[c-nsp] 1841 suitable for BGP?

[Anton Kapela]

> > 2. the CEF path somehow executes at a higher priority
> >   than processes like BGP and process-switching
> >   (huh, is process-switching an actual process?);
> >   hence CEF-switching performance would be less
> >   affected by ordinary processes (compared to
> >   process-swiching)
> 
> That is exactly the case.
> 
> CEF is interrupt-based forwarding, ie it'll interrupt any 
> running process and do packet forwarding, when there are 
> packets to forward in fast-path.

Indeed.

For all those 'stuck' on cpu-forwarded platforms (vs. tcam, etc), useful
knobs such as 'scheduler allocate' and 'process-max-time' exist to make
using these platforms more amicable while used interactively. 

I generally set a equal share for network int time & process time on
faster boxes (npe400 and newer) and for progressively lower end boxes,
usually 2x the network interrupt time allowed for process time. Also,
process-max-time of ~50ms seems to be reasonable, although others have
warned of stability issues on certain S train releases while using
smaller intervals. The net effect of said adjustments is that the CLI is
always reasonable via ssh amidst bgp scan timers of 30 seconds, fast
ospf hellos, and many kpps flowing. Even with these knobs, IOS won't
ever feel like a RTOS, so don't expect night/day results.

I would have imagined most folks here have fully reviewed the available
documents from Cisco regarding these two commands and their use, but
after reading a few of the previous messages in this thread I suspect
the following URL's would be of broader interest:

http://www.cymru.com/Documents/performance.html

http://www.cisco.com/en/US/tech/tk648/tk361/technologies_tech_note09186a
0080120f48.shtml#router_self_protect  

http://www.cisco.com/en/US/products/sw/iosswrel/ps5187/products_command_
reference_chapter09186a008017d03b.html#wp1061953

-Tk