> Okay.  How about: Push is simple and known to work; but it costs.
>  Pull is harder, but may be cheaper.  Still a very broad statement,
>  but it allows the following question:

I think it's not particularly useful to have a high-level discussion
about whether something as general as "push" vs. "pull" will work or
not. Push vs. pull are just two approaches. They both have plusses and
minuses. To decide which makes sense (and in what context), you have
to look at the overall problem, requirements, and a whole bunch of
factors. And different folk may weigh various factors differently.

I'm not saying push can't work. In fact, I'm sure it can. But just
because it can work, doesn't mean other approaches should be ruled
out.

> Is the savings (in whatever dimension) that Pull offers over Push
>  worth the effort?

I think so. At least, I don't want to see it ruled out up front.

> >> I could ask you what happens to packets while the Pull is being
> >> responded to, or a bunch of related questions. I won't.
> > 
> > They get queued. Or dropped. Or possibly something else. Yes, there
> > are implications to that. But not necessarily a show stopper either.

> It was a show-stopper for Ipsilon.

Now there is a soundbite. But I'm not sure exactly how that applies
here...

> >>> In my view, this puts an unnecessary load on NVEs.
> > 
> >> Let's talk instead about the "unnecessary load". Can someone quantify
> >> this?  Is it CPU? memory? messaging? What's the bottleneck or pain
> >> point?
> > 
> > Some or all of the above.
> > 
> > If typical VNs are smallish, I agree that an NVE can preload full
> > tables with no problem. But what about for very large VNs? Should the
> > architecture *force* such preloading of full tables, even if the
> > working set of routes is actually very small?

> I thought I picked a fairly large VN, with 10,000 members.  Even if
> you take a larger one, say 100,000 members, the memory cost is (imo)
> small.

It's not just memory, there is also the signalling cost.  And if you
have 10K VMs, no doubt at any one time some (more?) are moving around,
with the corresponding signaling (using push), going out all the way
to *all* NVEs.

> <snipped, see next email>

OK.

> > Note also that we may be looking at the problem from different
> > perspectives. For example, in a single data center, I can imagine a
> > centralized directory service holding the complete address mapping
> > information for all the VNs in the DC. An NVE in such cases can query
> > such a mapping system with very very low latency.

> Perhaps so (I would love to get more details.)  But what did you
  gain?

Um, some large data centers already have centralized components. Like
the entire VM orchestration system.  Adding more to that
centralization doesn't necessarily change the single point of
failure/risk factor in a substantial way. E.g., in the ARMD
discussions, one large operator said they'd like to centralize more of
that info, and reduce the signaling (which was effectively
redundent). There was little benefit in dynamic use of ARP, for
instance, when all addresses and their locations were already known
centrally.

Thomas

_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3

Reply via email to