What's the problem with RAID 0?

[dgerol]

Most high-end motherboards these days come with a built-in RAID controller. In fact my ASUS P5WD2 Premium comes with two.

I'm considering setting up a RAID 0 array of 4 x 250GB hard disks for maximum performance, but I have read a lot of posts on this and other forums where people seem to not like RAID 0 at all...

The opponents of RAID 0 seem to come in two different flavours:

a) RAID 0 is not reliable because if one disk fails then you lose all your data
b) RAID 0 does not really improve performance at all

My thoughts are these...

a) Most computers have only one hard disk - thus having a potential failure rate of 100%. How is this any better or worse than a RAID 0 array? If one disk fails in both cases you're stuffed - so how is RAID 0 worse? With disks getting larger all the time, a backup solution is getting more and more difficult to implement. How are you going to back up 500GB of data? On DVD? I don't think so.... Thus most new computers with one hard disk will not be entirely backed up. OK... same with RAID 0.

b) I can't believe that 4 disks reading and writing in parallel cannot have much impact on perfurmance - just seems nutty to me to say so. Especially if you are doing large file transfers and video encoding etc...

Sure, RAID 1 + 0 is good, but what if you don't want to fork out the cash for two disks and not get the space benefits? Also RAID 5 is great, but what if you don't want to spend the extra processing for computation of parity information?

Looks to me as though RAID 0 is the best solution for performance.

Any thoughts?

[mathias]

The opponents of RAID 0 seem to come in two different flavours:

a) RAID 0 is not reliable because if one disk fails then you lose all your data
b) RAID 0 does not really improve performance at all

No, there aren't two "camps" of people who don't like raid, some(or most?) people here don't like it for both those reasons.

a) Most computers have only one hard disk - thus having a potential failure rate of 100%. How is this any better or worse than a RAID 0 array? If one disk fails in both cases you're stuffed - so how is RAID 0 worse?

You don't seem to understand probability very well. If the probability of a hard drive failing on it's own within a certian numer of years is x, and the probability of a big power fluctuation(that is either a power surge or something causing the power supply to break) killing the hard drives is y, then the probability of a single hard drive failing is x + y, and the probability of a striped raid system failing is x + x(1-x) +y.


And comparing raid to a system with one hard disk is completely unfair. A fair comparison is to a system with the two drives, with the OS(or the majority of the OS) and the swap file or partition on seperate drives.

[jackylman]

a) Most computers have only one hard disk - thus having a potential failure rate of 100%. How is this any better or worse than a RAID 0 array? If one disk fails in both cases you're stuffed - so how is RAID 0 worse?

A potential failure rate of 100%??!! That means the disks wouldn't start.

Let's assume after 1 year, your disks have a 99% chance of still working properly. (I'm ignoring bad sectors etc. and assuming a binary state of operation, working or not working). Then your RAID has a 98% chance of still being operational (0.99 * 0.99 = 0.9801) after 1 year as opposed to 99%.

Also, you could look at the increased heat from having two drives close together as a factor that might contribute to earlier failure. Not to mention the extra cost of the disk, extra electricity to run the disk and more electricity to cool your house becauese of the heat it generates. Oh yeah, HD's are noisy, even the better ones. All the reports I've seen show little benefit to RAID and NCQ on the desktop, unless you load windows all day long.

But it's your choice....

[teddysalad]

dgerol:

You make a good point saying that today's large size hard drives are a challenge to back up, however installing 4 drives in RAID 0 is suicide.
If you have one drive you risk failure, yes, but with 4 you just quadrupled the risk. If any one drive fails your data on ALL drives will be lost.

I think I understand where you are coming from cause I too desire to have 1 or 2 terabytes storage in my PC. But I stop and think really how much space do I need to do my work. Cause HDs do fail and everything not backed up is always at risk.

If you really need a terabyte or more on line at once and you desire faster throughput, you should consider RAID 5 w/parity. If you loose a drive you can restore everything relatively easily.

