Re: V880 LVM and Oracle performance
From: Juhan Leemet (juhan_at_logicognosis.com)
Date: 10/11/04
- Next message: Geoff Lane: "Re: frustration with Solaris"
- Previous message: Mike: "Re: 2,000 dollars for a compiler!?"
- In reply to: Rodrick Brown: "Re: V880 LVM and Oracle performance"
- Next in thread: Darren Dunham: "Re: V880 LVM and Oracle performance"
- Reply: Darren Dunham: "Re: V880 LVM and Oracle performance"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: Mon, 11 Oct 2004 17:51:21 -0200
On Mon, 11 Oct 2004 15:32:19 +0000, Rodrick Brown wrote:
> On 2004-10-11 10:46:58 -0400, Artur Karazniewicz
> <karaznie@[n-o-s-p-a-m]aster.pl> said:
>> I'm about to install new V880 with Solaris 9. We are thinking about adding
>> LVM support on this box (for better management) and RAID. Unfortunately we
>> have no big experience with LVM. This box is dedicated for our app server
>> and oracle database. We migrate from V280R and hope we will have much
>> better performance on new boxes (especially when it comes to I/O). I wonder
>> how LVM (and software RAID) can impact Oracle database especially regarding
>> I/O.
>>
>> Usually for the best performance we try to follow practice to put Oracle's
>> indices and datafiles on different disks. It gives us better performance -
>> especially when heavy disk operations are performed by Oracle. We plan
>> install regular database (won't use raw devices).
>>
>> Does LVM and RAID have big impact on I/O performace (of course it has, but
>> how big is it?) in scenario described? What options/settings would You
>> prefer in situation above?
>
> Ofcourse it does quick and simple answer you are writing against more
> spindels so the I/O across that volume would be much greater than say
> writing to one dedicated disk.
>
> Give it a shot for yourself and see.
>
> metainit d50 -r disk1 disk2 disk3
> newfs /dev/md/rdsk/d50
> mount -o noatime,directio,logging /dev/md/dsk/d50 /u01
This is a S/W RAID5 definition, isn't it? It is not so obvious, and I
don't quite understand which way you are suggesting the performance will
go (up or down)? I agree that if the OP tries it he should be able to see.
My own experiences (informal measurements) showed that S/W RAID5 can
multiply the number of I/O operations (for writes) by up to a factor of 5.
I am not sure how that depends on the number of disks in the RAID5
metadevice? I had 5 (coincidence?). Darren Dunham in his post mentions
similar numbers. The net effect seems to be that writes (throughput and
hence also latency?) could often be 5x slower. I believe that this is
because the RAID software has to read the previous data and/or parity info
from disks and combine it with the new data to be written back out to
disks. Reads can be faster, because (as you point out) they are spread out
over a number of spindles (no parity change).
I would say that for database use (many writes?) S/W RAID5 is not a good
idea. In my own case, I am using S/W RAID5 for a code repository (mostly
reads, few writes) so it works OK for me. I would not use it for a normal
database. If you can get hardware RAID, that could be a different story.
For databases I think RAID1 (mirroring) would be good. It would improve
reliability (against disk failures), and improve read performance (by
spreading over spindles), and not significantly penalize writing (just
parallel write commands, as long as SCSI bus data bandwidth not hitting
limits). I would do some benchmarking with expected usage patterns first.
-- Juhan Leemet Logicognosis, Inc.
- Next message: Geoff Lane: "Re: frustration with Solaris"
- Previous message: Mike: "Re: 2,000 dollars for a compiler!?"
- In reply to: Rodrick Brown: "Re: V880 LVM and Oracle performance"
- Next in thread: Darren Dunham: "Re: V880 LVM and Oracle performance"
- Reply: Darren Dunham: "Re: V880 LVM and Oracle performance"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|
