[c-nsp] High CPU util on Cat4506

[Phil Mayers]

Jee Kay wrote:
> On 20/11/06, Jee Kay <jeekay at gmail.com> wrote:
>> I've got (very) high CPU util on one of my Cat4506s (SupIV running
>> 12.2(20)EW1). I have a strong suspicion that it is being caused by the
>> switch doing a bit of multicast forwarding (~20-30Mb/s).
> 
> This appears to be because some of the high-rate sources are stuck in
> 'Registering', which I assume means every packet gets punted to the
> CPU to get encapsulated and shunted off to the RP.
> 
> The RP is an anycase RP consisting of a pair of Juniper M7is.... Not
> all sources in a particular group are affected, and not all sources
> from a particular interface are effected.
> 
> Anyone have any good ways of tracking down what looks like a PIM
> Register problem?

What is the size of the data packet being encapsulated? You may be 
seeing MTU issues.

The old PIM spec was ambiguous on this, but the newly issued PIM spec 
instructs designated routers to do this with "large" packets:

IP + PIM + register + IP(id=xyz, morefrags)
IP + PIM + register + IP(id=xyz, frag=14xx)

i.e. the tunnel between the registering router and the RP is to be 
treated as link with MTU = pathMTU - sizeof (IP+PIM+register)

Packets with dontfrag set are to be ICMP-frag-needed back to the source 
by the designated router (yes you did read that right, yes a router is 
now required to return an ICMP in response to multicasts in some limited 
circumstances)

Some cisco routers (certainly 6500 + sup720 + ios 12.2 SXF2) do not do 
this - they do this:

IP (id=xyz, morefrags) + PIM
IP (id=xyz, frag=14xx) + PIM (tail of payload)

The Juniper WILL drop them. We've seen CPU spikes as a result. Setting 
register-rate-limit helped, but ultimately a fix from Cisco or 
increasing the MTU between the DR and RP is the only fix.

I have been lax in not opening a TAC case about this.