I originally wrote:
I've never worked with oracle b/4 and would appreciate hearing from
anyone
that's run benchmarks or has opinions of raid 1 vs. raid 5 for oracle
databases (not the redo and rollbacks). Our configuration is 2 X 8200,
ASE,
LSM, and redundant HSZ50's, DU4.0A Thanks.
I got responses from:
Mullin, Stephen [SMullin_at_gecits.GE.com]
Mike Sullivan [msulliva_at_emc.com]
sanghvi_at_proto.wilm.ge.com
Paul Grant [paul.grant_at_kscl.com]
Robert Schuhl [robert.schuhl_at_ald-vt.de]
Joseph Thvedt [jthvedt_at_brookings.net]
There seems to be a lot of folklore, myth, and religious views on this
topic.
What I had done was gathered up the output of 'volstat -g
oracle_groups', sorted
the data a couple different ways (increasing to decreasing: number of
reads, number of
writes, blocks of reads and blocks of writes), performed some
calculations (I took the
ratio of the number of reads to the number of writes and the ratio of
the blocks
read to blocks written) and suggested to our DBA that some of the data
would lend
itself to a raid5 configuration (where the number of reads to writes had
a ratio > 5).
This was a month's worth of data which I thought would show the
characteristics of
existing data to help design new databases. Our DBA said our existing
database wasn't
designed properly so the data is meaningless, and somebody told the DBA
that he doesn't
like raid 5 so there is quite a reluctance to even discuss the data.
Without knowing
a lot about oracle, I'd think you could point to data access
characteristics like I
gathered and make reasonable suggestions for disk configurations
(regardless of how
the database is designed, certain data is going to have certain
characteristics and I
don't see that changing, though the characteristics of the index and
control files may very well change). Responses I received follow, I've
made a couple comments
along the way in square brackets[]. Thanks to all who replied. Jeff
Stephen Mullin wrote:
Long time no talk... I've left the DEC CSC and now am doing SAP/Oracle
on 14 DU servers (2 clusters).
I'm LSM mirroring RAID5 on separate HSZ's, works well and from tests
I've done Not much impact due to Digital internal HSZ code. With large
number of
RAID5 members,I think more than 4 it initializes the container into
RAID3 i.e.
a six disk RAID 5 actually does 5 as Stripes and 1 as parity
The tests I did seem to confirm this is still the way they do it, I
learned of it years ago... on a five minute write got 20sec difference.
(24 HSZ50's, 128meg cache with 64k chunksize, 256 transfer size)
BTW:
Mirrored RAID 5 has the benefit of being "Self Healing" too, less mirror
repairs, and transparent to cluster failovers...etc...etc.
Mike Sullivan wrote:
The answer depends on the benchmark. A write intesive benchmark
which does many sorts/inserts/updates/deletes would be better on
RAID 1 where as a read intesive benchmark may work just fine on
RAID 5 volumes. Another area to look at is how random or sequential
are the queries to deciede if you would like to implement some form
of striping.
sanghvi_at_proto.wilm.ge.com wrote:
We have not noticed any differnece between mirroring and raid 5.
Paul Grant wrote:
Raid 5 has data redundancy if you lose 1 disk (max). Raid 1 has full
data redundancy,
as long as you don't loose a pair in one cabinate then there should be
no problem.
Design your raidset to go across channels (ports) rather within a single
channel....
Raid 5 has 4 writes per write where as RAID 1 has 2 writes per write.
Therefore using
raid 1 will be and robust and Raid 5 will be slower and much less
robust.
[when mirroring across different scsi buses, the host sees 2 writes
where it would
only see 1 write for raid5ing to a single scsi, with the HSZ calculating
parity and
doing the additional writes]
You also loose around 30% of your capacity due to parity requirements on
RAID 5. You
loose 1/2 of the capacity on RAID 1.
[DEC raid5 contains 3 to 14 disks so the capacity 'lost' ranges from 1/3
to 1/14 or
from 33% to 7%]
Depending on the number of disks and the use of your then the advantages
are varied.
Cost will be your overall consideration however as disks are "cheap" in
relative terms
I would advise mirroring with RAID 1....... The extra cabinates etc. are
also a
consideration......
Robert Schuhl wrote:
We're using Oracle with RAID 5. Permance may be a bit slower than on a
JBOD, but it saved already my live. Be sure to use disks capable working
in
a RAID, not all do so (DIGITALS do of course), some always use their own
cache, and that can lead to data loss. ANd of course: you need a kind of
battery backup, on the controller or external, power loss is the biggest
problem for RAID's.
Joseph Thvedt wrote:
RAID 5 can have its place in a database installation, but used
indiscriminately, it will probably lead to disappointing performance.
It offers a very attractive cost per unit storage. If your application
does even a moderate amount of writing (and what application doesn't?),
you'll find that RAID 5 probably isn't the way to go. But, if your
application can stand it, RAID 5 provides fault tolerance at a
relatively low cost. You might consider mixing & matching different
RAID and non-RAID disks for different parts of the database.
[this is what I tried proposing to our DBA]
An excellent discussion may be found here:
http://sunsite.cs.msu.su/sunworldonline/swol-09-1995/swol-09-raid5.html
[this was a good article]
Received on Fri Nov 21 1997 - 19:45:45 NZDT