Fwd: Re: Reed's views, was [e2e] Cannara's views

[Steve Deering]

At 10:00 PM -0400 4/16/01, John Day wrote:
>All of that reflects our understanding at the time, that Ethernet
>addresses would have to be assigned and be globally unique, just
>as IMP addresses were and later IP addresses.  It was only later
>that we realized that data link (and Ethernet addresses) only have
>to be unambiguous within the scope of the layer they are in.

Perhaps that was your (mis)understanding at the time, but it is very
unlikely to have been the understanding of the designers of DIX 10 Mbps
Ethernet, given that there already existed other LANs (such as the
Xerox 3 Mbps Ethernet) that worked fine with only locally-unique
addresses, including serving as part of higher-layer networks
(such as the IP and Pup internets).

>...Consequently, not only did they not [need?] 48 bits, they didn't need
>to be centrally administered either.  In fact, they only needed to be
>administered and assigned by the owner of that particular Ethernet
>segment.

*Avoidance* of the local administration hassle was one of the main
benefits of the Ethernet addressing scheme.  A factory-assigned,
globally-unique serial number satisfies the local uniqueness property
required on a given LAN, without requiring users or network managers
to assign addresses, or the use of protocols to dynamically generate
locally-unique addresses, both of which approaches are more error-
prone and resource-consuming.  I find it hard to believe that this
was not well understood and an important reason for the choice of
the 48-bit addressing scheme (as opposed to your conjecture that
the DIX Ethernet designers made the address bigger than necessary
because they misunderstood the uniqueness-scope required for a LAN
address).  If only I had a decent filing system that allowed me to
find my yellowing paper copy of the mid-80s article that gave the
rationale for the 48-bit design, I could confirm that...

Steve