Any effect of sub-optimal setting of Kernel SEM tunables?

From: Rich J <rich242j_at_gmail.com>
Date: Mon, 22 Jun 2020 09:28:47 -0500
Message-ID: <CAANsBX15pHXLznchNaWJWy056LSNmhc4D6TyT3P_W1aO2hogOA_at_mail.gmail.com>



Hey all,

So I'm migrating from the almost-auto-tuning kernel of AIX to OL7. I have five 19.6 instances on a VM with 288GB of RAM, with a total PROCESSES count of 19000 (two with 5000, three with 3000). The Kernel semaphore tunable is the 19c pre-install default of "250 32000 100 128", so SEMMSL*SEMMNI=SEMMNS.

Various MOS articles point to setting SEMMSL to the maximum value of any instance PROCESSES setting, or that setting "+10". The oldest and most-referenced article, MOS 15654.1, says "no smaller than the smallest 'processes'". And then it follows that up with "This is not a requirement though." Not helpful...

"ipcs -s" lists 81 semaphore sets of 250, with a couple other sets of a few semaphores each. This should leave plenty of room for the 32000 maximum, although Oracle's allocated many more semaphores than I would have expected, based on my interpretation of the documentation and MOS articles.

What's the advantage, if any, of reducing the number of semaphore sets from that 81 to 5 or 6 by increasing SEMMSL?

Thanks,
Rich

--
http://www.freelists.org/webpage/oracle-l
Received on Mon Jun 22 2020 - 16:28:47 CEST

Original text of this message