[e2e] Re: crippled Internet]
wenyu at cs.columbia.edu
Sun Apr 29 15:14:06 PDT 2001
Because I missed the initial part of this discussion, can someone please
point to me the exact places (urls, etc.) where delays on Internet2 is
very high (e.g., exceeds 500ms)? I browsed through the Surveyor and NLANR
page, and found a few anomalies myself, I figured it may be beneficial for
others to know. I am including the exact date(s), peer(s) so that you can
also see for yourselves easily:
In most of the surveyor's data, the (90%) delay is "capped" by 300 ms.
And more importantly, if packet loss during a measurement interval (1
minute) exceeds 10%, its 90% delay becomes meaningless (or infinity),
therefore it will be shown as a "capped" 300 ms.
I checked a few paths, such as, Advanced->Columbia (e.g., Nov 10 2000),
there are 3 spots with capped 300 ms delay. Its same-day loss stastics
graph (where loss% > 10%) matches well with the delay graph. Its same-day
reverse path is even worse, with a lot of high loss spots and
corresponding capped delays. Its same-day traceroute suggests that the
routing is through Internet2 (via nn2k and abilene). Therefore it is
really strange that packet loss can be so bad on Internet2 (even within
US). If you look at previous/next day stats, you'll find the situation
had lasted for a few days, then faded out on Nov 13. The route did NOT
change, though. So I wonder what had happened.
Other paths, e.g., Stanford->Columbia, Jun 22 2000, there are a few
high loss spots (strange that 90% delay is not capped in this
graph). It seems that on Internet2, even during normal operation,
loss% is not very close to 0, more likely on the order of 0.1%, but
they seem to appear only once in a while rather than being random
(whether conforming to a Bernoulli or even Gilbert model). Also, from
Jun 27 to 28, delays in both directions increased by about 25 ms and
then popped back. The route changes do not coincide or match well with
the delay changes. I cannot explain it.
It is strange that so far I did not find any Surveyor data available for
any path in year 2001. So I wonder if surveyor is still operating.
For the NLANR site, I looked at some of the summaries, e.g., from U.
The loss% between Alaska and Florida State U and GA Tech is > 5%, which
is not so good for an Internet2 path (if it is). Alaska <-> CSU San
Bernardino has a high stdev and max in delay, although one may doubt
whether the path is Internet2. So I looked at stats from GA Tech, which is
surely on Internet2:
GA Tech <-> Columbia has loss > 7%, stdev is low, though. GA Tech <->
Dartmouth has 6.74% loss and 512 ms max delay and 13 ms stdev. And now I
wonder what is happening here. You may have noticed loss% to CSU San
Bernardino is 100%, so it is possible that the end-points are down during
the measurement, and we cannot tell whether a 7% is network loss only or
partially due to end-points. But delay variation cannot be caused by a
dead end-point (though a slow end-point may be responsible, but we
wouldn't know), and there are plenty of sites in this page that has large
stdev and max delay, like Vanderbilt U., Virginia PolyTech, all of which
are supposedly on Internet2.
Finally, GA Tech end-point seem to suffer from >5% losses for all the
paths associated with it, while other end-points have relatively lower
loss rates. So perhaps the losses are "real", and it may have occurred at
the access network level. But I would have to know more about their test
data to be sure of their validities.
-------------- next part --------------
An embedded message was scrubbed...
From: "Christian Huitema" <huitema at exchange.microsoft.com>
Subject: RE: [e2e] Re: crippled Internet
Date: Wed, 25 Apr 2001 19:15:37 -0700
More information about the end2end-interest