Re: OFA Architecture with RAID

In the case of a large datawarehouse style database I find that RAID5 really does have a place. Sure, you accept that the weekly (and/or monthly) update is going to be slower, but you live with that. What you get in return is a good degree of fault-tolerance for a reasonably cheap price.

Of course a two or three way real mirror (RAID0) is the best route (and you can even break a mirror for backups), but that means 2 or 3 times the cost of the disk for you database, which can be large for a large database. I think RAID5 is good in these situations where you just don't have that sort of budget.

Steve Phelan.

gremlin wrote:

> I agree, I have always heard that RAID 5 is noticeably slower. What if I have, for example, a
> large database that is primarily just read-only? I'd update the data (significantly) once per
> week, but the users would be doing queries, not generating a lot of new data. Is RAID 5 and
> advantage, neutral, or a disadvantage in this case? Just curious.
>
> Mike
>
> Steve Phelan wrote:
>
> > Well, not really. The problem with RAID5 arrays is their generally very poor write
> > performance. Of the 3 RAID5 arrays I've tested this has been 50-200% slowdown compared to
> > tqking the same drives in a striped array (i.e. RAID0). A write cache can get round this, but
> > this is a dangerous route to take if you lose power to the server - potential database
> > screw-up.
> >
> > The problem is that much of your database may actually be very write intensive, especially
> > redo, archive, temp and maybe your datafiles if that's what your application is like.
> > Distributing the I/O is good, but mixing eveything together and hoping that the RAID5 array
> > mix will some how sort it all out is not the way to go. A database's I/O is a far more
> > complex mix, and you've got to be prepared to make some intelligent loaction decisions.
> >
> > Like I said, best thing is to try it and monitor it. Start moving things around when you see
> > bottlenecks and then re-monitor.
> >
> > Regards,
> >
> > Steve Phelan
> >
> > (Oracle 7 & 8 OCP)
> >
> > gremlin wrote:
> >
> > > Well, that was my thinking with the one mount point. If the file system is RAID (5 or 3)
> > > then you really don't get anything by having different mount points, they all end up as
> > > one in the end anyway! I would think this would give very good performance, if you have
> > > n physical drives and n processors then the I/O is distributed as much as you can get it,
> > > isn't it? If you split it up (say, index tables on 2 drives and data tables on 3
> > > others?) then you still wait for the slowest one anyway I would think. Not that I claim
> > > to be any kind of expert.
> > >
> >
> > [snip]
>
> --
> -----------------------------------------------------
> To reply to me via email, remove the "__NO_SPAM__" in
> the header email address.
> -----------------------------------------------------
Received on Sun Feb 28 1999 - 05:12:09 CST