Re: RAC in NAS

  • From: "Mark Brinsmead" <pythianbrinsmead@xxxxxxxxx>
  • To: kutrovsky.oracle@xxxxxxxxx
  • Date: Mon, 31 Jul 2006 21:06:18 -0400

Slabtop -- yeah, discussed it to death (in a slightly different form)
already.
(I.e., in the form of "/proc/slabinfo".  "slabtop" is a good means of
watching
this value "on the fly", although Nuno's suggested use of "watch" provides
a more focused test.)

Its a decent way to test for AIO, but not conclusive.  Well, "negative"
results are conclusive, but "positives" aren't.  (Other applications, or
databases, on the server could be performing AIO and causing the "slab"
stats to change, and making it only *appear* that the subject database
is using AIO.)

The presence of IO_SUBMIT vs. PWRITE system calls *would* be
a conclusive test.  Sadly, we don't always have 'strace' available, but
I won't quibble over that right now.

I had *suspected* (from earlier comments in this thread, and maybe a
small snippet or two on MetaLink) that IO_SUBMIT might be the thing
to look for.  I guess you're confirming that suspicion.  ;-)

(Oddly enough, having raised an SR on this subject, Oracle support
has yet to mention IO_SUBMIT.  Of course, about half of the other
"answers" they provided were wrong, so even if they *had* mentioned
IO_SUBMIT, I may have overlooked it...)

Really, though, a simple, quick, definitive test might be in order.
Something
like V$ASYNCH_IO, or maybe V$IO_OPTIONS, or...

I *know* the information I'm looking for is buried in the database.
*Somewhere*.

Oh well, I now know how to use 'strace' to test for both AIO and DIO.  (On
Linux, anyway.)  That should be enough for my immediate needs -- and
then some.  But I still think it would be nice if Oracle Corp provided us
with
a method that was both reliable and portable.  I'm pretty sure they could
if they wanted to...


On 7/31/06, Christo Kutrovsky <kutrovsky.oracle@xxxxxxxxx> wrote:

have you seen "slabtop" ?

--
Christo Kutrovsky
Senior Database/System Administrator
The Pythian Group - www.pythian.com
I blog at http://www.pythian.com/blogs/


On 7/29/06, Nuno Souto <dbvision@xxxxxxxxxxxx> wrote: > Mark Brinsmead wrote,on my timestamp of 29/07/2006 11:55 AM: > > > But here's the half where you're wrong. (Or at least, not completely > > "right".) Those hypothetical > > "other" applications can just as easily be resposibly for the > > *increases* in the slabinfo stats as > > they are they would be for "non-zero" values. After all, those values > > had to increase from zero > > *some* *time*, didn't they. ;-) > [....]




Cheers, -- Mark Brinsmead Staff DBA, The Pythian Group http://www.pythian.com/blogs

Other related posts: