Wietse,
Please try the following, as asked half a week ago:
postconf -e smtp_connection_cache_on_demand=no
postfix reload
and report if this makes a difference.
Wietse
I have tested this since yesterday night.
I got some problems with Linux per user number of processes limit. I
fixed it. I also increased some delivery concurrency figures, and now I
can see up to 1300 processes delivering emails to the qmail servers.
I had a few minutes shot today at a rate of 6300 emails per minute. I
ran a full hour at 180,000 emails per hour. The outbound line was saturated.
CPU is about 30% loaded, no Wait I/O, no swap, memory is large.
I think I will reach about 600,000 emails per hour if I fix some timeout
on the qmails (replace by postfix?). Maybe I could reach 1 million?
The full architecture that I plan will include 2 to 3 clustered postfix
relays and 50 2nd level qmails(or postfix) delivery servers, each with 3
to 5 IP addresses, and upgraded outbound internet connection.
With your help, I better understand now the impact of timeout and
concurrency parameters. In fact, delivery was blocked because postfix
was trying to reuse connections, so was waiting each email to complete
to send the next one. Also, because hundreds processes were created at
start time to manage inbound messages, there were no slots to fork
processes to deliver messages on the other hand. Same problem caused
very slow DNS and EHLO, because no available slots to fork.
Of course, if you want me to post my conf, I will with pleasure.
Many thanks to you, to Victor and Stan.
Patrick
