[e2e] How to increase the flow's rate reliably???

Jeevandra jeevandra at swift.ee.umist.ac.uk
Sun Aug 19 13:30:12 PDT 2001


Hi there,
I am a research student trying to develop a rate-based transport protocol primarily catered for streaming video on the internet. At the moment I am employing a rate adjustment mechanism based on LIMD. My protocol in a nutshell, the sender transmits packets at an everage rate and the receiver monitors the reception of these packets. After a period (typically after a SRTT), like RTP the receiver sends a report back to the sender. If there is no packet loss experienced, the sender increases its rate linearly otherwise it halves its rate.
At the moment, what I was hoping you guys could advise me is on the rate increase step. I realise that during the congestion avoidance state, most TCP implementations increase their window by one packet after each RTT. As a result, I decided to employ a pretty similar rate increase control action as shown in the equation below.

Rate_new = Rate_old + (TLastUpdate* PacketSize)/(SRTT*SRTT)

where TLastUpdate = Time elapsed since last rate update.
I believe this would mimic TCP's rate increase step quite resonably. This or something similar has also been used by others in their rate based transport protocol. 

However, I've found in my simulations that such a mechanism could induce severe oscillation in the rate particularly when used with rate-halving multiplicative decrease. This is mostly obvious when a huge PacketSize is used and the path's RTT is small. A PacketSize = 1000 bytes and a RTT = 100 ms would give a rate increase of 80,000 bits/s for each RTT. I am sure a continual increase by such rate could be quite drastic for a narrow bottleneck link and as a result the flow might be vulnerable to regular packet loss and rate halvings.

I was hoping if anybody could share a light on this and advise me on a more suitable rate increase technique, particularly for streaming. Has anyone done any work on the implications and stability of this 'TCP like' rate increase? I do realise that the PacketSize, RTT, SRTT and the Bottleneck Bandwidth plays an important role in deciding on the suitable amount to increase the rate at each control step. Has anyone come up with a more adaptive rate increase technique? 
Your help and advice is very much appreciated. I thank you all in advance.

Regards,
Jeevandra Sivarajah
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.postel.org/pipermail/end2end-interest/attachments/20010819/92b634a1/attachment.html


More information about the end2end-interest mailing list