[c-nsp] High SNMP CPU with SXH. Is SXI any better?

[Dale W. Carder]

Hi Jeff,

On Dec 12, 2008, at 9:55 AM, Jeff Fitzwater wrote:
> 	We are running 12.2SXH2a on sup720-CXL and have been having  
> consistently high (80-100%) CPU on route processor when retrieving  
> either the ARP table or the Bridge-mac table.   No matter what  
> program we use HP NNM or an snmp script, the CPU route process goes  
> from 20% to 90+% with the SNMP process being the top dog when you do  
> a "sho proc cpu sort".

We see this too.  I'm guessing you also have a non-trivial
amount of directly connected hosts?

> It appears that internally the route processor is doing a lot of  
> crunching to get this table data, specifically the ARP and Bridge  
> Mac table.   I remember something about the format it's in and it  
> had to be converted when retrieved with SNMP.

See RFC 1905 4.2.2(1) which requires lexicographical ordering of
retrieved values.

So, if IOS stores the arp/cef/whatever datastructure in memory in
any other format, which seems likely, it would have to sort the
table every time to spit it out via snmp.

Now, this is of course compounded by the sup720 RP having a
processor that lags behind current commodity chips by at least
6 years.

> Q.   Does anybody know if there is any change with SXI and SNMP  
> queries?

I wouldn't expect anything to change unless the sorting algorithm
were dramatically improved or unless IOS specifically maintained this
table in a better fashion.

Maybe if Cisco hadn't alienated their customers with the 6500/7600
split, you would have an RSP720 today.

> I also remember reading something about a different way to retrieve  
> this data locally on the router and push it to a host, but cannot  
> find any reference to it now.  Any ideas on this?

I haven't looked into it, but perhaps you can find a cisco specific
mib, maybe cef or mls specific that doesn't have this performance
penalty?  Otherwise, I bet a query via clogin outperforms the snmp
table.

Dale

--
Dale W. Carder - Network Engineer
University of Wisconsin / WiscNet
http://net.doit.wisc.edu/~dwcarder