Le 16/02/2010 17:47, donovan jeffrey j a écrit :
DNS round robin is bad, it works but is defective for real load
balancing. The client choose the IP to use, this is "random", and
after can use the same ip for a while... this is not random.
Again, I am doing every days exactly what required at the beginning of
this thread.
I have one Postfix server who simply relays all emails to send to a farm
of 40 mail sub-servers.
To load balance, I simply use a local DNS who manage a local domain. All
40 sub-servers are identified as equivalent MX of this local domain. The
Postfix server just ask the DNS what is the MX of this local domain, and
get a name /ip from the DNS.
I am very satisfied of this load balancing.
Again, I said before that the statistics show a difference of less than
2% of the traffic to the sub-server who work the most and the one who
work the less. It is, as I think, very very well load balanced.
When a sub-server fails, some messages are stuck in the Postfix, but a
very small percentage. Immediatly after you restart the sub-server, or
you put it out of the DNS, all messages are processed.
I don't think there is a need for more precision in the load balancing.
I don't think there is a need for keepalive, or any expensive device to
do it.
Patrick
