[Foundation-l] [Internal-l] Relocation Announcement

[Gregory Maxwell]

On 9/24/07, Delirium <delirium at hackish.org> wrote:
> It's also a negligible improvement on technical grounds. If you look at
> the time it takes between a user requesting a page on one of our sites
> and them getting the result, trans-oceanic packet transmission times are
> not anywhere near the bottleneck.

A user? ... you mean a reader?  If they have a decent broadband
connection then latency overwhelmingly dominates their load times.

TCP/HTTP usually takes several round trips to connect and fetch an
object. Fetching locally to one of our squids. TCP transfer ramping is
very much latency sensitive for shorter transmissions. Pipelining
helps, but the connections to clients are still short lived even in
the best cases.

Forking curl and fetching http://en.wikipedia.org/wiki/George_W._Bush
from a system close to the knams caches takes and average of 0.017
seconds and I expect most of the time is measurement error and the
time curl takes to start.

Making the same request from the squids across the ocean takes an
average of 1.215 seconds.

There is no bandwidth bottleneck in between. The squids on each side
are pretty much the same. Latency is the reason for the difference,
and crossing the ocean adds a lot.

A lot of things can be improved, but we can't increase the speed of
light. Closeness counts.

This is not not the difference between okay and sllloowww..  It's the
difference between okay and 'faster than it would be if it were all on
my computer already'. :)

The situation isn't the same for logged in users because they don't
get the same benefits of distributed front end caches for text.
Performance for them matters too, but they are a infinitesimal portion
of the total users and requests.