---------- Forwarded message ----------
Date: Wed, 23 Sep 1998 18:38:48 -0400
From: Matthew J Zekauskas <matt@advanced.org>
To: Neal Cardwell <cardwell@cs.washington.edu>
Cc: Guy T Almes <almes@internet2.edu>,
Matt Zekauskas <zekauskas@advanced.org>
Subject: Re: Modeling of impact of loss on network performance
At 11:05 AM 9/23/98 -0700, Neal Cardwell wrote:
>Hi,
>
>Impressive graphs! Could you outline how you obtained the one-way delay
>and loss measurements? You can't use pings, right? Do you have software
>running at each of the sites?
>
>thanks,
>neal
>
Neal,
At the various locations we have placed Surveyor measurement machines.
Each machine is a Pentium-based (200-400MHz) PC, running BSDI, that is
equipped with a GPS card. We use the GPS cards to obtain synchronized
clocks on all the systems.
Among machines we send timestamped test packets. When the packets arrive
at the other side, we can compare the timestamp in the test packet with the
current GPS time, and thereby obtain the delay. Right now, the machines
sit in "mostly" a full mesh - most of the machines are in a full mesh but
there are a few that peer only with the machine at Advanced.
- the GPS clocks are rated to synchronize within 1 microsecond
- we find that our error is +/- 50 microseconds given the
current software
- the test packets are 12 byte UDP packets, therefore are 40
bytes total
- the port number is only fixed when the measurements start (so
you don't know which port nubmer is used in advance)
- the packets are sent out on a Poisson schedule, currently (and
for most of our collected data) with an average of two per
second.
- a packet is declared as lost if it doesn't arrive within
2 minutes.
- we run modified traceroutes in parallel, also on a Poisson
schedule but with an average of once every 10 min, since
they are more invasive.
For the graphs themselves, we divide a day into 1 minute intervals, and
then compute the minimum, 50th percentile, 90th percentile & percent
packets lost for each interval.
That's the 1000 foot view; let me know if you have more questions.
--Matt