physical distances vs. net distances

[Paul A Houle]

Jesse Keating wrote:

>
>>
>>For the most part, intra-continental routing has little difference to
>>your download speed, unless it is something like Alaska to Florida or
>>Seattle to New York type thing.  Usually very very high bandwidth lines
>>are used in such long jumps before breaking down into lower bandwidth
>>and dispersing out over an area.  However INTER-continental traffic has
>>to go under the pond if you will and there is a lot of latency (and
>>cost) there.  This is why using a mirror in your country or continent
>>will be much more preferred than crossing an ocean to get your files.
>>    
>>
    I've been watching intercontiental internet performance since '89,  
and I can say that performance (for long transfers) depends more on the 
load factor of the connection than anything else.  Around that time,  
Finland had a particularly good connection to the US,  and funet had 
faster downloads than most sites in the US.

    Generally,  usage of an international link varies diurnally and 
weekly -- it was horrible (often 30% packet loss) connecting to the 
states from Germany in 1999 during the day,  but on a German holiday,  
the transatlantic line was so fast that the US backbone felt faster than 
it had felt a year ago to me.

    Around 2001 I was working with a startup in Brazil,  and we noticed 
a different pattern:  the link to the US was blisteringly fast during 
the work day,  but slowed to a crawl around midnight -- it turned out 
that Brazillians logged on en masse after 11 pm,  when phone rates dropped.

    The deployment of Random Early Drop around 2000 or so made a big 
difference in international line performance.  Back in 1999,  I'd make 
graphs of the Germany-US link connection saturating in the morning,  
ping times increasing from about 400 ms to 1200 ms,  horrible packet 
loss kicking in only when the buffers were filled up.  The Brazil-US 
connection circa 2001,  on the other hand,  would start dropping packets 
at a low rate long before saturation,  which kept the connection running 
smoothly if slowly.

    In some countries,  like India and South Africa,  I get the 
impression that the international connection isn't the bottleneck to 
most sites,  but that the domestic network (at least between us and the 
places that we've got equipment) is overloaded.