7206VXR router with Packet output discards

Unanswered Question
Jun 18th, 2008

My remote site customer has a router that is having an unusually high number of

discard packets compared to other Internet routers with much higher utilization.

Their engineers are thinking that this is due to the fact that they have a 10 Gig

network and Gigabit-connected servers behind the router with a default 40-packet que-depth on the OC3 FIFO interface. They are thinking that it may not be deep enough to hold the packets during brief bursts coming in the Gig interface before going out their 155Mbps WAN interface.

What would be your opinion on changing the output queue depth on router's POS2/0 to a higher value,like 400? The max setting is 4096.

The configuration would be:

int pos2/0

hold-queue 400

!

Their engineers are wanting to know if this would be an acceptable remedy to this

problem or if there is a Cisco-approved number for the hold-queue for an 155Mbps

circuit behind a 10-gig network.

I have this problem too.
0 votes
  • 1
  • 2
  • 3
  • 4
  • 5
Overall Rating: 0 (0 ratings)
Loading.
Joseph W. Doherty Wed, 06/18/2008 - 08:47

40 packets (assuming 1500 coming from Ethernet), would not allow a single TCP flow to ramp up to full OC3 line rate for much WAN latency. Theoretically, the FIFO queue should accommodate the BDP (bandwidth delay product). However, such a large size isn't required for multiple flows, assuming they don't all start at the same time, and such a large queue could introduce highly variable latency.

What you might try would be either activation of WRED on the POS interface and/or CBWFQ FQ.

e.g.

interface pos 2/0

random-detect

or

policy-map CBWFQ

interface pos 2/0

service-policy output CBWFQ

or

policy-map CBWFQ

class class-default

random-detect

fair-queue

interface pos 2/0

service-policy output CBWFQ

Although WRED queue size often defaults about the same as FIFO default queue size, WRED's is elastic. I.e. bursts should pass, only sustained congestion should cause packet drops.

FQ usually defaults to larger queue size than FIFO, but more importantly, each flow gets its own queue. I.e. one large flow burst should have minimal impact on other flows.

rlortiz Thu, 06/19/2008 - 07:15

Joseph,

Thanks for the great input. Will look into your suggestions and give feedback later on.

Actions

This Discussion