[e2e] Question on traffic patterns for AQM design

Wei Wu wuwei at sdp.ee.tsinghua.edu.cn
Sun Sep 30 19:53:00 PDT 2001


Hi,Zhang Miao

Before we design AQM algorithms, we should make clean that what our goals are. 
Network congestion can be divided into 2 classes according to their time scale:
persistent congestion, which is typically longer than 1RTT and transient congestion, 
which is less than 1RTT. AQM targets on detecting and avoiding persistent one, while
accommodating transient ones. So just think network traffic go through a low-pass 
filter (critical frenquency is less than 1/RTT), the output of the filter is what 
we want to control. If AQM controller can adapt itself to track the output 
traffic, AQM can response well enough then.

Yes, the parameter settings of PI controller cannot work well for all network 
dynamical conditions, especially when the buffer size is low or the network load 
is heavy. The problems are the static parameter settings. We believe it is hard to 
design an AQM algorithm, which can suit for all network conditions with only one 
set of parameter setting. So the algorithms should be adaptive, adaptive to 
network dynamics. Adaptive RED validates this point. We are working on an adaptive 
PI controller recently. Preliminary results show that it can speed up the response
of PI controller much.   

Best Regards
Wei Wu
wuwei at sdp.ee.tsinghua.edu.cn
|=====================================================================|
| Wei Wu, Complexity Engineering System Labs (CESL)                   |
| Electronics Engineering Department, Tsinghua University             |
| Beijing, P.R.China                                                  | 
|                                                                     | 
| Office:Room 401 Division 10,Eastern Main Building of Tsinghua Univ. |
|       (tel) 86 10 6277 3228                                         |
| Dorm  :Room 300 Apartment 14, Tsinghua Univ.                        |
|       (tel) 86 10 6277 8632                                         |
|=====================================================================| 

In your letter 2001-9-14 9:45:00 
>Hi, Lynne
>   Some more for comment.
>   My question comes from the comparison between RED (proposed by Sally Floyd)
>and PI controller(proposed by C.V.Hollet, etal.). One shortcoming of RED is that
>as the number of flows increases, the average queue length is also increase.
>With introduction of integral factor, PI controller eliminates the "steady-state
>error" in RED. Unfortunately, there is no free meal. PI controller responses much
>slower than RED. This tradeoff has been analyzed in books on control theory(such as
>"Feedback Control of Dynamic Systems" by Gene F. Franklin, etal.)
>   Whatever RED or PI, some time is needed for the end system to response. If we
>don't consider the time for the controller to adapt to the changes in traffic,
>the shortest response time is the RTT. Howerer, if the Internet traffic changes
>very dramatically in very short time (approximate RTT or even shorter), it is 
>nearly impossible for end system and for controller at the gateway to response.
>   Here I have another question on the understanding of queue length. Many papers
>use average queue length as a metric. When traffic changes frequently, the queue
>length oscillates. Is it meaningful to have a little average queue length while the 
>queue length ocillates between zero and maximum queue length(ie., buffer size)?
>This will introduce delay jitter which is harmful for some applications.
>   Anyway, using AQM is better than only using droptail. With more information on
>the traffic patterns, we can make better tradeoff on AQM design.
>
>>The upshot - Dr. Cerf himself has commented that unless one reduces retransmission, 
>>congestion events will overwhelm. Timescales for active queue management must be fine 
>>grained, not coarse grained. TCP at the fine timescale integrated throughout the 
>>Internet is the only solution.
>
>Using fine grained timescales in TCP can't solve the problem.
>It is difficult for TCP to estimate the RTT.
>Fine grained timer also adds burden to the end systems.
>
>*****************************************************************
>*    Zhang Miao                                                 *
>*    Ph.D candidate,Department of Computer Science & Technology * 
>*    Tsinghua University,Beijing,China(100084)                  *
>*    Tel: (8610)-62785822                                       *
>*    Email: zm at csnet1.cs.tsinghua.edu.cn                        *
>*    Web: http://netlab.cs.tsinghua.edu.cn/~zm (domestic only ) *
>*****************************************************************






More information about the end2end-interest mailing list