Oracle FAQ Your Portal to the Oracle Knowledge Grid
HOME | ASK QUESTION | ADD INFO | SEARCH | E-MAIL US
 

Home -> Community -> Usenet -> c.d.o.server -> Re: NUMA init.ora settings

Re: NUMA init.ora settings

From: Dave <dherri_at_acxiom.com>
Date: 7 Sep 2001 14:01:10 -0700
Message-ID: <d31a72dc.0109071301.1c0fa9c9@posting.google.com>


Andrew Mobbs <andrewm_at_chiark.greenend.org.uk> wrote in message news:<WPu*fEI5o_at_news.chiark.greenend.org.uk>...
> Dave <dherri_at_acxiom.com> wrote:
> >Folks, I'm running an Oracle 8.1.7.2.0 database on a GS160 16x16
> >(Tru64 5.1) and have noticed 3 hidden NUMA-related params:
> >_NUMA_instance_mapping, _NUMA_pool_size, and
> >_enable_NUMA_optimization. The latter is set by default to TRUE, but
> >the other 2 are currently "Not specified". Does anyone have
> >experience with these, such as suggested ranges of values and ways to
> >test their effectiveness?
>
> I didn't modify any of those when I ran a GS320 benchmark. The NUMA
> optimisations apparently do things like give separate data structures
> for each QBB.
>
> I saw a marked benefit in binding volume processing processes from our
> application to individual QBBs. Forcing locality between the client
> and server processes seems to be much better than just letting the
> scheduler handle it. Of course, if your application involves only remote
> connections, then that's irrelevant.

I'm rather new to Compaq's NUMA architecture. Specifically, how did you bind volume processing processes to individual QBBs? Did you create processor sets and then reassign the PIDs to particular sets, as in using the pset_* and runon commands?

Although the box is a 16 x 16, we're only licensed for 8 processors for Oracle. We've created a pset of 8 processors (8-15). I then used aliases for all main Oracle executables to include the runon command, such as "alias svrmgrl='runon -p 1 svrmgrl'". I also created a script to run through cron every minute and check that all Oracle user processes are running on pset 1. If not, it reassigns them.

I also had our Unix admin define rad_gh_regions[2] and [3] to 1024, which I assume means that QBBs 3 and 4 have 1GB of RAM assigned to them for the SGA. I'm also assuming processors 8-15 are really on QBBs 3 and 4, but I don't know of a way to verify this.

I got a little long here. Am I on the right path? I am a little curious why Oracle gives those params but doesn't set them. Received on Fri Sep 07 2001 - 16:01:10 CDT

Original text of this message

HOME | ASK QUESTION | ADD INFO | SEARCH | E-MAIL US