[c-nsp] increasing input-hold queue

[Rodney Dunn]

The Input queue is simply a way to prevent any one ingress
interface from overwhelming the system buffers reserved
for packets going to process level.  You have a fixed
number of system buffers (that can grow and trim dynamically)
but at a given instant in time if you had a flood of packets
from multiple interfaces a single interface can't have more
than the input hold queue depth waiting to be processed.

So before I suggest anyone tweaking input or output hold
queues or buffers I want them to prove to me they
understand why the default isn't sufficient.  And to
do that you have to answer what packets are getting
punted to process level if it's an issue with the input
hold queue.

sh buffers input-interface <name> packet

will tell you that.  You have to do it over and over
to catch packets getting punted.

sh int stat   will give you an idea of the % of packets
that are being switched under iterrupt vs. punted.

A 7609 (assuming sup2/msfc) should process all
traffic in hardware (except valid punt traffic) so the
default queues should be just fine because the MSFC
CPU should have plenty of headroom if the hardware is
doing all the packet switching.

If it's the output hold queue you either have:
a) congestion
b) a bug

Same goes for buffer tuning. 
I have only seen one time that I suggested
a customer do buffer tuning after working in TAC for 6.5 years.
It was a bridging environment
with tons of PVC's and they needed extra buffers to hold
the replicated packets.  Most people tune buffers incorrectly
and to mask an underlying problem with a bad design.

I think a RPM-PR is the same as a 72xx so it's no different
than an MSFC in regards to how it handles packets are
process level.

Overruns, discards are not a process level problem and
therefore without more information about the segment
type, utilization, etc. I would no suggest tuning anything.
The system may be working just fine as it is.

All that rambling being done :) I usually tell people to
double the queue depth and monitor.

Rodney

 
On Thu, Nov 11, 2004 at 03:43:47PM +0200, Kim Onnel wrote:
> Dear list,
> 
> Using output interpeter ony my <sh int x> result, i get alot of
> overruns, discarded packets, so while looking around, i found that i
> should increase the input hold-queue, however, i didnt find any
> informaition regarding how much should i increase it, if there is any
> adverse effect, and such,..
> 
> Any ideas on a Cisco RPM-PR  , 7609 ?
> 
> Regards,
> 
> -- 
> ~Kim
> _______________________________________________
> cisco-nsp mailing list  cisco-nsp at puck.nether.net
> https://puck.nether.net/mailman/listinfo/cisco-nsp
> archive at http://puck.nether.net/pipermail/cisco-nsp/