[e2e] Comparing different TCP flavours?

Prof.Anurag Kumar anurag at ece.iisc.ernet.in
Mon Dec 9 20:14:58 PST 2002


	On your points (i) Dynamic arrival of finite size flows, (ii)  
Average performance over the finite sized flows, the following paper of
ours may be useful.

Arzad Alam Kherani and Anurag Kumar, ``Stochastic Models for Throughput 
Analysis of Randomly Arriving Elastic Flows in the Internet,'' IEEE 
Infocom 2002.

	However, we do not study what happens if the randomly arriving
transfers randomly use different TCP "flavours".

	This and other related papers (including two on closed loop buffer
analysis with finite size randomly arriving TCP transfers) are available
on my webpage


and we have much ongoing work in this area, both on modelling and control.

Anurag Kumar

On Mon, 9 Dec 2002, Martin Karsten wrote:

> During the years of TCP development, have there been systematic comparative
> studies of TCP performance of different TCP flavours - Tahoe, (New)Reno,
> Sack - in a "closed setting"?
> It seems to be easy to find work comparing TCP flavours with a focus on the
> performance of individual flows under a significant amount of "black-box"
> background traffic. Also, there is some work on the behaviour of "many" TCP
> flows, but only considering a single TCP flavour (i.e. no systematic
> comparison). Same for investigating individual parameters, like buffer
> space, etc. Finally, there is all the work on TCP variants over wireless
> links.
> However, I am wondering whether there has been any (simulation) research on
> comparing plain TCP performance where
> - all traffic is created by monitored TCP sources (i.e. no black-box
>   background traffic)
> - there is dynamic arrival of (finite-size) TCP flows, rather than a fixed
>   set of long-lived flows that run from the beginning to the end of a
>   simulation experiment
> - overall network efficiency is measured, e.g. in terms of when (on average)
>   each/all/most/etc. TCP transactions are completed, rather than studying
>   the temporary rate allocation(s) to individual flow(s)
> - TCP flavours and configuration parameters are systematically enumerated
> Of course I understand that the above goals are not trivial to accomplish.
> Alternatively, I would be delighted to be convinced that the above is not
> necessary, because the relative overall resource efficiency can be
> established otherwise.
> I'd be grateful for any pointers.
> Many thanks,
> Martin

Anurag Kumar				     |e-mail: anurag at ece.iisc.ernet.in
Professor             			     |Phone: (+91)-80-360 0855 or
and Associate Chairman 			     |       (+91)-80-394 2387
Dept. of Electrical Communication Engg. (ECE)|Fax:   (+91)-80-360 0991 or
Indian Institute of Science (IISc)           |       (+91)-80-360 0683
Bangalore, 560 012, INDIA                    |http://ece.iisc.ernet.in/~anurag

More information about the end2end-interest mailing list