Jerome Grandjanny wrote:
>
> Hello all Oracle gurus,
>
> May be this question has been already discussed.
>
> I'am in charge of designing a disk layout for a new database.
>
> I need to decide wether I use RAID 0 (stripping) or RAID 5 for
> the tablespaces containing the data and the indexes.
>
> The database must run 24/24-7/7.
>
> In some books about Oracle, I read that using RAID 5 is not always a
> good choice for data (possible performance problems during write
> operations).
>
> In the other hand, in case of a disk crash, a hot plugging RAID 5
> system can offer a non-stop recovery mechanism.
>
> Please, let me know your opinion, your experience, ...
> Any advice welcome.
> Thanks.
> Sincerely,
> Jerome.
I am revisiting this question, as this is a very interesting one, and as
a lot of sites invariably get into this question and there is no single
correct answer. Plenty of threads were opened on this topic in the past
in various news groups. I am attempting to summarize my findings here
after going through these articles:
- There is no standard on the functionality and implementation of the
various RAID versions, namely 0,1..7.
While mostly its agreed that RAID0 is striping, RAID1 is mirroring and
RAID0+1 is striping and mirroring, other versions differ in functions
and algorithms from vendor to vendor. For example with some vendors
RAID4 is striping with a single parity disk and RAID5 is striping with
parity on a round robin basis. Another vendor may not even have RAID4 in
his product line, he may be terming these versions as RAID5 and RAID7
respectively. And for the same reason one can never say RAID5 is bad,
RAID5 of this vendor is bad is more appropriate.
- The RAID technology is improving over years, vendor are coming up
with newer and newer ways to improve the performance. So, the
performance statistics one year back could be quite different from what
is now. Currently the prominent methods vendors are adapting to improve
the performance are on-board caching (used as conventional cache by some
and used to piggy bag the blocks by others) and parity on a round robin
basis which reduces the disk contention (of the parity disk if one
exclusivley assigned). The other improvemnets are fault tolerant level
(more than one failure), hot spare and non-volatile memory
(write-behind).
- And you have another option in RAID's: soft RAID. While this is very
cheap compared to hardware RAID's, I did not come across any article
advising in favour, reason being the CPU overhead is excessive.
- Its generally known that parity RAID's (RAID3 and up) slower than
RAID0, RAID1 (degree of slowness varies with each vendor), improvments
are made on parity RAID's to match the performance. While is true that
if you run RAID0/1 in these improved configurations they will run even
faster, reason to go for parity RAID solution is that you have now a
fault tolerant system with the accepatble performance level (may be
equalling that of a simple striping or mirroring) and at a reasonable
cost (than say mirroring a larger database).
- Measuring the performance is another great task. A single task
benchmark may favour a RAID configuration, running several tasks may
produce a discouraging result. Think time between i/o requests (real
life situations) also may vary the performance. For this reason the
results/opinion found in various articles greatly differ depending on
the benchmarking scheme they have used. Simulating the real time
application (not a easy task) or running the real application (again not
a easy task) will be the best test.
- Next million buck question, once you have decided a RAID
configuration, is: what is the stripe size. I didn't find any definite
answer here as well. Some favour large stripe size (64K and upward),
arguing that the blocks will be found in contiguos in any-one disk which
will reduce the disk seeks. Some favour the smaller stripe size to take
advantage of parallel reads (of disks).
And more and more interesting and useful information in these articles.
You can very easily get confused and lost or come up with a good answer.
Received on Sun Aug 17 1997 - 00:00:00 CDT