Also, RAID 0 does not guarantee extra speed. All the magazines were touting RAID 0 for games, until somebody there actually TESTED it. They found there was NO speed increase with the RAID 0 set-ups. The game software did not support it.

So you may not actually get those therotical speed increases you are counting on.

Good luck! Keep us posted on what you do and how you make out.

Cheers


P.S. Mathias: Thanks for the algebra lesson!

[dgerol]

Hmmm... perhaps you are right and RAID 5 is the safer way to go.

Does anyone know the performance difference between RAID 5 and RAID 0? I understand that you lose in effect one disk-worth of space to parity information. How does parity computation affect the CPU?

[chylld]

apparently raid 5 can be faster then raid 0. but hardware raid 5 controllers aren't that cheap

nevertheless, winxp has a built-in software raid 5 that offers about half the performance of hardware raid 5, that's the route i plan to take later on (it's free!!)

[swivelguy2]

If you're concerned with disk performance, do one of the following things:

A. have no paging file.
B. keep a static-sized paging file on a dedicated partition at the very beginning (the outermost/fastest part) of your hard disk.

The performance increase from a fast, unfragmented page file is usually much greater than any performance gain from a RAID setup.

The reasons RAID 0 is useful is because it allows you to have one logical disk even with (in your case) 4x250 GB drives, without wasting any space - but at the cost of reduced reliability.

In any RAID array, the early partitions of the array are the fast ones, and the late partitions are the slow ones. If you didn't do RAID at all, you'd have fast and slow places on each disk.

[PositiveSpin]

apparently raid 5 can be faster then raid 0. but hardware raid 5 controllers aren't that cheap

nevertheless, winxp has a built-in software raid 5 that offers about half the performance of hardware raid 5, that's the route i plan to take later on (it's free!!)

I think you'll find that the software RAID 5 is only (officially) available in Windows 2003 server - XP doesn't offer it.

[chylld]

I think you'll find that the software RAID 5 is only (officially) available in Windows 2003 server - XP doesn't offer it.

yeh that's why i linked to the thg article that shows you how to enable raid 5 on xp. only have to change 3 files

[dgerol]

My ASUS P5WD2 has RAID 5 as part of the ICH7R Southbridge, so I think I'll try it out and see how it goes.

[lenny]

a) Most computers have only one hard disk - thus having a potential failure rate of 100%. How is this any better or worse than a RAID 0 array?

The probability of a hard drive failing is X, where X is between 0 and 1.

The probability of one or more drives failing in a two drive array, where the probability of one drive failing is X, is :

X * (1 - X) + (1 - X) * X + X*X = 2X - X^2

Probability of drive 1 failing * Probability of drive 2 not failing +
Probability of drive 1 not failing * Probability of drive 2 failing +
Probability of both drives failing.

If X = 5%, probability of a failure in a 2 drive array is 9.75%. If X = 50%, probability of a failure in a 2 drive array is 75%.

With N drives in your array, you don't quite increase your risk by N. But for small values of X it is fairly close.

This little excursion into simple probability is brought to you by insomnia.

[wim]

You don't seem to understand probability very well. If the probability of a hard drive failing on it's own within a certian numer of years is x, and the probability of a big power fluctuation(that is either a power surge or something causing the power supply to break) killing the hard drives is y, then the probability of a single hard drive failing is x + y

it's always a laugh to see someone say "you don't understand X very well" and then go on to talk some nonsense about X.
anyway, you can't add probabilities like that, to see the mistake in your claim suppose x was 0.6 and y was 0.7. then x+y (which obviously must still be <= 1) isn't even a probability

edit: while i wrote my post, lenny has pointed out the same fact whilst resisting the urge to nag and correct - very admirable. the appropriate addition rule (with assumption of independent probabilities) is indicated in his post

[xarope]

I'm probably rehashing what most people have said already, but here goes:

