On Wed, Dec 25, 2024 at 01:08:29AM +0300, Vitaliy Makkoveev wrote:
> As I understand you tried the diff which keeps timeout and task
> reinitialisation. Could you try the last one which uses barriers
> instead?

I started the test before you posted the updated diff.  I started
a new run afterwards and then went to bed.

> Also, could you share your performance tests to me?

The sources are public, but it needs some major setup of machines
to get it running.

https://github.com/bluhm/regress-all
perform.pl is running the actuall tests.

At the end the results are on a web page.
http://bluhm.genua.de/perform/results/2024-12-23T10:08:33Z/perform.html
The newest diff is in the middle column.

I think it is unrealistic that you rebuild this setup.  It needs
two OpenBSD and two Linux machines.  They have a lot of special
cables and configuration.

Better you send me the diffs and I test it.  Expect that it takes
some hours.  I have to read your mail, wait until the machine is
free, wait until the test run finishes, analyse the results, and
reply a summary of what I see.

bluhm