Steve,
I don't have any great answers for you, but I thought it might be helpful to let you know that I have seen the same thing, and share some info I found. Long story short... 1) Lower-end hardware is easier to overload and may drop a little of everything if flooded beyond resource capacity, and 2) You should see better performance with stateful traffic than with non-stateful.
Short story long...
We were at the Cisco Proof of Concept lab in San Jose running a QoS simulation, using a tool that would blast non-stateful TCP traffic over a link. For such purposes, non-stateful TCP traffic is the same as UDP in that the traffic does not respond to network conditions by throttling/windowing up and down, it just keeps blasting at a constant speed, which for us was 4x the capacity of the link.
And what we saw was that some of the smaller PA's just could not keep up, even if our QoS configs were correct. We had a 2651 with a WIC1DSU/T1 PA that would drop a little of everything when we tried to blast 6 megs of non-stateful traffic through. The Cisco lab guys got one of their colleagues on the line who is a QoS guru, and he said that he had seen it before, and we were simply overloading the hardware. That T1 PA has finite resources and we were asking too much of the interface. It was trying to drop as much best-effort traffic as quickly as possible but it just could not empty the outgoing buffers fast enough, so all traffic classes were impacted.
However, when we blasted 6 megs of stateful traffic through the same link, the PA would do just fine, and QoS behaved as expected. But that is because the stateful traffic responded to packet drops and throttled/windowed down, just as actual (non-UDP) traffic should.
But on higher-end WAN interfaces, we did not see the same problem. We could blast through 4x link capacity, and QoS treated both stateful and non-stateful traffic as expected.
I know this is no grand solution, but I thought the info might be helpful. Good luck!
Best Regards
Robert Hyde
