Disk Drive Configurations
Dave Goldsmith
NT server supports four disk drive configurations two of which provide
fault-tolerance.
Disk Mirroring - keeps a duplicate copy of all data on a different disk
drive. Allows for complete redundancy of data if a hard drive fails at the
cost of 50% of your usable hard disk space.
Stripe Set with Parity - stores the data on multiple hard disks and uses an
additional hard disk to store parity information. Provides performance
improvements and the ability to regenerate data on a member disk if it goes
bad.
What have people found works better for them based on experience?
Dave Goldsmith
dgol...@osec.com
Rowdy Ace
Dave,
You are correct about these two types of disk configuration, but I submit
to you that Stripe sets with parity are not that great of a performance
increase. Remember that pulling data off of a stripe set (or RAID 5) disk
includes a calculation, as it does writing to the disk. You get better
read performance from a Mirror set, because it only reads from one place.
I use Disk Mirroring for System drives. As for data drives, I use RAID 5
hardware, that takes the load off of NT and let's it do what it was
designed for, being an OS. If you have a small company and think you can't
afford RAID 5 hardware, then please remember when using Stripe Sets, if you
lose a drive, the Microsoft documentation says that you need to replace the
bad drive with one of Exactly the same make, model and size. This means
buy a couple of spares. Hope this helps.
Good Luck!
--
Todd M. Tolbert, MCSE
AKA Rowdy Ace
Dave Goldsmith <dgol...@osec.com> wrote in article
<67ubjn$mv0$1...@winter.news.erols.com>...
Prima
Rowdy Ace wrote:
All true if you want to try running a strips set with IDE. Have you
ever heard of SCSI? The latest SCSI bus you can support more then a
dozen hard drives, AND access them all at the same time! The
calculations involved in a RAID read is insignificant, it is only the
writes that can theoretically suffer. Doing two things at the same time
does not take longer. It saves time. Try it tomorrow morning brush you
teeth while your taking a wiz and you'll find you are two minutes early
for work!
This means that using no extra hardware a 4 disk stripe set with parity
will run almost four times as fast as a normal drive. NT mirroring does
not (last time I looked) support simultaneous reads from both drives,
although there are some versions of mirroring that do.
By using hardware, however, you take the load of the CPU, and my end up
with a better version of Mirroring. I believe NTs RAID5 is technically
quite advanced.
Personally I use strip sets without parity and then mirror the resultant
drive (yes its actually possible) This gives you the performance of
RAID3 (no calculation), and a fully redundant mirror. The result is a
very expensive fully redundant drive that can be accessed about as fast
as RAM on a 486.
Rowdy Ace
I guess this all depends on how much disk space you need. In my
environment, I have about 120 GB of RAID 5 disk space. Having NT take care
of the Stripe sets is ludicrous. The hardware raid takes the load off of
the CPU's (as you said), which to me is a significant performance gain. My
point was that if you're going to do RAID 5, before you just do it with NT,
you need to look at the Hardware type to save NT Cycles.
--
Todd M. Tolbert, MCSE
AKA Rowdy Ace
Prima <prim...@netvigator.com> wrote in article
<34A363A7...@netvigator.com>...
Tom Cooke
Personally, I'd trust 1+0 best; mirror your disks across two
controllers, then stripe over mirror pairs. Triple mirroring is even
better, as you can take one set offline to back up, then catch it up...
In article <67ubjn$mv0$1...@winter.news.erols.com>, Dave Goldsmith
<dgol...@osec.com> writes
>NT server supports four disk drive configurations two of which provide
>fault-tolerance.
>
>Disk Mirroring - keeps a duplicate copy of all data on a different disk
>drive. Allows for complete redundancy of data if a hard drive fails at the
>cost of 50% of your usable hard disk space.
>
>Stripe Set with Parity - stores the data on multiple hard disks and uses an
>additional hard disk to store parity information. Provides performance
>improvements and the ability to regenerate data on a member disk if it goes
>bad.
>
>What have people found works better for them based on experience?
>
>Dave Goldsmith
>dgol...@osec.com
>
>
--
Tom Cooke
Tom Cooke
<stuff>
Hang on, if you stripe _then_ mirror, and one drive goes down, you lose
your entire mirror and protection. If you mirror drive pairs, then
stripe, if a mirror goes down you are still mirrored on all other
pairs...