- RAID 0 is high performance (whether minimally on a desktop, or maximally on a server) that almost any other RAID solution (search for writeups on the Areca RAID6 implementation which seems to give almost the same performance but better availability, of course this is implementation dependent)
- RAID 0 means any of the disks in the stripe set fail leads to total data loss, and associated high cost of data recovery, whereas a normal disk failure has multitude of free or cheap disk scanners which can attempt ro recover data for you. There is a company which sells "cheaper" products to allow recovery of raid5 data (the name escapes me right now), I'm not sure if this applies to raid0.
- Performance improvement is dependent on your I/O bottleneck characteristics, e.g. for people who use Photoshop with large RAW images, or video editing etc, may choose to use a RAID 0 partition for the scratch/temp disk on the grounds that losing that partition will not cause catastrophic loss of data.
- If you really have the money, and wish to gain x (x=unknown )percentile points of performance improvement on purely disk I/O, try h/w based RAID 1+0 (i.e. mirror, then stripe), which allows more combinations of drive failure than RAID 0+1 (and of course RAID0), with higher performance than almost any other RAID solution. I say almost, since H/W RAID performance is also dependent on implementation (e.g. what CPU/ASIC etc they are using for the XOR implenemtation, cache size/type etc).

I'll say this though, if time is not a mitigating factor (i.e. you can are willing to reload windows/re-ghost), and data is not important, then sure, go RAID0 to your heart's content. I know lots of gamers who do (and also listen kindly to their gripes when their 733t raptor raid0 that pwnz, dies).

Personally I have yet to go that step myself, having lost many hours recoverying other people's data and reloading/reinstalling OS of choice... no I'm not in that business, these are just friends/family I'm helping out. My business side recommends 1+0 for the multi-terabyte databases our application requires (largest one so far is probably 24+ terabyte), although for cost reasons most companies run RAID5 with the index tablespaces on RAID 1+0.

[xarope]

hmm, on reading my post above, some other points did not come across:

- raid0 implementations on motherboard are typically either not very well done (=> performance suffers, or data corruption occurs), or else will impact on CPU (in which case what you gain in I/O perf you lose in CPU perf). Hence, YMMV, I'd suggest research on your particular MBs raid chipset to find out if this will impact on you.
- to gain truly good performance, you'd most probably have to get a proper HW raid0 card, which might cost $$$, and again mitigates your time<->money tradeoff in going for a motherboard raid0 solution.
- It seems to me that the bulk of complaints on steam (the half life 2 internet distribution model) was not that it didn't work well, it was that when people started up half-life to play, they ended up having to wait for updates (and hence cracks/hacks abound where they could turn it off or make it pretend it was in offline mode). You should consider that as well, is it worth x% disk i/o perf if your PC is down additional y% of the time due to the raid0.

anyway, as mentioned, go ahead, try it out and if performance doesn't improve the way you think it will (remember performance is never purely about disk i/o, you need to ensure overall improvement, not just improving disk i/o at the cost of cpu load etc) you can treat it as a lesson learnt.

[merovingian]

This is depressing, first you're posting in the wrong place to get answers about storage performance. Read this first...
http://www.anandtech.com/storage/showdoc.aspx?i=2101
Then post on their forums. What is it you want to do with this set up? Serve data, load apps faster? App loading is dependent on access time mostly, if you want to go fast then get a raptor (4.5 or 4.7ms access times if I remember correctly), if you have the money to blow get a 15K rpm scsi drive with a scsi card, they hit 3.3ms access times. Good luck.

[lm]

It seems most of the posts about probability are either wrong or just complicated.

If the probability of a single drive to fail during some defined period of time is P then the probability of one or more drives out of N drives failing in that same period of time is:

Code: Select all

1 - (1-P)^N

Example: If the probability of one drive to fail in one year is 1%, then the probability for it to not fail is 99% and so the probability for 4 drives not failing is 99%^4 = 96% and so the probability of one or more of those 4 drives failing during one year is 4%

Now suppose we want to use that 4 disk raid system for 3 years. The probability for it not to fail is (99%^4)^3 = 88.6% so the probability of one or more drives out of 4 during 3 years would be 11.4%.

