Tracy R Reed wrote:
Andrew Lentvorski wrote:
S3 has good bandwidth but sucky latency. My guess is that those
systems all try to write synchronously. They probably get clobbered
by the round trip delays.
You keep going on about latency but I don't really understand why. You
are just uploading a file. Where does latency get involved? Are you
saying they take too long to send a tcp ack for each of your sent
packets or something?
Both TCP and object transaction.
TCP latency is sufficiently inconsistent that nobody can use S3/EC2 for
streaming applications yet.
Object transaction allocation is also sufficiently inconsistent that you
need to be able to maintain a large stream of objects in flight
without waiting for any particular object to complete.
You need something that will put a lot of transactions on the wire
without waiting for the previous ones to complete.
We are doing backup and mass storage. Not anything transactional. I
never have more than one transaction to put on the wire and it takes
anywhere from minutes to hours to complete depending on the upstream
bandwidth I have.
All you do is transfer a single file so object overhead is nil. This is
a single transaction and TCP on both systems probably has a pretty deep
queue. You can lose a packet in this case for a long time (100's of
milliseconds) and everything will still fly.
-a
--
[email protected]
http://www.kernel-panic.org/cgi-bin/mailman/listinfo/kplug-list
