Oracle FAQ Your Portal to the Oracle Knowledge Grid
HOME | ASK QUESTION | ADD INFO | SEARCH | E-MAIL US
 

Home -> Community -> Usenet -> c.d.o.server -> Re: Performance Question: RAID or Individual drives ?

Re: Performance Question: RAID or Individual drives ?

From: Art S. Kagel <kagel_at_bloomberg.net>
Date: Mon, 13 Jul 1998 15:00:59 -0400
Message-ID: <35AA596B.464F@bloomberg.net>


SV wrote:
>
> If you think what you said is the whole story
> then you're wrong.
>
> There are plenty of other things which can
> influence I/O perfomance.
> SCSI controllers (Fast/Wide or Ultra Fast/Wide, or narrow ),
> RAID disk cache configuration (write-back or write-through),
> software or hardware striping,
> OS, etc.
>
> You mentioned neither of them, so your numbers
> are pretty much useless for any intent or purpose
> other than for your particular system.
>
> About the only thing I would agree with out of the box
> is that parallel query does seem to improve total I/O
> disregard of other things.

I have to agree with Sergei. I work more with Informix than Oracle but in my testing, both with Informix and with "C" test programs (to eliminate the database from biasing the test), I have found that the controller and RAID software/firmware can have a BIG effect on performance. Multiple intelligent Ultra Fast/Wide controllers with lots of cache (ie enough for several entire stripes say 4MB) can result in improved performance for all RAID configurations. That said Motox is right about the write penalties of RAID5. However, he did not seem to test the best configuration of all for databases: RAID10 or 1+0 as Sergei mentions, which is a stripe of mirrored pairs.

We use a 30 drive stripe spread over 4 Ultra Wide/Fast controllers and get 30MB/s write throughput for write with 50 processes WRITING to the array, read throughput is much higher! Each process, obviously, is seeing much less throughput than that (about on a par with Motox's testing, but adding more and more processes continues to increase total throughput; I just gave up testing when our performance goals were surpassed. This is exactly the behavior that one gets with modern database parallelization such as Oracle and Informix provide.

Also noone mentions RAID3 which is as fast for reads as RAID5, does not suffer the same write penalties, and gives better protection than RAID5 which suffers from not correcting media corruption.

It looks to me like Motox's testing is using a single Wide or Fast only controller and is hitting his controller's I/O limit (10MB/s).

BTW: All redundant RAID configurations (1,3,4,5) suffer from performance degradation during single drive recovery of over 90%. The sole exception is RAID10. Since only one mirrored pair from the RAID set is being rebuilt in RAID10, and since only one drive needs to be read to do that, the performance degradation due to recovery is <1/N where N is the number of mirrored pairs in the RAID10 set. Therefore, one thing to consider heavily is how much performance can you afford to give up if a single drive fails and needs to be replaced and recovered? If the answer is not much then bite the bullet and spend the dollars (or whatever) and go with RAID10.

Art S. Kagel Received on Mon Jul 13 1998 - 14:00:59 CDT

Original text of this message

HOME | ASK QUESTION | ADD INFO | SEARCH | E-MAIL US