Of course what we don't know is the probability of one drive to fail during one year.

And shouldn't this be about silencing stuff?

[chylld]

And shouldn't this be about silencing stuff?

yeh that's something i've been thinking about and is one of the drawbacks of my plan to go raid 5 - seeks could potentially be much noisier because all of the hdd's will be active at the same time.

[Cerb]

Hmmm... perhaps you are right and RAID 5 is the safer way to go.

Does anyone know the performance difference between RAID 5 and RAID 0? I understand that you lose in effect one disk-worth of space to parity information. How does parity computation affect the CPU?

Software RAID 5 will be slow. Can't say for windows, but a 360GB (4x120GB) software array in Fedora Core (ClarkConnect 3.1, actually) can get about 8-10MB/s writes (reads are good enough to not notice any problems) with a AthlonXP 1600+. RAID 0 is significantly faster until you get a hardware controller with cache.

Cheap controllers that do 0,1,0+1, are pretty much doing it in software.

If that Tom's link is right, the software RAID gets similar performance in Windows, and it's more like 1/3, not 1/2, of decent hardware (of course, God forbid they do a real test, either, like zipping files or running the basic office benchmarks).

Fine for a simple file server, not good for much else.

[wim]

OT. (read if you like maths)

It seems most of the posts about probability are either wrong or just complicated.

most..? excuse me, but i can only see one reply which is wrong, and none which are complicated. your explanation - whilst correct - is the same as jackylman said originally (except with n=4 instead of 2). and you are saying the same thing as lenny, except he's summed the probabilities directly where you've taken certainty (1) minus the complement (1-P)^n. but same thing, logically, expand see for yourself: 1-(1-P)^2 = 2P-P^2
further reading binomial theorem

[chylld]

all of these numbers flying around enticed me to think about the problem myself so here's my crack at it.

instead of putting a single number to the probability of failure for a hard drive, i thought about it as a curve over time, i.e. probability of failure over time, where p(fail) = 0.5 @ time = MTBF. by using simple probability math one can combine curves and get a combined probability of 1 or more failures over time. so p(>=1 fail) = 1-((1-p(fail))^(number of hard disks)), which i think is relevant because a single failure in a raid0 array is of concern.

so, on the assumption that the p(fail) over time curve is linear, i made this graph:



series 1 is for 1 hdd, series 2 is for 2 hdd's, etc.

now if we assume that a failure will occur when p(>=1 fail) > 0.5, we can draw an imaginary line at p = 0.5 and it's easy to see that the more drives you have, the sooner along the timeline you have to go to get an intersection.

now of course i have no idea what the curve for p(fail) over time actually looks like, but if it's anything remotely linear then the MTBF of a raid 0 array is frighteningly shorter than that of a single drive!

but obviously most of us know that already just wanted to draw a graph to help myself, if not others, get around the numbers.

[wim]

cool chylld
but no way should series1 be linear...i reckon it would probably be shaped more like your series5, divided by 2

[mathias]

it's always a laugh to see someone say "you don't understand X very well" and then go on to talk some nonsense about X.
anyway, you can't add probabilities like that, to see the mistake in your claim suppose x was 0.6 and y was 0.7. then x+y (which obviously must still be <= 1) isn't even a probability

edit: while i wrote my post, lenny has pointed out the same fact whilst resisting the urge to nag and correct - very admirable.

Okay, so I forgot to multiply the probabilities of a hard drive(s) failing on its own by the probability of it not getting electrocuted(1-y). (or vice versa)

My equations are still more accurate, the probability of electrocution(y) is very small(probably significantly smaller than x), so lumping it in with x creates much bigger inaccuracies. That is:

x + x(1-x) +y

is much closer to

(x + x(1-x) )(1-y) +y

than

z + z(1-z)
or
1-( (1-z)*(1-z) )
where
z = x(y-1) +y = x + y(1-x) = 1-( (1-x)*(1-y) )


which is equivalent to
(x + x(1-x) )(1-y) + y + y(1-y)


(z represents the combined probability of the drive failing on its own or through electrocution)

[merovingian]

Hmmm... perhaps you are right and RAID 5 is the safer way to go.

