Steve Casner's paper, which you cited, and Sue Moon's paper at
http://an.kaist.ac.kr/~sbmoon/paper/infocom2004.pdf, both report very
limited variation in delay within the ISP network. Sue's paper goes on
to describe points of variation on the order of ten and 100 ms in some
detail as well as reporting the general case of low variation in delay.
But most people don't live within the PE-PE domain, where these studies
were done - they connect to the backbone ISP through an access carrier
or through an enterprise network, or connect via some longer path. So
responding defensively "give me numbers" and citing as proof of your
case a paper that only looks at the path within the ISP has the effect
of shutting down and making an end-to-end discussion appear to be
invalid when Casner and Moon in fact only perform a measurement of a
part of the path.
uh, fred. it was vicky who made the comparison i2 to internet,
not i. i2 does not include site links, and some are good and some
are bad.
it is common wisdom today that the internet backbone is not where
congestion occurs, but rather the customer tails. though one
should always be suspicious of common wisdom, this particular bit
seems pretty well supported, pings from uganda's makerere
university notwithstanding.
you/ve been pushing qos for a long time, fred. but, in the current
situation, where the tails are the issue, signaling back from dest
to source is still the big gap. imiho, from the ops perspective,
only sally's ecn has made any useful approach. sadly, we may be
able to judge the actual demand for e2e qos by ecn's very slow
deployment. i think this is unfortunate, as ecn is pretty cool.
but, in this community, the question would seem to be how long the
current situation will prevail, where it is far simpler and less
expensive to throw bandwidth at the backbone, as opposed to
spending even more on opex-eating complexety and ever more complex
and expensive routers. i suspect it'll be a while before we even
see cotton balls being blown, and a very long while before new
ducts. i.e., raw bandwidth costs will likely stay low. even the
price of lighting it is declining.
this has been discussed recently, both here and in simon lam's 2004
sigcomm award paper (recent ccr). so, i think we should
o encorage i2 as the usg's way of subsidizing higher ed [0] and
providing a playpen where big spikes and other traffic
anomalies are not discouraged
o encourage qos research
o keep the real internet as simple as possible, after all, it is
fools such as i who have to run it
randy
