Re: shmmax sizing recommendations

  • From: Don Seiler <don@xxxxxxxxx>
  • To: Yong Huang <yong321@xxxxxxxxx>
  • Date: Thu, 19 Nov 2009 12:39:17 -0600

1. Yes memory_target=0.  I'm aware of the issue with AMM and HugePages.

2. Even if no harm, I'd feel better setting the shm* to values that
the host would be able to handle.

3. Thanks for that note.  That seems to be the general concensus re
multiple fragments.

Don.

On Thu, Nov 19, 2009 at 12:13 PM, Yong Huang <yong321@xxxxxxxxx> wrote:
>> While troubleshooting an ORA-04030 issue, Oracle Support also took
> ...
>> Oracle 11.1.0.7 64-bit on OEL 5.3.  The server has 64gb of physical
> ...
>> SGA sizes of ~42 gb and HugePages configured for 44gb.  The largest
>
> I highlight three points in your message: ORA-4030, 11g, HugePages. You
> probably already know that to use HugePages in 11g you need to turn off
> AMM by setting memory_target to 0. Can you double check? Make sure SGA
> *is* using HugePages, otherwise you would squeeze the memory available
> for not only SGA but also PGA to a small number.
>
> I have a note about HugePages:
>
> http://yong321.freeshell.org/oranotes/HugePages.txt
>
>
>> In the past I have set SHMMAX to either 1) the physical ram or 2)
>> theoretical limit.  Since this is simply an upper limit, it really
>> has no performance "hit" for setting it too high.
>
> I remember Jim Mauro in his "Solaris Internals" 1st ed says shmmax is
> just a number checked to see if you attempt to create a shared memory
> segment larger than that. There's no other use. So there's no danger
> setting it too large. I think it's the same on Linux. Here's my setting:
>
> [oracle@dctrpcora1a ~]$ sysctl kernel.shmmax
> kernel.shmmax = 4294967296
> [oracle@dctrpcora1a ~]$ ssh dctrpcora1b /sbin/sysctl kernel.shmmax
> kernel.shmmax = 68719476736
>
> On both RHEL boxes, I have only 4 GB physical RAM. But I tested setting
> shmmax to 68 GB on one of them without any issue.
>
>
>> There is some overhead in having multiple shared memory segments vs.
>> a single one (for any given instance) but I dont have any numbers
>> in my pocket to qualify that.
>
> Steve Adams in his O'Reilly book says having multiple fragments in a
> shared memory segment slightly slows down instance startup and server
> process creation, with no other negative impact. But if you have very
> frequent process creations, it's better to have one big shared memory
> segment.
>
> Yong Huang
>
>
>
>



-- 
Don Seiler
http://seilerwerks.wordpress.com
ultimate: http://www.mufc.us
--
//www.freelists.org/webpage/oracle-l


Other related posts: