[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [ns] Hoe[] idea
* Mark Allman ([email protected]) [200300 13:40]:
>
> As for ssthresh estimation, it is implemented (if setting the
> "newreno_changes" variable, anyway). Well, sort of implemented
> anyway. The Hoe paper is not particularly verbose on the actual
> algorithm, so I just added something simple long ago (and I see that
> it has survived).
>
> IMHO, investigating bandwidth estimators such as the Hoe proposal is
> better done in real networks. For instance, when I added this
> mechanism to ns I got some pretty good results. But, when these
> mechanisms were evaluated "in the wild" the mechanism didn't work so
> well (see On Estimating End-to-End Network Path Properties,
> Allman/Paxson, SIGCOMM '99
> http://roland.grc.nasa.gov/~mallman/papers/estimation.ps).
I would be inclined to agree with Mark. I have done some
work in this area using simulators (REAL and 'ns') and found the
environment to be a little sterile. Work with estimators such as packet-pair
is more suited to "wild" networks due to their fluid and dynamic nature.
However, it is quite easy to implement a simple dual-packet ping against
a remote host in 'ns', and take the result to initialise the congestion
window for a Reno/Vegas/Whatever source. From what I recall, this is
approximately what Hoe suggested (please correct me if I'm wrong).
There was some more interesting work done in 1999 using packet-trains --
i.e. chains of probe packets to estimate bottleneck bandwidth. However,
I'm not sure whether it was published or not. I can dig out the
reference if anyone's interested.
I have a couple of papers on the use of packet-pair in transport
protocols at http://www.scs.leeds.ac.uk/atm-mm/publications.html
thanks,
rik wade
--
+------------------------------------------------------------------------+
| [email protected] mob: +44(0)7771 766621 |
| Internet Systems Developer tel: +44(0)1132 076054 |
| Planet Online Limited fax: +44(0)1132 346065 |
+------------------------------------------------------------------------+