Re: Memory Sizing Advice
Date: Mon, 12 May 2008 13:38:48 -0700 (PDT)
On May 11, 8:00 am, Pat <pat.ca..._at_service-now.com> wrote:
> On May 10, 8:00 am, "Mike Jones" <mj20..._at_hotmail.com> wrote:
> > "Pat" <pat.ca..._at_service-now.com> wrote in message
> > > The question I have is, is there any downside to me buying, say, a 32G
> > > box and setting the SGA size at 20G? Will I actually end up harming my
> > > performance with an over-large SGA (assuming I have enough physical
> > > memory to keep the box out of swap)?
> > I am in agreement with bhonaker. I do not see any issues with doing this.
> > However, as others have indicated, this may not solve your problem. But I
> > don't see it causing harm. Is it possible you could comment on the
> > configuration of your current system as well as the configuration of the
> > replacement system?
> > Mike
> Current system is:
> 4 X Intel(R) Xeon(TM) CPU 2.80GHz (8 cores)
> 8 G RAM
> 1.1 TB of storage (SCSI RAID 10 array, direct attached)
> Red Hat AS4, 32 bit
> Oracle 10.2.0.3 (32 bit)
> 2.3 G SGA
> Replacement system is:
> 2 X Intel(R) Xeon(R) CPU E5310 @ 1.60GHz (8 cores)
> 16 G RAM
> 1.6 TB of storage (SCSI RAID !0 array, direct attached)
> Red Hat
> Oracle 10.2.0.3 (64 bit)
> 12 G SGA
I pretty much agree with the others in this thread, though I have a quibble with a statement in David's first post: "> > A smarter solution is to add an index to reduce I/O by orders of magnitude.
Possibly. It's also possible that this is a highly transactional configuration, effecting large volumes of data changes. An index won't help there"
What I've seen in most transactional configurations is that indices help a lot. What often winds up happening is certain objects thrash the SGA, so Oracle spends a lot of time doing housekeeping on the buffers there, as well as all the other things David mentioned. What I've seen on my systems is that there are usually just a few objects responsible for most of the thrash (which makes sense, since most activity is the mission critical app), and separating them into their own cache (particularly a recycle cache) helps enormously. See the performance tuning manual for the basics about looking at V$BH, and google about for other commentary about it. But definitely don't listen to anybody who advocates using tablespaces with different block sizes for performance.
If you look at what is happening on your current configuration, I'd expect you'll see there's not all that much size involved in the candidates for recycling. Then when you look at the total PGA usage you'll see how much your users really need. Check out your advisor views.
The more modern Oracle handle large SGA better than in the olden days, so using an oversize SGA isn't really the big deal it was in the past, as long as you aren't way stupid in something.
> I've got the hardware budget left to bump up the memory on the new
> system a bit if that'll help and I'm tempted to do it (hence this
My experience indicates it is best to grab what you can when you can,
> One totally off topic question I do have though is, does anybody have
> experience with the new 4 core line of intel chips? My operations guys
> are procuring these now instead of the old dual core Xeons, but they
> make me nervous since the new ones report lower bogomips than the old
> ones. I'm aware that the "bogo" in bogomips stands for bogus, so I'm
> totally wiling to be convinced the new chips are, in fact, faster, but
> I'd love to hear any real world experience here.
Tuning becomes complicated because it involves twiddling many different bottlenecks, so it is entirely possible to make things worse in an unexpected way. For example, making the SGA larger could take the bottleneck away from I/O and put it onto cpu, suddenly making bogomips important, and making everything slower in both configurations, even more so in the "faster" one. That's when it is important to already understand methodologies like Cary Milsap publishes. I'm not disagreeing with Mladen, in fact I agree with what he posted, but I'm pointing out that lots of factors are involved that can make a specific situation "interesting."
-- @home.com is bogus. "...the Internet is the best way to disseminate accurate information..." - Mark Cogan, attorney for alleged baby killers. http://www.worthingtondefense.infoReceived on Mon May 12 2008 - 15:38:48 CDT