Re: RAC in NAS

Agreed.  You can live without AIO.

You'd be *amazed* at the push-back I got from various sources, though, when
I suggested turning
on I/O slaves.  After all, that's "obsolete", isn't it?  ;-)      (Not!)

But it *sucks* when you don't *know* you're living without it.  I recently
saw this happen to a client
who had migrated from DAS to NFS (without doing enough homework).  The first
hint that AIO was
not in use manifested months later with the deployment of a new application
and the sudden appearance
of huge numbers of block_buffer waits, incomplete checkpoints, and all
manner of other bad things
resulting from the (Synchronous) DBWR being unable to write fast enough.

Sadly, unless (and until) you know exactly *where* to look, it is almost
impossible to determine whether
or not the database is actually *performing* Asynch I/O.  (As most here
undoubtedly realise, looking
for DISK_AYNCH_IO=TRUE does *not* do the job!)

(Almost) just for chuckles, I opened an SR with Oracle support, asking
questions like "how can
I test whether my DB is doing Asynch I/O on Linux?" and "knowing that Asynch
I/O is unsupported,
what are the risks of doing so anyway?".  After almost two weeks, the
questions are unanswered,
even though I was able to answer them myself with less than an hour of
surfing Metalink and Google.

For the record:  "cat /proc/slabinfo", and "there is no risk -- the Linux
kernel will stop you" respectively
appear to be the answers to these questions.  And hat's off (again) to
Werner Puschitz for his excellent
website -- the information I found there seems to be a good deal better than
Metalink...

Anyway, this is a large part of where the homework comes in.  Those who fail
to do it probably won't
even suspect that AIO is not being used until *long* after they start using
NFS.  Of course, that might
beg the question: if it has no observable affect, is it *really* a problem?
But I'll leave that one to the
philosophers to ponder.  I *know* I won't bother to ask OSS. ;-)


On 7/26/06, Mladen Gogala <gogala@xxxxxxxxxxxxx> wrote: 


[...]

Asynch I/O is not that important. Oracle can emulate it using I/O slaves.
Granted,
it's not as good as the real thing, but you will not sufer much, either.
Direct I/O
is much more important and it is supported. FC5 is the sign of things to
come. It
does support full NFS4 version, with cient caching and async I/O included.
EL5 is
likely to have those features, minus bugs, discovered by free beta
testers, like
me.



--
Mladen Gogala
http://www.mgogala.com




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

Other related posts: