Oracle FAQ | Your Portal to the Oracle Knowledge Grid |
Home -> Community -> Usenet -> c.d.o.server -> Re: Async I/O
Unless you are running mostly read workloads, you should not use
RAID-5. Spend a little more money and run RAID 0+1, you will
see great improvements in terms of performance.
I might not have been clear in my previous message. If you specify async I/O with JFS on 10.20, we know for sure you will not get async I/O. So you will not get the expected performance boost. But by doing so it may have bad side effects.
Async I/O is never a problem for data integrity if you want to use it with database running on raw or VERITAS Quick I/O files (available on Solaris 2.5.1 and 2.6 and soon on HP-UX 11). So you see there are no contradictions. I will say the same thing "on or off the record". :-)
David Lai
VERITAS
opus wrote:
> Again, I thank all for responding as everyone has been very helpful:
>
> David, one more question if I could and some clarifications. I
> am using VxFS on HP but I just reference it as JFS. Further to
> that the database files are on a RAID5 configuration.
>
> You say the behavior is UNKNOWN but then you say data integrity
> isn't a problem and you don't get any performance gain to
> begin with. If the behavior is unknown isn't this a contradiction?
> I know you work for Veritas so how about some thoughts
> "Off the Record" then...
>
> Opus
>
> David Lai wrote:
> >
> > Hi Opus,
> >
> > All I can say about it is since it is not a supported config, the
> > behavior is UNKNOWN. It is unsafe at best to use async
> > I/O with JFS on HP-UX 10.20. I don't believe you will suffer
> > in terms of performance because I don't believe you get any
> > gain to begin with.
> >
> > I also noticed there are some discussions about how your
> > data integrity will suffer when using async I/O. It is not true.
> > To put it simply without going into details, all DBMSs today are
> > designed to take advantage of OS level kernelized async I/O for
> > better performance without compromising data integrity.
> >
> > I will also clarify one point. JFS is a "lite" version of VxFS. The
> > full function VxFS on HP is called OnlineJFS.
> >
> > David Lai
> > VERITAS Software
> >
> > opus wrote:
> >
> > > Hi David, and thanks for responding ....
> > >
> > > So it doesn't work at all or would I get some sorta error message or
> > > it (as in the next reply) a matter of ensuring data integrity
> > > and that could be a potential problem. So far, I've had no
> > > problem but then... There's no way we're going to 11.0 anytime
> > > soon I doubt. So you recommend turning it off period then? Will
> > > I see some preformance degredation in disk writes or nothing
> > > as it doesn't really work?
> > >
> > > thanks again....
> > >
> > > David Lai wrote:
> > > >
> > > > JFS is the name for VERITAS File System (VxFS) on HP. Indeed it
> > > > does not support async I/O on 10.20. So you should not be using
> > > > async I/O right now. In a few month we are releasing
> > > > the VERITAS Storage Edition for Oracle on 11.0. This product
> > > > includes VxFS 3.3 that allows async I/O on files (with the VERITAS
> > > > Quick I/O driver). As a matter of fact, these Quick I/O files behave
> > > > and perform just like raw devices. On Solaris we have seen over 100%
> > > > performance improvements over traditional cooked files for TPCC
> > > > type workloads. We expect the same on HP as well.
> > > >
> > > > -- David Lai
> > > > VERITAS Software
> > > >
> > > > opus wrote:
> > > >
> > > > > Hello all,
> > > > >
> > > > > I'm running 7.3.4 on HP 10.20 and I've enabled async I/O in the kernel
> > > > > and in the init file. someone in Oracle support said to do that many
> > > > > moons ago even though it says only use async i/o with raw device files.
> > > > > I use JFS files. Now someone else from Oracle support sends me FAQ
> > > > > saying I shouldn't use asyn I/O with JFS but there's no reason as to
> > > > > what can happen if you do. So far been running like this for quite
> > > > > awhile. Any thoughts?
> > > > >
> > > > > I've also been running an Oracle 8.0.3 instance using async i/o (it
> > > > > defaults to that outta the box fer gawd's sake - all ya gotta do is set
> > > > > it up in Unix) for awhile as well.
> > > > >
> > > > > Reply to the newsgroup please, and thanks if anyone knows..
> > > > >
Received on Tue Mar 23 1999 - 12:18:03 CST