Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

RE: EF% vs. BCN

Hi Harry, Raj,
 
I absolutely agree with you that there is a tradeoff between fairness and stability vs. bw utilization. This periodicity is exhibited by not only by CBR traffic but also by TCP traffic (over 90% traffic on internet).
 
TCP flow control mechinism increases window size lineraly after the round trip time (RTT) when acknowledment (ACK) messages indicate successful packet delivery. If  ACKs indicate a missing packet, TCP reduces the window size by half (exponential backoff). TCP flows eventually adapt to the available network bandwidth and flow bandwidth oscillates (depends on the version) around the available network bandwidth. Thus periodicity is exhibited. During congestion times if network drops multiple packets, many flows backoff. Thus now even though network bandwidth is available for flows to use, but flows can't use it because of TCP flow control. When flows come out of slow start, they increase their bandwidth at the same time till congestion happens again. Thus global synchronization of flows lead to periodic congestion.  

Backward congestion control mechanism is implemented to push congestion to the upper layers across the nodes in the ring. This congestion control mechanism is basically is an ABR approach which is implementing another level of flow control. In presence of two control loops, end to end TCP flow control loop and BCN control loop, it is not clear how they will interact.  TCP may not come out of slow start to take full advantage of the available bandwidth. 

These are important points to think about.

 

-----------------------------------------------

Sanjay K. Agrawal, Ph. D.

Luminous Networks

sanjay@xxxxxxxxxxxxxxxxxxxx

 
 -----Original Message-----
From: Harry Peng [mailto:hpeng@xxxxxxxxxxxxxxxxxx]
Sent: Thursday, August 17, 2000 5:55 AM
To: Khaled Amer
Cc: Raj Sharma; Reflector RPRSG; Vinay Bannai; Sanjay Agrawal; Charles Barry
Subject: Re: EF% vs. BCN

 
Raj:

Interesting traffic pattern....
If I understand your question, the EF BW is reserved and leaked into the Ring
It appears as reserved time slots; hence a periodic function?

The question is what is the congestion threshold and when to send congestion notification. The time scale for congestion detection is important. If buffers are used, then it is some buffer threshold.  Otherwise, is congestion
detected at cell level, burst level, call level, or days etc.

There is one trade-off for fairness: stability versus maximize BW utilization.
-If you response too fast going into congestion, (reacting to small burst), you are susceptible to instability and you  BW efficiency may be reduced.
-If you response too slow, you will not be fair, as some node may get starved.
-If you response too fast in existing, you may not be fair again, and you oscillate.
-If you response too slow in existing, you loose efficiency.

Regards,

Harry
 
 

Khaled Amer wrote:

Raj, Not sure I understand how you reached to the conclusion of 'the traffic at any node would be a periodic function'. Can you shed some light on this? Khaled Amer
President, AmerNet
Architecture Analysis and Performance Modeling Specialists
Phone: (949)552-1114             13711 Solitaire Way, Irvine, CA 92620
Fax:     (949)552-1116             e-mail: khaledamer@xxxxxxx
----- Original Message -----
Sent: Wednesday, August 16, 2000 9:36 AM
Subject: EF% vs. BCN
I would like to findout if anyone has giventhought to Explicit Forward (EF) traffic versus BCN mechanismsused on buffer insertion rings.Assume that x% of the BW is provisioned for EF.As EF traffic into the ring it will given highest priorityforwarding and hence to culminate in all EF packets arriving together. The BW consumed by EF traffic overtime can become a time function. Example x(t) = sine or square wave function of t with an amplitudeof x.As a consequence available BW for the rest of the traffic at any node can become a periodic function. There could be interplaybetween this perdiodicity and the periodicity of the congestion notification messages.Has anyone simulated this or have given thought tothis?
-- 
Harry Peng               
------------------------------------------------------------------
Dept: 1E11              
Email: hpeng@xxxxxxxxxxxxxxxxxx
ESN:   39-52277           
Phone  613-765-2277
Fax:   613-768-4904 
Web:   http://skywww/~hpeng/
-------------------------------------------------------------------