Hi Wietse, all,

Closing the loop on this thread.

After a lot of experiments and trial and error, we found the cause was OpenDKIM. When we took it out of the mail path and used an alternative for DKIM, the system got fast again.

With OpenDKIM in the path, we saw 100+ smtpd processes. After replacing it, the count dropped to as low as 12–15 for the same traffic. The kernel and the other Postfix settings were unchanged.

So Wietse was right: the accept-lock waits were a symptom, not the cause. The milter was adding time to each SMTP session, so each smtpd stayed busy longer. That drove up the number of processes during bursts and left many idle ones waiting on the lock afterward. My DNS and lock tracing was looking at the wrong layer.

Thanks for pointing me toward per-client service time.

Regards,
Rajesh Mishra

On 01/09/26 4:50 pm, Wietse Venema via Postfix-users wrote:
Again, the number of (smtpd+cleanup) processes during peak load
depends on the traffic volume and the time to service an SMTP client.

Apparently, your newer servers take more time to service an SMTP
client. Slow DNS can be one such cause. A file system with a slow
fsync() implementation can be another.

You need to find out what causes the DIFFERENCE in latencies of the
old and new servers. Translation: this involves measurements on
BOTH old and new servers.

The locks are a red herring. These are 'after the peak' processes
waiting for a client connection. Those proceses go away after a
timer goes off.

        Wietse
_______________________________________________
Postfix-users mailing list -- [email protected]
To unsubscribe send an email to [email protected]
_______________________________________________
Postfix-users mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to