Asymmetry of transatlantic traffic [Re: [e2e] congestion correlation?]

[Simon Leinen]

>>>>> "ao" == Andrew Odlyzko <amo at research.att.com> writes:
> [...] substantial asymmetry in the traffic load in the two
> directions.  This is true even within the US. but very noticeable on
> trans-Atlantic and trans-Pacific connections.  As an example, if you
> look at the data for JANET, the British academic and research
> network (available at <http://bill.ja.net/>) you will find the
> following ratio for (traffic from the US to Britain) / (Britain to
> US traffic):

>  month         ratio
> --------        ----
> March 97        1.26
> March 98        1.98
> March 99        2.05
> March 00        3.27
> March 01        2.70

Apart from the 1997 numbers this corresponds well to our measurements
(traffic between the US and SWITCH, the Swiss academic network):

month        ratio   US -> CH  CH -> US
Mar 1997      2.06     817 GB    397 GB  
Mar 1998      2.35    1405 GB    597 GB  
Mar 1999      2.09    3150 GB   1508 GB  
Mar 2000      3.00    5961 GB   1984 GB  
Mar 2001      2.36   10877 GB   4614 GB  

> (An increase in this ratio also held true for some other links in
> the late 1990s, and I am watching with interest whether it will
> reverse course and shrink, as the ratio for JANET may be doing.)

Some wild guesses about reasons why the level of asymmetry changes...

At least in our case, most of the data is "content" from the US
requested from users within our customers (universities).  Somehow
there's always more of this going US->Europe than the other way, I
won't speculate why this is the case.

Anyway, when a transatlantic link is congested (in the US->Europe
direction during office hours in Europe), then this will limit the
amount of US->Europe traffic, but the Europe->US traffic will not
decrease much.  So a lower ratio may just reflect a more congested
link.  Thanks to a huge decay of prices for transatlantic capacity
over the last years, many European networks were able to get out of
the permanently-congested-during-daytime situation recently, which has
caused the ratio to increase.

If the ratio is coming back down now, this may be driven by
applications such as Napster and Gnutella, which contribute more and
more traffic (in absolute and relative terms), and which seem to
exhibit more symmetric traffic patterns than Web access.

> Capacities are generally completely symmetric.  This is true 
> throughout the telecom world (except for newer technologies
> at the edges, like DSL, cable modems, and satellite data), a
> legacy of the voice telephony world.  (Frame Relay offers 
> customers the option of asking for asymmetric PVCs, but this
> is just how the capacity is sold, as the Frame links themselves
> are symmetric.)  In my Globecom '99 paper I used this as an
> argument against QoS:  Basically the reasoning was that with 
> far less effort than is going into developing and deploying
> QoS, one could develop a transport infrastructure that would
> take advantage of the asymmetries in traffic patterns, which
> tend to persist for extended periods of time, and that this
> would provide at least as much of a gain in performance.

I find this argument quite convincing.
-- 
Simon.