3750: High CPU Utilization...Causes?

[w-sims]

I have a 2 switch stack of 3750-12S That have High CPU utilization primarily with the IP Input process. Any suggestions as to causes?

[royalblues]

can you please post the output of show version and sh processes CPU.

Narayan

[w-sims]

Attached is the requested info.

[glen.grant]

http://www.cisco.com/en/US/products/hw/routers/ps359/products_tech_note09186a00801c2af3.shtml

[pjosefson]

Did you ever resolve this? I have a customer with similar problems, and I *think* I've narrowed it down to worm-infected clients trying to infect random addresses (leading to LOTS of new flows having to be routed by the CPU).

[glen.grant]

That would certainly do it if you have infected clients . If this is the case I would find out why the antivirus isn't working on their machines .

[pjosefson]

As I would... unless they were CaTV broadband customers - I have no desire to instill into them the notion that I am available around the clock for support... or take any responsibility whatsoever for their antivirus solutions... :-)

Waving with my customer's AUP might do the trick, though, but as their staff is 1.5 people, I don't think they care much for spending time waving...

However... I DO suspect a middlebox (NetIntact PacketLogic) is doing some sending of gratuitous RST:s... meaning that even updated Windows machines maximum-connection-per-host is thwarted. I so do NOT like middleboxes (esp. since I didn't sell it in the first place)...

/Peter

[w-sims]

Is your client switch/routing Tivoli Service Manager (TSM) across this switch?

[pjosefson]

Not that he knows of. Broadband customers - individuals/households as well as commercial (problems most severe at night, when the commercial customers don't do much, so it shouldn't be them). But you tickled my curiousity. Why?

More info: The network is a mini-MAN in a town with almost exactly 10000 residents. The 3750 cluster is the core L3 equipment, access layer is L2 switches.

We MAY be oversubscribing the TCAM, as we have almost a hundred VLAN:s running L3 (as opposed to the 8 in the example in the sdm template). On the other hand, 100 / 8 = 12.5 and we are nowhere near 12.5 times the numbers supported by the template. We more or less only have a default gateway (upstream ISP) and then each directly connected VLAN.

Also, I feel I would see interrupt load then (as CEF in CPU would show as CPU load), not IP Input. Right?

/Peter