[c-nsp] Gameserver-Traffic and Cisco-routers - anything I should consider?

Saku Ytti saku+cisco-nsp at ytti.fi
Mon May 2 01:57:00 EDT 2005


On (2005-05-01 21:15 +0200), Gunther Stammwitz wrote:

> Ok there are a few drops but they're minimal. Maybe flow control would help
> to prevent them?
> 
>   Output queue 0/40, 0 drops; input queue 2/75, 217 drops
These drops are packets that were processed by GRP/PRP, ie. not a transit
packet. So they were packets like huge flood of BGP updates or so. You
can say:
conf term
int foobar
hold-queue 200 in 

 To increase the buffer to 200 packets from 75 and probably never see those
drops again, but I have to stress this that this would only explain your
packet loss if someone was pinging _at the_ GSR.

> We're using vrrp in order to provide a redundant gateway to the customer if
> our router fails. The 2nd device is a foundry-router that stay's backup
>(I've checked the logs).

 I think think might explain your drops, if you're running multiple VLANs
in same physical interface and VRRP in each of them that would increase
amount of packets that need to processed by the GRP/PRP, and the default value
of '75' hold-queue was thought of without VLAN's. So I'd encourage increasing
it bit. (But will not solve your problems in transit packet loss)

> Oh.... I've seen that the cpuload on the line card is at 41% which might be
> a little high.
> Any idea why? I'm not using acls. Show proc cpu doesn't show anything
> unusual...

 Why is this high? You're running E1 Line Card (which is well, crap. It's 
software based line card!). ~All of your load (41%) is I/O load, which means
pushing packets. E1 Line Card caps out at ~300kpps, you can fit one
gige port 2.8Mpps (1.4Mpps for both directions). And online games use
extremely small packets and lot of them. 
 E2 and up cards are ASIC based, you can push millions of packets
through them and the I/O load be just about 0% for them (only exception
packets are processed by LC CPU).

> execute-on slot 2 show proc cpu
> ========= Line Card (Slot 2) =========
> 
> CPU utilization for five seconds: 42%/41%; one minute: 41%; five minutes:
> 41%

 To summarise, I don't see anything alarming there, but personally
wouldn't offer this kind of service using that hardware. For low budged
upgrade I'd recommend replacing all E1 cards with used E2 3GE cards. For
higher budged go with new E3 4GE ISE cards.

-- 
  ++ytti


More information about the cisco-nsp mailing list