[c-nsp] Microsoft NLB vs Cisco

[Tim Durack]

On Sat, May 10, 2008 at 4:48 PM, Peter Rathlev <peter at rathlev.dk> wrote:
> On Sat, 2008-05-10 at 12:09 -0400, Tim Durack wrote:
>> Anyone using Microsoft NLB Multicast mode for a cluster?
>>
>> It requires a static arp entry on Cisco, as the cluster ip resolves to
>> a multicast mac, which can't/shouldn't be learned via arp.
>
> I find that a very irritating requirement of the MS NLB. :-)
>
>> So we do something like: "arp a.b.c.d 0100.5e7f.xxyy arpa"
>> Apparently this results in software switching the adjacency on a
>> Sup720, which is painful to say the least.
>>
>> Any suggestions?
>
> I guess you're referring to CSCee49121 "static ARPs dont create adjs
> when used with routes pointing at intf". I thought this was only a
> problem if you used it like this:
>
> ip route 10.11.12.13 255.255.255.255 Gi1/1
> arp 10.11.12.13 030b.adc0.ffee Gi1/1
>
> Is the problem also there without the route statement? We use it against
> two MS NLBs, and we don't see any problems. The traffic doesn't seem to
> be software switched, but apart from consulting Feature Manager and
> looking at the CPU interrupt usage, I'm not completely sure how to check
> it. How do you do it?

No static route - maybe that's the difference.

Educated guess work. CPU is running >90%. Install a CoPP policy
dropping the traffic, and CPU drops back to a more normal ~30%.

Monday I plan to try a SPAN against the rp, and see what is hitting
it. I need this to tune CoPP anyway.

> Regards,
> Peter
>
>
>