psuedo-swap and oracle

[Rich Owens_2]

Should psuedo-swap (swapmem_on) be utilized on an Oracle database server? I have read conflicting opinions on the subject.

[harry d brown jr]

I've seen a few posts on it and most agree that it should be on, but I'd like to see some other opinions on it also.

live free or die
harry

[Sanjay_6]

Hi Rich,

Here is a thread from ITRC on this,

http://us-support2.external.hp.com/cki/bin/doc.pl/sid=4cb8cd970246cdd1ee/screen=ckiDisplayDocument?docId=200000024605263

Here is some detail on the kernel parameter swapmem_on,

http://docs.hp.com//hpux/onlinedocs/os/11i/kcparams/KCparam.SwapMemOn.html

Hope this helps.

Regds

[Tracey]

What I've read about it here on the forums is that to have it on is "basically" harmless, but to have it off, and then the system need it, you are SOL. Currently my 11.x system is using it with sampmem_on and I am experiencing pretty good performance.

[Roger Baptiste]

hi,

I have experimented with this value and based on that i would suggest to turn
it on . There is nothing to lose, since it is putting your memory to good use.
The only exception being, if you have less physical memory on the system (1Gb or less), because then you wouldnt be bothered with swap more than 2Gb. But since yours is a DB server , i guess this is not the case.
So, go for it!

Sometime in October, there was a good discussion on this topic. I would suggest your to go through it. I could not get the link, but i am sure Sanjay can dig it up in a jiffy ;-)

-raj

[Rich Owens_2]

Thanks to all for the replies.

[James R. Ferguson]

Hi Rich:

First, the standard disclaimer -- your mileage may vary... Now, make your own choice based on the following.

Knowledge Base document #KBRC00001129 (What is Pseudo Swap?) gives a good an explanation of pseudo swap in the first place (see attached for your convenience).

The best rationale for *not* using it is provided by Stephen Giullo and Doug Grumann in the Performance Cookbook white-paper:

http://h21007.www2.hp.com/dspp/files/unprotected/devresource/Docs/TechPapers/UXPerfCookBook.pdf

Regards!

...JRF...

[A. Clay Stephenson]

Hi Rich:

My answer to this would be it depends. The ideal situation is to have enough memory so that you know you will never swap. In that case, I find that my Oracle servers are fine with only 256MB-512MB of mirrored primary swap, swapmem_on=0, AND about 1x memory size configured as filesystem swap (mirrored lvol) at very low (hi number) priority. The nice thing about this is that you don't blow a bunch of disk that does nothing. Filesystem swap eats up nothing until it is actually used.
Normally filesystem swap is a bad thing but it this case it's there as a safety net and essentially as a resource enabler.

I especially dislike the old rules of 2x-3x (4x-6x) swap rules when I intentionally buy enough memory to never swap in the first place.

If you are truly swapping, then swapmem_on=1 does nothing for you and in this case you need device swapping over as many paths as possible.

The only time that swapmem_on=1 should be set is the case where a machine has much more memory than swap; it then becomes a resource enabler. Having said all this, in machines with large amounts of RAM, I have never seen big differences in performance with the paramter set either way.

One final note: If you do have a small primary swap then you need to set up a dedicated unmirrored dump space but this is still much better than blowing a tomn of disk than won't really be used.