dovecot - Feb 2023 - Roundcube <-> Dovecot timeouts under very moderate load on big Dovecot nodes

Due to a recent topic regarding the performance of the Roundcube webmailer with
Dovecot as a backend, I set up imapproxy successfully, shrinking the response
latency of the webmailer heavily, which is great.

However, I still encounter strange timeouts on some requests when Roundcube
tries to access Dovecot. Be it for authentication or to FETCH data. I run the
tests/benchmarks via distributed JMeter from 4 nodes, they access the webmailer
which in turn accesses the Dovecot via imapproxy. There are 16 different
mailaccounts spread over two dovecot nodes with the first one acting as a proxy
if the account resides on the second. The problems start at around 100
concurrent users and gets worse the more users are added.

The timeouts make up for a small percentage of the requests (0.0something
percent). The timeout is reached after 30 seconds of Dovecot not answering, due
to the timeout set in the webserver the webmailer is running on.

Now I am wondering how to further debug this and what could be the reason. Some
requests towards Dovecot are certainly taking longer than 30 seconds, while the
average request takes mostly less than 1. The Dovecot is running on two big
hardware machines with 48 cores, 96G memory and SSD storage. The load on the
Dovecot nodes while the test is running is minimal.

Best regards,

M. Schmidt
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<https://dovecot.org/pipermail/dovecot/attachments/20230218/5cd6d8a0/attachment.htm>