RAID 0 is usually a foolish choice for desktops

"Leythos" wrote in message
...
In article ,
says...
In fact usually a stripe size of
16K or 32k for two drives is usually optimal. Also Anand doesn't seem
to
understand the definition of what a stripe is.

Stripe size is optimal based on the type of data being accessed, which,
as most people tend to overgeneralise, is different based on what the
primary function is for the system.

As an example:

1) Database - larger stripes are better

Wrong. Stripe units 2x-3x the average record size is optimal.

2) Large images or other LARGE files - larger is better

Wrong, The stripe size that optimizes sustained transfer rate is best here
and often that's NOT large stripe size.

3) Most home (non-soho/non business) systems - smaller
4) Video editing - larger

No.

5) Web Page Design - smaller
6) Games - smaller

And the list goes on.

And you make these up how?

"Leythos" wrote in message
...
In article ,
says...

"Leythos" wrote in message
...
In article ,
says...
RAID-1 offers fast reads, but slower writes, but RAID-0 is just
not
worth the loss of all data on BOTH drives.

Just do backups and u are fine.

Most people don't even purchase something for backup, most don't even
know what a tape drive is, most people don't even know what Windows
Update is, most people don't even run the update for the AV program.

If you build an 120+GB RAID-0 Array, do you know that it would cost
more
than the price of a new computer to back it up to tape?

Tape, god no. Use a SATA drive in a removeable tray ~$220 for 250GB.

Removable drives are great options for off-line backup. I just installed
a 1.2TB array/server to act as a backup storage solution. Everything in
the office is backed up to it, about 2 weeks worth of backups, and then
the tape drive is used once a week to backup the latest Friday (takes
more than 1 tape.

Forget the tape and use a removeable HD.

Milleron wrote:

If more proof of this old contention is needed, there's a cutting-edge
review by Anand Shimpi on AnandTech.com:
http://www.anandtech.com/storage/showdoc.html?i=2101

As always in these tests, the real-world improvement achieved with
RAID 0 varies between 0 and 4%, which is simply imperceptible. The
price that's paid is two-fold: (1) the difference between the RAID
and a single drive of the same capacity, and (2) the DOUBLING of the
chance of a hard-drive failure.

For the life of me, I can't understand why so many users decide to
install RAID 0 on desktops.

For the life of me, I cannot understand how RAID0 can claim to be RAID
at all.

RAID = Redundant Array of Inexpensive Disks.

RAID0 has no Redundancy, therefore it's not RAID.

RAID 1 is another matter entirely, but, as Anand says, "If you haven't
gotten the hint by now, we'll spell it out for you: there is no place,
and no need for a RAID-0 array on a desktop computer. The real world
performance increases are negligible at best and the reduction in
reliability, thanks to a halving of the mean time between failure,
makes RAID-0 far from worth it on the desktop."


Ron

SCSI can be cheaper.

An ultra 320 raid controller costs big time. A top end 8 drive SATA raid
controller will cost about the same per disc.

73 GB SCSI drives are (here) cheaper than the raptors and have comparable
stats (the raptors come out very well in many benchmarks). So with 14 drives
on one RAID controller (you never should as this exceeds the IO performance
of the controller by a long way) you have one nice clear solution especially
if it is dual channel. I have been playing with an Intel controller recently
(ex LSI or something) and it has all the bells and whistles... yum.

Needless to say, the raptors will come down in price a little faster than
the SCSI, and a new drive design or two is no doubt already in the wings. So
things are improving! Gone are the days of 5, 10, 20 MBs and other slow
SCSI, the old yich clunkity clunk IDE drives. Thank goodness. Now we have
IDE drives with 1 year warrantee. That sends a shiver down my spine - if the
drive doesn't have a 3 year warrantee then its no good.

- Tim




"Winey" wrote in message
...
On Fri, 2 Jul 2004 22:46:34 +1200, "Tim" wrote:

Agree 100%. It's a good article. SCSI is still smoother esp with a dual
CPU
system. Perhaps the NCQ drives will help iron things out, then dual core
opterons will bring a smile to everones faces.

What are NCQ drives?

Glad you still like SCSI. If you look at the pricing for some of
the high-perf 73 GB ATA drives, you're going to pay about what the
same drives cost in SCSI-land.

