Re: Neighbor Unreachability Detection is too impatient

[Mikael Abrahamsson <[e-mail]>]

On Wed, 25 May 2011, Philip Homburg wrote:

> So you are saying (with a bit exaggeration from my side) that some links 
> in data centers have so many hosts that a network disruption of a few 
> seconds will cause a multicast storm big enough to meltdown the network.
>
> With modern CPUs and a Gbit/s network that seems very odd to me.

I have an example from reality:

At AMSIX, there are a few hundred routers on the same L2 network. 
Connected to this we have a router from a major well known vendor. This 
vendor does not filter IPv6 multicast messages on the linecard, but 
instead punts all of it to RP, and AMSIX L2 infrastructure treats all 
multicast as broadcast, so all multicast (mainly ND traffic) is sent out 
on all ports.

This traffic to the RP was enough to cause IPv6 BGP sessions to go down 
because it was continous ND messages coming in all the time (few hundred 
PPS).

ND is chatty, and it does cause problems on large networks. One might call 
this multiple faults on multiple levels (bad linecard code, bad L2 
platform, bad low-performing CPU on the RP), but it does cause problems in 
real life regardless.

-- 
Mikael Abrahamsson    email: [e-mail]