Catalyst 6509 seeing output drops in one of the port-channel interface

[zulkarnain]

Hi All.

We have a 6509 and configured 4 interfaces bundle together into 1 port-channel. We are seeing 1 of the port having more output drops than the other 3. What would be the issue? How to make the output drops decrease? If we set the port with "hold-queue 4096" out would it be help? Any impact on the buffer and the cpu of the switch?

[Joseph W. Doherty]

How does that port's usage compare to the other ports in the same Etherchannel?

Realize that if you have a very bandwidth usage intensive flow, it will only use one link, and Etherchannel doesn't take into consideration port usage.  I.e. You can be "unlucky" such that you might have just one or more busy flows that "land" on a particular port, and if it's the same two hosts, especially w/o including 6500 hashing using UDP/TCP ports numbers, the same transmission port will be used over and over.

What might be done, is trying a different hashing algorithm, but improvement is not guaranteed as much depends on the attributes of your traffic.

As to using hold-queue 4K, I forget whether a 6500 really uses that (as much, on switches, is done all with special hardware).  You could try it.  Increasing queue depth often mitigates drops due to transient congestion (like with microbursts), but will not mitigate sustained congestion.  It also can introduce additional queuing latency.

[zulkarnain]

Hi Joseph. 

Thanks for the reply. As of now we have apply hold-queue 1024 output and still seeing a frequent drop. I guess this would be a hardware limitations issue.

[Joseph W. Doherty]

Could very well be.  On a 6500, not all line cards, with same "speed" ports, are equal.

For 6500 line cards, often multiple ports (a bank of them) share the same ASIC (and hardware resources, like buffers).  So, if some ports are very "busy" and others aren't, you can sometimes improve resource allocation by distribution of busy ports.  For example, rather than having busy ports on port 1..4, have them on ports 1, 9, 17, 25, etc.  Also, have busy ports on "better" (there are reasons Cisco lists some line cards for users vs. servers) line cards and/or distribute across line cards.

Lastly, with the 6500 series, you also have to understand how best to use classic, fabric and classic/fabric line cards, within the same chassis.  (There also the question of whether DFCs, on line cards, would mitigate some drops.)