Re: Q: Is RAID/striping good with Oracle???

"DAVID DEVEJIAN" <djd_at_bank2000.com> wrote:
>
>
>R Glenn Stauffer <stauffer_at_crabapple.cc.swarthmore.edu> wrote in article
>
><Random header information snipped>
>
>> Common wisdom has it that there is a significant write penaltly with
 Raid5.
>> I've also heard some suggest that you will lose the entire raid volume if
 you
>> lose two drives - I'm not sure that is any more likely than losing both
 drives
>> in a mirrored set.
>>
>
>To write to any point on the disk, the system must write to both the data
>disk that the data is bein placed on, and the parity disk. The effect of
>this is that you cannot write to two disks (data disks) at once.
>
>As far as the loss of a drive, yes, a RAID 5 array cannot recover from a
>two drive failure, as far as how unlikely this is?, well that depends on
>how many drives you are using. For example If you use three drives in a
>disk array, the probability of an unrecoverable loss is (d^2)/3 where d is
>the probability of loss on one drive, by contrast if you are using 4 drives
>mirrored (the same amount of data) the probability is approx. (d^2)/2.
>Since it can happen on either of two drives. Thus the likelyhood of an
>unrecoverable failure is 3/2 times more likely in a 3 drive raid 5 array.
>However, for many applications, this probability is so small as to be
>insignificant, or at least bearable. However, there are applications where
>you want the increased security.
>
>> If you run RAID5 on a transaction system without a caching controller
 with a
>> fairly large cache, you will see a performance hit. If your system is a
 read
>> intensive DSS, that shouldn't be as true (of course, Oracle is still
 doing
>> writes for sorts, etc). I read something when I was configuring our
 server
>> that you need something on the order of a 40mb write cache on the
 controller
>> to eliminate most of the RAID5 write penalty. Of course, then you have
 to
>> weigh the cost of more drives (RAID0+1) vs more costly controller
 hardware
>> for Raid5.
>
>Interesting thought, I had not heard that 40mb figure before, and I would
>have thought that it would be very system specific.
>
>Just my $0.02
>Dave Devejian
>djd_at_bank2000.com
>

If you use disk caching to make up for the cost of Raid 5 you've just moved your point of failure to the disk cache board. If it dies with unwritten data in it you might be worse off than if you didn't have it and lost your two disks. I can't speak for ORACLE specifically, but most database products write disk blocks out in a certain order and thus expect the first ones to be on disk when the last ones are written.

My 2 cents too!

Cheers
Jay Walters Received on Thu Aug 29 1996 - 00:00:00 CEST