Re: Address Family Support in OSPFv3

[Quaizar Vohra <[e-mail]>]

Sina,

The example you take picks a small number of adjs. If you go to
a larger number of adjs, you will run into logical
fragmentation. And once the total adjs grow over 4 logical
fragments, the AF approach and SIN approach will result in
the same number of LSA fragments.

I concede the fact that you could have a large network
where each router has only a small number of adjs. Then
the gain you mention works. But I doubt that anyone
will run IPv4 unicast using OSPFv3 for a very long time.

As far as total memory is concerned, both approaches would
turn out to be quite equivalent in any topology if the
metric values are different across all AFs. Sharing router
LSA links for multiple AFs will be painful from the
implementation point of view.

Also with prefix LSAs you can't share the prefixes between v4 and v6
simply because they are different.

So the gain is O(3) even in the best case. Big deal :). This is true
whether you talk about the amount of memory or amount of LSAs. I feel
that most implementations of the AF approach will end up choosing not
to implement overlapping router LSA links for multiple AFs, simply
because of the implementation complexity. So you will end up
with any gain only in terms of the number of LSAs (not in total
memory). Even the later doesn't work if routers have large number
of adjacencies.

Not worth all the implementation pain :).


Quaizar


 > Quaizar
 >
 > ->
 > ->Sina,
 > ->
 > ->I very much agree with Acee here. Optimization at the cost of
 > ->implementation overhead is justified when the optimization
 > ->results in significant performance gains. The gain with AF
 > ->approach is less than O(2), in fact much less in almost all cases.
 >
 > I do not agree to "much less than O(2)" performance gain and let me get
 > an example to clarify that
 >
 > Let's consider the following
 >
 > - a network participating in IPv6 unicast& multicast and IPv4 unicast &
 > multicast
 > - a worse case scenario where there is No overlap between any topologies
 > - we would consider 1450 byte as the limit for generating new fragments
 > for router LSA, prefix LSA
 > - a router have in average 20 adj in each AF (so 80 adj in total as we
 > considered no overlap)
 >
 > You could see that even in worse case scenario our proposal compared to
 > the multiple instance has a
 >
 > Gain O(4) for each Router LSA
 > Gain O(4) for each network LSA
 > Gain O(4) for each prefix LSA
 > Gain O(4) for each link lsa
 >
 > You could see that we are far away from "much less than O(2)"
 >
 > Sina