[altq 1294] The more traffic come, the less delay.

[Hiroyuki KATO <hiro@nemoto.ecei.tohoku.ac.jp>]

Hello.

I am a university student and currently
having a research about time-delay in IP-networks.
The reason why I give you this letter is that I have a strange output when I made a simple experiment of CBQ.
The weird thing which actually happened was as follows...

I am using ALTQ3.0 on FreeBSD4.3 release.
I connect 3 PC and make a bottle-neck as follows.

   --------           --------            ------
  | source |---------| router |----------| sink |
   --------  100Mbps  --------   10Mbps   ------

this CBQ has two classes, such as a real-time 
class and a background class.
each of the class has 50% of bandwidth.  
altqd.conf is below.
---------------------------------------------------------
interface dc0 bandwidth 9600K cbq

class cbq dc0 root NULL pbandwidth 100

# real-time traffic
class cbq dc0 real_cls root pbandwidth 50 borrow
        filter dc0 real_cls 0 15000 0 0 17

# background traffic
class cbq dc0 back_cls root pbandwidth 50 borrow default 
-----------------------------------------------------------

I also used 'rude & crude' as a set of packet generator and packet
moniter.
http://www.atm.tut.fi/rude/

First, rude made 12 background flow , each flow was 1Mbps UDP.
Then, rude made realtime flow increase to 6Mbps by 1Mbps.

When real-time traffic was less than 5Mbps,
no congestion occured and its delay was little(about 3ms).

when real-time traffic was incresed to 5Mbps,
a congestion was occured and the time-delay was increased 
to approximately 50ms.
I am pretty sure that it happened because of queueing descipline. 

However, when real-time traffic was increased to 6Mbps,
the time-delay was decresed from 50ms down to 40ms.

I do not know why this happened.
I would appreciate if someone gives me a comment about it!
Thanks,

Hiroyuki Kato

----

hiro@nemoto.ecei.tohoku.ac.jp