Noons wrote:
> Fabrizio wrote:
>
>
>>disk_async_io? I was referring to OS aio but since we are here:
>>disk_async_io is enabled by default on 9i (the OP version).
>
>
> Yes, but not active! Check out the Metalink stuff on
> aio:
>
> HPUX: although it is enabled from Oracle, it is not available
> in HPUX cooked IO. You can only use RAW. And in fact, for
> ia64 HPUX 11 boxes there is currently a bug with 9ir2 that
> causes a log dump to be produced in $ORACLE_HOME/rdbms/log
> every minute! Due to Oracle forcing the use of an unauthorised
> AIO in HPUX. There is an arcane workaround, but the whole
> thing is iffy and shaky to start with. DAMHIKT, 150000 log
> dumps later...
>
> RHAS: you need a number of patches to both RHAS3 and 9ir2 to
> make aio and dio work even though it is enabled by default in
> Oracle. In fact, even if you use filesystemio_options=setall,
> you STILL won't see an strace of open(O_ASYNC) unless you got
> ALL the patches in there. Same for dIO. This is of course for
> cooked IO. The situation is slightly better for RHAS4 (kernel
> 2.6), but 10g is the only one certified there...
>
> Solaris: I strongly suspect Sy is spot on, it only works
> well in 9ir2 with RAW.
>
> To fix all this mess you can of course invest in Veritas file
> system and its extensions, but watch your pocket!
>
> I really, really wish Oracle would clean up this TOTAL MESS of
> cooked aio/dio in 9ir2! Quite frankly, their attitude of only
> fixing 10g is causing a lot of bad feelings with sites that
> cannot yet upgrade but need the features NOW. Once again
> Oracle marketroyds are doing their best to shoot themselves
> and their product in the foot! Some things never change...
>
I'm not going to comment on HPUX, as I few experience there, but aio on
fs on solaris works. I don't have a testcase (no spare sun machines for
tests here) but matalink note #48769.1 is a good place where to look.
That explain why truss is misleading when checking for async i/o.
For redhat (and this is one of the reasons I prefer suse) the main patch
for aio is 3016968 which does only insert a new libaio amongst the
oracle libraries, overwriting the default OS behaviour.
(For direct i/o and "concurrent i/o" the discussion is far more
complicated).
I envy you endurance on 9i... honestly I gave up... I'm migrating
whatever I can toward 10g. Oracle defeated me.
Regards
--
Fabrizio Magni
fabrizio.magni_at_mycontinent.com
replace mycontinent with europe
Received on Thu Jun 23 2005 - 02:46:18 CDT
