Re: rfc4941bis: On the use of multiple addresses

[Lorenzo Colitti <[e-mail]>]

On Fri, Jan 31, 2020 at 3:06 PM Pascal Thubert (pthubert) <
[e-mail]> wrote:

>  Limiting the amount of addresses is an obvious implication of SAVI in a
> managed network. When we did RFC 8505 the Sec-Dir review insisted to have
> text  to protect against DOS attacks. The SAVI draft did not go through
> that full cycle and is not complete to that regard. So I think the original
> text was a lot better.
>

The fact remains that if the SAVI RFC contains no text in support of a
statement, it should not be cited in support of the statement. Maybe we can
say that if address tracking technologies are in use, and network devices
run out of memory to store address mappings, then creating further
addresses will result in some of them not being usable. This is not a
policy issue, it's a scaling issue.

And the size of the cache is mostly irrelevant to the ND churn these days.
> When addresses are dropping for the lack of rmemory there’s really an
> issue.  What’s relevant is the amount of new addresses joining the network,
> the number of nodes in the network that need to resolve it, and what nodes
> do when they sleep cycle or roam.
>

Ok, but that text is too vague. What specific problems would you like to
mention here? From the two lines you just wrote it sounds like when nodes
create many addresses, there is lots of ND traffic, and:

   1. Networks that inspect all ND (e.g., if SAVI in use) can run out of
   CPU cycles because there is too much ND traffic.
   2. If there are very battery-constrained devices on the network, the
   power used to listen to ND packets can meaningfully impact battery life.
   (In the real world, this does not happen to laptop or phone class devices,
   but it can probably happen to lower-power devices.)

We measured (on the wire) several hundred multicast per second in average
> over 90 minutes during a keynote with a large audience. /64 per host is
> indeed tempting in such environment. But it boils down to IPv4 with 8 bytes
> addresses. We can do better.
>

500 pps of multicast sounds like a lot, but if the network is very large,
well, ND scales linearly with the size of the network (assuming L=0 which
for a network with thousands of hosts it pretty much has to be). The only
meaningful measurement here is what percentage of airtime was used by ND,
and that shouldn't vary too much from network to network. As for better,
well, it depends. From a host perspective, being able to create addresses
without having to ask the network is desirable.