[p2p-hackers] Modelling Latency and Packet Loss on the Public Net

Michael Rogers m.rogers at cs.ucl.ac.uk
Mon Jun 12 19:35:29 UTC 2006

Hash: SHA1

Hi Will,

I'm writing a p2p simulator at the moment and I've been facing similar
questions. Here's what I've come up with.

If you're looking for the distribution of the mean latencies of
different links (rather than the distribution of the latencies of
packets on a given link), Figure 5 of [1] shows what looks to me like a
lognormal distribution with median 140ms, 20th percentile 70ms and 80th
percentile 280ms. Lognormally distributed samples are easy to generate:
exp(x) where x is normally distributed.

As for packet loss, er, I'm at a loss. Figure 3a of [1] gives
distributions for upstream and downstream bandwidth (unfortunately not
as neat as the latency distribution). But even if we know the bandwidth
of the access link, the amount of packet loss you experience will depend
on the amount of background traffic, how the background traffic reacts
to packet loss, and how your traffic reacts to packet loss.

If your traffic is TCP-friendly and you can estimate how many other TCP
or TCP-friendly connections will be sharing the access link then it
should be possible to work out the steady-state packet loss... possibly
by rearranging Equation 1 in [2] to give p as a function of T, s, and R?
I have a feeling that t_RTO can be estimated as 4*R but I can't remember
where I read it. :-)

That's as far as I've got - any pointers appreciated!


[1] http://www.cs.washington.edu/homes/gribble/papers/mmcn.pdf
[2] http://www.icir.org/tfrc/aimd.pdf
Version: GnuPG v1.4.3 (GNU/Linux)


More information about the P2p-hackers mailing list