Timwi wrote:
Gerard Meijssen wrote:
I also do not understand your preoccupation with data centres in the
USA. We are an international organisation. Our customers are all over
this earth.
I do not understand everybody's preoccupation with data centres around
the world. Our "customers" are exclusively Internet users, and the
Internet is global. That servers are faster when they are
geographically closer is a wide-spread myth.
Timwi
Let me help you understand. Your argument does not take into account how
the real Internet works.
The individual servers certainly run at the same speed wherever they are
in the world, but the service does not, because the Internet is not a
free, perfectly reliable, zero-delay network with zero packet loss and
infinite bandwidth.
Consider the effects on TCP dynamics of:
* increased round-trip delays and timing jitter from long paths
* the increased probability of congestion in long paths, rather than
short ones
* the increased probability of having non-congestion-related packet loss
on a long link, compared to a short one
Now also consider:
* the costs of local vs. global transit, given a limited budget
* the likelihood of being offered these for free by donors, given their
relative costs
* the relative costs of transit vs. peering, and opportunities for flow
balancing
and hence the likelihood of having more non-overbooked high-quality
connectivity to more places available at a given price, which in turn
will make the service more resilient against, for example, DDOS attacks.
In particular, you should consider
* the remarkably high performance boost Wikimedia obtained from adding
the Paris cluster, of a mere three machines.
* why most of the root servers are actually globally distributed clusters,
* why Google distribute their servers globally
* and why Akamai and other distributed hosting outfits continue to be
able to sell their services, rather than their customers maintaining a
single big cluster.
Oh, and finally, the killer arguments: the decreased probability of
routing instability and connectivity problems, due to a smaller average
number of forwarding and AS hops, and the removal of a single point of
failure in the event of a data centre disaster.
-- Neil