Re: [v6ops] Thoughts on draft-yourtchenko-ra-dhcpv6-comparison

[Owen DeLong <[e-mail]>]

On Dec 16, 2013, at 13:46 , Ole Troan <[e-mail]> wrote:

> Owen,
> 
>>>>> Second: some of the text you have now deals with link-layer performance. However, I think that since these are configuration protocols, the attributes that are important are primarily semantics, not performance or implementation.
>>>> 
>>>> Lorenzo, this is kind of a puzzling position to take.   There are lots of ways to do things with really nice semantics that fall on their face for performance reasons.   So I think performance questions are in scope.
>>>> 
>>>> Oh, of course - in general, they are. But here the protocols used are so similar that there is little substantive difference in performance.
>>> 
>>> could you expand on that?
>>> I see the scaling properties on a multicast capable link very different between the two.
>>> take a multicast capable link with 10K nodes on it, where power has just come back.
>>> a few hundred RS/RA messages, at least 40K DHCPv6 messages.
>> 
>> I'm trying to think of a multi-user network technology still in use today where 40,000 packets over a couple of minutes time is likely to be a significant load and I'm failing to come up with one.
> 
> indeed, I wasn't thinking of link-congestion, but processing time. on the DHCP server and the first-hop switches and routers.
> do you agree that RS/RA and DHCP scales quite differently in this scenario?

Meh... I think most modern DHCP servers could actually take that kind of load without too much problem. In any real world scenario, 10,000 hosts that are all DHCP clients on the same segment isn't all that likely. However, even if we forego that reality for a moment, they won't likely be homogeneous which means that they will take different amounts of time to get to where they want to make their DHCP requests. Can we agree that the spread would be at least 45 seconds and more likely somewhere in the neighborhood of 120 seconds from the time of the first host requesting to the last?

If so, then you're talking about 40,000 packets (10,000 requests) over 45 seconds = <1,000 PPS. If your switch and/or DHCP server can't handle 1,000 PPS from 10,000 nodes, I'm mystified as to how you expect that network to work when the power hasn't failed recently.

I think in terms of the switch, there are a number of problems that are more impactful than DHCP that would prevent me from attempting to put 10K nodes on a single network segment.

So, yes, I suppose, technically, there is a valid argument that the scaling properties of RA and DHCP are different. However, I think it's kind of the networking equivalent of Zeno's paradox... At some point you have to just accept that close enough is close enough and splitting the hair any further is pointless. I think this goes beyond that point.

> (then we can argue later if the differences matter or not).
> 

I guess I'll concede that there is a theoretical difference, but I remain unconvinced of an actual difference. I'm pretty sure that the theoretical difference doesn't really matter.

To know whether the actual one did or not, I'd have to see an example where it might.

Owen