Re: LL32 Multihoming

    Date:        Fri, 1 Aug 2003 12:33:16 +0100
    From:        "Philip Nye" <[email protected]>
    Message-ID:  <006601c35820$b38d1030$131010ac@aldebaran>

  | Particularly as
  | there had been a suggestion on this list that consideration of the hardware
  | addresses of other interfaces might somehow came into the algorithm - which
  | I would say contravenes independence.

Whether or not that suggestion is adopted, no, it doesn't.

You're also being confused about an independent execution of the algorithm
for each interface, and the interfaces being independent.   Those do not
mean the same thing.

To make this clear, consider two brothers applying to enter some club
(or college or something).

If the entrance admission algorithm is run independently for each
brother, then each of them stands an equal chance (depending upon
whatever other issues there are) of being admitted.   That does not
in any way prevent the criteria for admission containing "has no
brothers" or "has no twin brother" or "has no brother who is married"
(in the latter case, it is possible that one of the two may be eligible
for entry, and the other not be, but those are still based upon
independent executions of the algorithm).

On the other hand, if the algorithm isn't run independently for each,
then there can be rules like "brothers cannot be admitted together, the
tallest of multiple applicants is to be selected".  To test for things
like that, you have to know that the two are both applying at the same
time, andthe characteristics of each application, one application can't
be evaluated without knowing whether or not the other exists, what it
contains, and what its result is (in that admitting one of the two,
would automatically result in the other being excluded, and which is
which depends upon what is in the applications).   In that case, the
algorithm isn't being run independently on each brother.

Note, that the similar rule "no admission if a brother (or any relative)
is already a member" is OK - that doesn't result in any kind of
cross-dependence in the algorithm, as each application is tested, the
membership list is simply consulted to see who already exists.   That
may result in the 2nd of the two being rejected, but that is still
based upon independent algorithm execution.

In none of these cases are the brothers independent of each other
(necessarily anyway) - you can assume they are highly co-dependent.
That has nothing whatever to do with the independence of the
algorithm used to make a decision about them.

For LL addresses, and their assignment algorithm, all this means that
the algorithm to select an address on an interface works, according to
its rules, on one interface without caring what another instance of
the algorithm happens to be doing on a different interface.   That means
there cannot be rules like "if you get a conflict with the address
being tested, then go back and redo the address assignment on the other
interface" - that wouldn't be independent.  On the other hand, you can
have rules which say things like "look at the mac addresses for all
interfaces in the system", as that affects nothing.   You can also "look
at all LL addresses assigned to the system [and don't duplicate one, or
whatever]" that is also OK for independent execution.

kre