Re: [RAM] Re: the separation of ID/RLOC

[jnc at mercury.lcs.mit.edu (Noel Chiappa)]

    > From: Christian Vogt <christian.vogt at nomadiclab.com>

    > from a delay perspective, a pure pull model won't support efficient
    > provider fail-over. A push model might; it depends on the actual
    > mechanism.

In a way I would have thought just the opposite, actually - but in actuality
I'm not sure it makes any difference. Here's why.

First, I would think pull might be faster, because in the push model, when
the binding from identifier (in the sense 'location-independent name') I to
locator (in the sense 'location-dependent name') L changes from L1 to L2,
that knowledge starts near L1/L2 and has to propogate across the network. My
sense in that in a push system, which has to propogate that binding to
everyone in the network, that might happen slower than a pull system, in
which only people who want the updated binding have to get it.

Second, I'm not sure that matters, though, because when that binding is
updated from I->L1 to I->L2, a lot of boxes out on the far side of the
network are going to have the old I->L1 binding cached, and ongoing traffic
is going to be trying to use it - and of course it won't work. So, just as in
any caching scheme, you have to have a mechanism to invalidate cache entries
when they are outdated.

I suspect that that cache-invalidation mechanism will be the dominant one, in
the sense that most entities which are communicating with I are going to find
the new binding through the cache-invalidation mechanism (whatever that is),
rather than through the normal distribution mechanism (whether push or pull).
So my sense is that the push/pull difference is not going to be significant.


Now, you do have an interesting question on the cache-invalidation issue; do
you depend on an explicit detection-update mechanism (i.e. distant box D
sends packet to I, arrives at L1, I is no longer there, L1 sends an error
message to D with L2, and D updates its cache), or whether you use a push
mechanism to update everyone (which is basically what classical routing does).

Pull plus cache-invalidation probably has less total signalling overhead than
push, but push has the advantage of not having the resolution latency when a
new communication starts. (Push plus cache-invalidation probably has the
fastest response of all, but the highest signalling overhead.)

	Noel

_______________________________________________
RAM mailing list
RAM at iab.org
https://www1.ietf.org/mailman/listinfo/ram