--W--


- Tim

"Milleron" wrote in message
. ..
If more proof of this old contention is needed, there's a cutting-edge
review by Anand Shimpi on AnandTech.com:
http://www.anandtech.com/storage/showdoc.html?i=2101

As always in these tests, the real-world improvement achieved with
RAID 0 varies between 0 and 4%, which is simply imperceptible. The
price that's paid is two-fold: (1) the difference between the RAID
and a single drive of the same capacity, and (2) the DOUBLING of the
chance of a hard-drive failure.

For the life of me, I can't understand why so many users decide to
install RAID 0 on desktops.

RAID 1 is another matter entirely, but, as Anand says, "If you haven't
gotten the hint by now, we'll spell it out for you: there is no place,
and no need for a RAID-0 array on a desktop computer. The real world
performance increases are negligible at best and the reduction in
reliability, thanks to a halving of the mean time between failure,
makes RAID-0 far from worth it on the desktop."


Ron

"Tim" wrote in message ...
SCSI can be cheaper.

An ultra 320 raid controller costs big time. A top end 8 drive SATA raid
controller will cost about the same per disc.

73 GB SCSI drives are (here) cheaper than the raptors and have comparable
stats (the raptors come out very well in many benchmarks).

No, comparable SCSI HDs are NOT cheaper.

So with 14 drives
on one RAID controller (you never should as this exceeds the IO
performance
of the controller by a long way) you have one nice clear solution
especially
if it is dual channel. I have been playing with an Intel controller
recently
(ex LSI or something) and it has all the bells and whistles... yum.

Needless to say, the raptors will come down in price a little faster than
the SCSI, and a new drive design or two is no doubt already in the wings.
So
things are improving! Gone are the days of 5, 10, 20 MBs and other slow
SCSI, the old yich clunkity clunk IDE drives. Thank goodness. Now we have
IDE drives with 1 year warrantee. That sends a shiver down my spine - if
the
drive doesn't have a 3 year warrantee then its no good.

Nonsense. The warranty length is simply a price point decision and say
nothing about reliability. Many ATA HDs have 3 year warranties.

"Leythos" wrote in message
...
In article ,
says...

"Leythos" wrote in message
...
In article ,
says...
In fact usually a stripe size of
16K or 32k for two drives is usually optimal. Also Anand doesn't
seem
to
understand the definition of what a stripe is.

Stripe size is optimal based on the type of data being accessed,
which,
as most people tend to overgeneralise, is different based on what the
primary function is for the system.

As an example:

1) Database - larger stripes are better

Wrong. Stripe units 2x-3x the average record size is optimal.

2) Large images or other LARGE files - larger is better

Wrong, The stripe size that optimizes sustained transfer rate is best
here
and often that's NOT large stripe size.

3) Most home (non-soho/non business) systems - smaller
4) Video editing - larger

No.

5) Web Page Design - smaller
6) Games - smaller

And the list goes on.

And you make these up how?

I'm not sure how to reply to you without ticking you off, but you are
wrong, larger average file sizes means larger stripe sizes for better
performance, which is exactly what I wrote.

Utter nonsense. There is NO first order relation between file size and
stripe size.

"Leythos" wrote in message
...
In article ,
says...
I'm not sure how to reply to you without ticking you off, but you are
wrong, larger average file sizes means larger stripe sizes for better
performance, which is exactly what I wrote.

Utter nonsense. There is NO first order relation between file size and
stripe size.

You are correct, there is no relation between file size and stripe size,
but there is a relation between file size, stripe size and performance.

Envision stripe size like you would cluster size and then you'll
understand.


You are simply WRONG. The concept of cluster size and stripe size have no
relation to one another. Cluster is a logical software concept of disk
storage allocation. Stripe size has to do with physical allocation across
drives.

Envision the difference between stripe size and stripe unit size. Stripe
size has nothing to do with performance but stripe unit size often does.
The optimal stripe unit size for streaming/large-file performance is one
that quickly starts and then maintains the continuous stream of data on all
the drives in a RAID 0 set. The goal is not to lose any revolutions. If
the stripe unit size is too large then the initial OS read request may not
be large enough to get all the HDs in the RAID 0 set starting a read. If
the stripe unit size is too large then the drive's read ahead may not be
sufficient to continue the stream until the next read request arrives. Fail
these criteria and you lose revolutions and therfore non-optimal streaming
performance. Stick to an area where you have some actual knowledge.

