[c-nsp] high cpu usage due to interrupts

[Anton Kapela]

> > So at least for unicast seems that is working, I assume that the 
> > output path cannot be forwarded with the VIPs...
> 
> No, that's not right. In a well setup dCEF 75xx 99% of all 
> packets would be dCEF switched that are transit through the box.

Right, but let us not forget this tidbit:

> 2 AIP controllers (2 ATM).

AIP is not dCEF capable (as it's not vip-based) and anything switched to
it or from it will always punt to the RSP, _regardless_ of the ingress
interface. So, packets entering the router from the Fast-E PA's on the
VIPs' won't ultimately be dCEF switched to the AIP; they'll always take
a trip through the RSP. This isn't to say the punts are 'bad' - they're
not 'horrible' so long as they can be cef switched once they arrive (and
you have enough cpu to do it at the rates you require..) - but clearly
this is why you burn CPU with traffic through your box whenever it
ingresses or egresses significant traffic on your ATM link.

> IOS (tm) RSP Software (RSP-JSV-M), Version 12.1(14),

Is this supposed to be a joke? :) Grab some S-train (especially if all
you need is basic v4, mcast, and netflow). Personal favorite is
12.0(32)S lately, (28) also has been reported by colleagues to be stable
as well.

Also, RSP4's can forward all the bits you'd probably ever need and
appear to be under $100 on eBay, more like 60. Check
http://tinyurl.com/39zowz for an example.

-Tk