[c-nsp] Improved queuing in 12.4(20)T?

[Per Carlson]

Hi.

I'm doing some QoS-testings and notice a remarkable change in the
latencies on a priority queue (as well as some improvement on other
queues) in 12.4(20)T compared with 12.4M (19, 19b and 21) and
12.4(15)T7. The scenario is H-QoS with a parent doing "shape average"
and a child with 4 queues:

class-map match-any Voice
 match  dscp cs5  ef
!
class-map match-any Business
 match  dscp cs3  af31
!
class-map match-any Network
 match  dscp cs6  cs7
!
policy-map Child
 class Voice
    priority percent 33
 class Business
    bandwidth percent 40
 class Network
    bandwidth percent 2
!
policy-map Parent
 class class-default
    shape average 8000000
  service-policy Child
!
interface FastEthernet X
 service-policy output Parent
!
end


When pushing traffic through the policy (Voice and Business within
contract and enough "class-default" traffic to trigger back-pressure)
I get the following latencies:

                   Voice     Business  class-default
12.4M/12.4(15)T:    13 ms      14 ms      126 ms
12.4(20)T      :   0.4 ms     8.5 ms      138 ms

There is no drop in neither the Voice nor Business queues, and about
21% in the class-default queue.

I have tried the same test on both a 871 (the WAN-port) and a 1841
with similar results. This makes me think there has been some major
improvement in 12.4(20)T, but can't verify this in the RN for
12.4(20)T.

There *is* a new H-QoS feature in 12.4(20)T called HQF
(http://www.cisco.com/en/US/docs/ios/qos/configuration/guide/qos_frhqf_support.html),
but I can't see that it's directly relevant.

Does anyone have some knowledge or insight to enlighten me here?

-- 
Pelle