vs single integrated plan (bottom up 
          vs top down)
         
  What is needed for forwarding the packet to its destination's location
  vs how you refer to/identify the object?

========

Design Space

  We are discussing the design of the address field in a packet, this is 
  NOT the same as the lengths of addresses.  VLA choice makes this field
  variable and extensible to meet a variety of requirements.

  VLA still means you need some answer for what is the maximum length
    required to meet requirements.

  Gains due to more efficient coding?

      AS numbers do not need to be the same length as unicast addresses
        cluster adds << globally unique unicast

      Local use addresses can elide representing the outing structure
        in globally unique and routable unicast addresses (this is the 
        common case for cable TV, parallel supercomputers, etc.)

========

One of the Following Assertions is True

    16 bytes is 
        enough forever
        enough until techno shock
        enough for 20 years
        NOT ENOUGH

    What if it isn't enough?

        Building in limits where they are not needed does not meet
        the need for being sure that it will work forever.

        Cost of using VL addresses is small, for each bit of coding you
        get a doubling of size.

========

Unfounded Claims

  variable length runs noticeably slower in routes

      - empirical evidence does not agree with this

  variable length addresses result in slow host implementations

      - need to effectively feed TCB lookup, can use hashing, binary
        tree search, etc.

      - structure of N+8 (locator + SYSID) where

          N = n * locator_extent_size (for n=0, 1, ...max)
              and locator_extent_size = 4 or 8

          gives you nicely aligned addresses that are a "natural
              computer" length

  Variable length addresses are resource wasteful:

      Compared to what?
          12 bytes of zero fill in an address?
      VLA is bandwidth and memory efficient both on the wire and
      in memory on routers and hosts.
      You code what you need to code.

========

Variable Length Encoding is Hard to Code

      Should we use fixed size packets

  Don't know how to extend addresses using VLA (the extend to the 
  right issue)

      Let's say we use 8 and 16 byte addresses.
      First 4 bytes of both can be the same for the same provider 
        (this is an administrative choice).

========

Summary

  VLA provides a flexible encoding scheme to meet the requirements and

      INSURES THAT WE WILL NOT RUN OUT

  Hard to imagine we are building old bugs into the new system.  
      16 byte fixed width choice is brittle and unnecessary.

  VLA can match requirements for encoding a variety of addresses
      in the packet field.

  VLA is efficient in matching these requirements in terms of 
      bandwidth and memory use.

========

2^128

IS NOT ENOUGH CUSTOMERS