[e2e] packet-pair probe implementation

[Zartash Afzal Uzmi]

Further, if the interest is primarily in bandwidth estimation, Kevin Lai
worked on "nettimer" for his Ph.D. in 1999/2000 time frame. If I recall
correctly, his work was also based on packet pair. It might be worth taking
a look at his work.

I looked at nettimer and packet pair a while ago for use in real world. I
also explored other possibilities. In my experience, simpler schemes (e.g.
total data/total time excluding the handshake period) gave reasonably good
estimates for low bandwidth connections (cellular, wireline 56K dialup,
etc.) but didn't work quite well on higher bandwidth connections. Perhaps,
the reason is that for high bandwidth connections, end systems are usually
the bottlenecks, and for the low bandwidth connections, bottlenecks
generally lie somewhere inside the network. I did not explore it further
since the work was supposed to benefit a product which had associated
timelines. Most connections where the product was supposed to be used were
low bandwidth and we were getting okay results anyways.

Zartash

-----Original Message-----
From: end2end-interest-admin at postel.org
[mailto:end2end-interest-admin at postel.org]On Behalf Of Rik Wade
Sent: Friday, May 02, 2003 12:21 PM
To: Atsuo J.
Cc: end2end-interest at postel.org
Subject: Re: [e2e] packet-pair probe implementation


On Thu, 01 May 2003, Atsuo J. wrote:

> Is the packet-pair probe (bandwidth estimation) technique (introduced by
Keshav and revised by Paxson) really useful in the realistic network?Does
anybody test it on the real testbed? Also, are there any real
implementation/usage of this technique at all? Your suggestions will be very
appreciated. Atsuo

I am not aware of any real world implementations of a packet-pair estimator
other than those mentioned. However, I did a great deal of work with
packet-pair as part of my PhD thesis in the Real and NS simulators. I have
an implementation of packet-pair for both simulators if anyone would
like the source code.

At least in a simulation environment packet-pair was very effective, at
least
until the congestion became so great that probe packets were regularly lost.
One concept which I believe has been enhanced in recent work is the concept
us using "packet trains" as opposed to isolated pairs. I recall reading
a paper (http://www.sics.se/~bengta/train.ps.gz) in 1999 on the subject.
This technique involves sending a stream of back-to-back packets, the
interarrival properties of which can be measured and manipulated to suit
your requirements. Searching Google for "packet train bandwidth probe"
brought up quite a few hits.

In my work, I combined a packet-pair startup phase with a pro-active
congestion avoidance algorithm (a slightly simplified Vegas-like routine).
The congestion window was also replaced with a token bucket model (the
rate of which was governed by the initial probe phase and then adjusted
according to RTT variation) in order to better facilitate burstiness at the
application layer. This proved highly effective in simulation, but I have
not progressed to practical implementation. The papers and thesis from
this work can be downloaded from http://www.rikwade.com. Source code is
available on request.
--
rik