Packet Drop 2950 vs 3560

paul.l.kyte · ‎10-20-2009

I have a problem on a 3560 with regards to packet throughput between an incoming 1Gb port and outgoing 100Mb port.

No matter what configuration I put on the 3560, QoS and otherwise, a burst of traffic coming from a device on the 1Gb port destined for a device on the 100Mb port always suffers from packet drops.

If I replace the 3560 with a 2950 with incoming 1Gb and outgoing 100Mb no such packet drops occur.

Can anyone tell me why the 3560 cannot perform as the 2950 does.

The burst of traffic is 30Mb of udp traffic, each packet being 1494 bytes, and lasts for 3.5 seconds. When analysed this is transmitting at about 70Mb per second.

The 3560 always drops packets no matter what I configure but the 2950 doesn't.

Any answers would be greatfully received.

Thanks,

Paul

Joseph W. Doherty · ‎10-20-2009

I recall(?) there might have been a "defect" for this that was corrected in later IOS versions. What IOS are you running on the 3560?

jbrenesj · ‎10-20-2009

Yes I recall something like that happening but when qos is configured, I believe you refer to CSCsc96037

IOS upgrade to 12.2(25)SEE1 will have the fix, it has something to do with tuning the weights for each queue.

paul.l.kyte · ‎10-21-2009

I'm using a higher level version than th e one you state, my version is 3560 12.2(35)SE5.

Any more ideas on why the 3560 is not performing as the 2950?

Thanks,

Paul

Joseph W. Doherty · ‎10-21-2009

Well, that leaves manual adjustment of the buffers. What have you tried already?

paul.l.kyte · ‎10-22-2009

I'm presuming its the egress buffers that are at fault. The 100Mb interface is set to qset 1 and this had a default of 25 25 25 25 for the buffers. Altering this to give greater buffer share to the queue that our traffic was using did not fix the problem, additionally we tried with no qos at all on the box and it still dropped packets.

The 2950 didn't drop any packets!

Is there any way to increase the actual buffer size rather than a percentage of buffer space?

thanks,

Paul

Joseph W. Doherty · ‎10-22-2009

Could be insuffient buffer space or perhaps threshold drop that's insufficient.

"Is there any way to increase the actual buffer size rather than a percentage of buffer space? "

I suspect there is, with the "mls qos queue-set output qset-id threshold" command, but it's never been really clear to me how to optimally configure.

Here's a TechNote that further explains 3750 QoS: http://www.cisco.com/en/US/products/hw/switches/ps5023/products_tech_note09186a0080883f9e.shtml#topic3

paul.l.kyte · ‎10-23-2009

Thanks for this reply and link. Looking at this its the same information that I have already read in other manuals. The commands for qos are really to share the buffer space, bandwidth and to set thresholds to stop any traffic hogging the space/bandwidth. I've not seen anything that states how to increase the overall buffer space.

Quite interestingly with no qos the 3560 drops packets but the 2950 doesn't. It appears to me that the 3560 quite simply cannot handle this type of traffic either because of some limitation within itself or because the total buffer size is low compared to a 2950.

I find it hard to believe an obsolete piece of kit can outperform a modern piece of kit!

Any more ideas on this?

Thnaks,

Paul

Joseph W. Doherty · ‎10-23-2009

I suspect you might be correct that buffer space allocation might not be expanded via configuration. Yet, if a certain queue is dropping packets, and additional buffer space can be acquired from other queues that don't need it, one would think this might solve your problem.

However, as I noted before, the buffer management on the 3560/3750, isn't clear to me especially on the reserved vs. common pool usage.