Tomasz J. Kaczmarski
Rowdy Ace (ttol...@flash.net) wrote:
: You are correct about these two types of disk configuration, but I submit
: to you that Stripe sets with parity are not that great of a performance
: increase. Remember that pulling data off of a stripe set (or RAID 5) disk
: includes a calculation, as it does writing to the disk. You get better
: read performance from a Mirror set, because it only reads from one place.
I'm not sure that a calculation is performed on reading. It
certainly is performed on writing and at least two hardware writes
are performed for every write request. Regardless of that fact, the
amount of time necessary to perform the calculation is negligible in
comparison to the amount of time necessary for a read. The
calculation performed is simple addition (in case of a write) or
subtraction (in case of a read). In the event when disks are several
orders of magnitude slower than the CPU it really doesn't make that
much difference.
: I use Disk Mirroring for System drives. As for data drives, I use RAID 5
: hardware, that takes the load off of NT and let's it do what it was
: designed for, being an OS.
I think this is a sound idea. I do this too.
Tom
Tomasz J. Kaczmarski
Prima (prim...@netvigator.com) wrote:
: All true if you want to try running a strips set with IDE. Have you
: ever heard of SCSI? The latest SCSI bus you can support more then a
: dozen hard drives, AND access them all at the same time!
Wow, this is kewl, I wonder how you got SCSI to work this way
because I don't see anything in the standards that says anything
about accessing anything at the same time. In fact they imply the
opposite.
: The
: calculations involved in a RAID read is insignificant, it is only the
: writes that can theoretically suffer. Doing two things at the same time
: does not take longer. It saves time.
If you're doing two things at the same time. In order to do that the
two disks must be on separate buses. However, it is NOT true in
general.
: This means that using no extra hardware a 4 disk stripe set with parity
: will run almost four times as fast as a normal drive.
Almost! The commands and data are sent sequentially to devices on
the same bus. They are also retrieved sequentially from devices on
the same bus. What may happen is that the SCSI device will receive
data from the bus at say 40MB/sec, buffer it while writing at
whatever speed it writes to disk, and while writingfree the bus for
other devices. You still won't squeeze out more than 40MB/sec from
the bus. In addition, if you send continuous requests so that the
buffers fill up (we'll talk about writes only, but similar is true of
reads) You'll be limited by the rate at which the devices write
from buffers to their physical media.
: NT mirroring does
: not (last time I looked) support simultaneous reads from both drives,
: although there are some versions of mirroring that do.
Of course not, there is only one kernel running in NT. At machine
language level machines are always sequential.
: By using hardware, however, you take the load of the CPU, and my end up
: with a better version of Mirroring.
You're right here. I agree.
: Personally I use strip sets without parity and then mirror the resultant
: drive (yes its actually possible) This gives you the performance of
: RAID3 (no calculation), and a fully redundant mirror. The result is a
: very expensive fully redundant drive that can be accessed about as fast
: as RAM on a 486.
Yeah, right. Why don't you pull all but several megs of RAM from
your machine (enough to boot NT), and use a large swap file. Why buy
RAM, disks are still much much cheaper and more versatile if you use
a dynamically sized page file.
BTW, in your scenario I pull any 2 drives from each set and your
system is down. It would make much more sense to mirror the drives
first and then stripe the resultants. This way you can withstand a
loss of up to half of the drives and in the worst case fail if two
mirrors (not just 2 random drives from each set) happen to die, an
unlikely scenario.
Tom
Roy Frederick
A read from a RAID 5 simply calculates what drive to read from and does
the read. A write has to read the parity sector and the sector to be written (if
it is not already in memory). The old contents of the sector to be written are
logically subtracted (exclusive or) from the parity sector. The new contents are
then xor'ed with the parity sector. Then the data sector and the parity sector
are both rewritten. All of this takes from 2 to 4 I/O operations, depending
on what might have been in buffers. A stripe set without parity improves
performance but makes you less fault tolerant. A stripe set with parity makes
you more fault tolerant but penalizes you fairly heavily on writes. Reads will
see almost the same improvement that the SSW/OP (like that?) does.
Roy