Large stripes are non-optimal for streaming. There's a middle ground that's
usually optimal and that depends on controller HW design, RAID drivers
design, OS design and the HD's internal caching and other behaviors and
settings.

"Leythos" wrote in message
...
In article ,
says...
Cluster is a logical software concept of disk
storage allocation. Stripe size has to do with physical allocation
across
drives.

Yep, I agree and we're not going to agree on it. It could be that we're
thinking the same thing and just not able to put it into type, but I'll
bow out of this thread with what I believe.

There are two things that I know about Stripes:

1) Number of drives in the stripe set has an impact on performance.

Yep. more means faster streaming.

2) Size of the stripe on each drive also impacts performance.

That's called the 'stripe unit' size and does affect performance. Not too
big and not too small is generally optimal for streaming.

Ron,

Shop around - the prices are near identical (raptor vs. 10kRPM SCSI).
Obviously don't go to a retail reseller that sells everything including
vacuum cleaners, go to a specialist reseller. Neither raptors nor SCSI are
cheap.

For IDE, warrantee may be a "price Point" for many drives - it is not for
example a price point for Raptors. You can always get sucked into buying
extended warrantees for products. The situation I deplore most is when a
retailer sells a product with a manufacturers warrantee of 5 years with only
a 1 year warrantee and will then sell an extended warrantee. Needless to say
you should not buy off such people.

Shop around. There is enormous diversity in quality of retailers / system
builders / resellers in terms of their own knowledge (none to excellent) of
technical aspects of products, their ability to support these products, and
the associated need to return product is less with the more capable
resellers as they tend to weed out the crap and not put it on the shelf.
There are certain manufacturers products which as soon as I see on the shelf
signal to me that the shop concerned knows little about computers and
quality.

- Tim






"Ron Reaugh" wrote in message
...

"Tim" wrote in message
...
SCSI can be cheaper.

An ultra 320 raid controller costs big time. A top end 8 drive SATA raid
controller will cost about the same per disc.

73 GB SCSI drives are (here) cheaper than the raptors and have comparable
stats (the raptors come out very well in many benchmarks).

No, comparable SCSI HDs are NOT cheaper.

So with 14 drives
on one RAID controller (you never should as this exceeds the IO
performance
of the controller by a long way) you have one nice clear solution
especially
if it is dual channel. I have been playing with an Intel controller
recently
(ex LSI or something) and it has all the bells and whistles... yum.

Needless to say, the raptors will come down in price a little faster than
the SCSI, and a new drive design or two is no doubt already in the wings.
So
things are improving! Gone are the days of 5, 10, 20 MBs and other slow
SCSI, the old yich clunkity clunk IDE drives. Thank goodness. Now we have
IDE drives with 1 year warrantee. That sends a shiver down my spine - if
the
drive doesn't have a 3 year warrantee then its no good.

Nonsense. The warranty length is simply a price point decision and say
nothing about reliability. Many ATA HDs have 3 year warranties.

"Leythos" wrote in message
...
In article ,
says...

"Leythos" wrote in message
...
In article ,
says...
Cluster is a logical software concept of disk
storage allocation. Stripe size has to do with physical allocation
across
drives.

Yep, I agree and we're not going to agree on it. It could be that
we're
thinking the same thing and just not able to put it into type, but
I'll
bow out of this thread with what I believe.

There are two things that I know about Stripes:

1) Number of drives in the stripe set has an impact on performance.

Yep. more means faster streaming.

2) Size of the stripe on each drive also impacts performance.

That's called the 'stripe unit' size and does affect performance. Not
too
big and not too small is generally optimal for streaming.

See, we were on the same page, I just wasn't using the proper terms for
it. I also was not talking about streaming and more thinking of
performance for specific types of files.

Big files and streaming is the same thing. A big file for this purpose is
one whose size is greater than the stripe size.

Multithreaded/multitasked small record random I/O is optimal performance
wise when the stripe unit is about 3x the average record size.