Does anyone know the performance difference between RAID 5 and RAID 0? I understand that you lose in effect one disk-worth of space to parity information. How does parity computation affect the CPU?

Software RAID 5 will be slow. Can't say for windows, but a 360GB (4x120GB) software array in Fedora Core (ClarkConnect 3.1, actually) can get about 8-10MB/s writes (reads are good enough to not notice any problems) with a AthlonXP 1600+. RAID 0 is significantly faster until you get a hardware controller with cache.

Cheap controllers that do 0,1,0+1, are pretty much doing it in software.

If that Tom's link is right, the software RAID gets similar performance in Windows, and it's more like 1/3, not 1/2, of decent hardware (of course, God forbid they do a real test, either, like zipping files or running the basic office benchmarks).

Fine for a simple file server, not good for much else.

I understand the opposite, I hear that software raid is faster but requires cpu usage. I assume anyone using raid 5 is also using a app drive, if so than there is no need to have hardware raid 5 for just 4 drives as the raid 5 is for media and will rarely be taxing the processor.

[xarope]

sw raid5 on linux can be faster than a cheap raid5 card without a dedicate xor engine and cache. If you have enough spare parts to build a linux box, cost is not a factor either, just the time spent in setting up linux and understanding about mdadm.

high end raid cards can garner better performance, but then they are USD$1000 or more, and you are "locked" in to that card...

[Cerb]

Hmmm... perhaps you are right and RAID 5 is the safer way to go.

Does anyone know the performance difference between RAID 5 and RAID 0? I understand that you lose in effect one disk-worth of space to parity information. How does parity computation affect the CPU?

Software RAID 5 will be slow. Can't say for windows, but a 360GB (4x120GB) software array in Fedora Core (ClarkConnect 3.1, actually) can get about 8-10MB/s writes (reads are good enough to not notice any problems) with a AthlonXP 1600+. RAID 0 is significantly faster until you get a hardware controller with cache.

Cheap controllers that do 0,1,0+1, are pretty much doing it in software.

If that Tom's link is right, the software RAID gets similar performance in Windows, and it's more like 1/3, not 1/2, of decent hardware (of course, God forbid they do a real test, either, like zipping files or running the basic office benchmarks).

Fine for a simple file server, not good for much else.

I understand the opposite, I hear that software raid is faster but requires cpu usage. I assume anyone using raid 5 is also using a app drive, if so than there is no need to have hardware raid 5 for just 4 drives as the raid 5 is for media and will rarely be taxing the processor.

I guess it could depend on the hardware. For one thing, I know a Duron 800 can't handle software RAID 5 very well.

It's hard to imagine a cacheless controller being much faster (it probably could be with a dedicated processor do do the parity--but even cacheless, that wouldn't be cheap). And yes, it is separate from the boot drive. I don't know if it's Red Hat or Linux in general, but the RADI5 array can't be bootable.

The main benefit of software RAID 5 is that real hardware RAID cards start around $200, and that's a PATA one that Promise is likely trying to get rid of. At Newegg, at least, you have to get close to $300 for a real SATA RAID card (3Ware and Adaptec both have them for about $300, with actual processors on the card and some memory). If performance matters, it's a good way to go. If you just need storage, and don't need perfect hot-swapping (I don't know how well hot-swapping and rebuilding an array works in software RAID), software will do the trick--you can buy another drive and some cool gadgets for that $300!

[Straker]

noone has stated the obvious yet - unless you're only using onboard IDE/SATA ports, any solution will also very quickly be bottlenecked by PCI. hardware vs software on speed alone is pretty much a non-issue unless you're only using a few drives or are using PCI-X controllers. PCIe works too but good luck finding anything but single and 16 lane slots on a consumer board.

remember the outer part of any modern, large HD can deliver >60MB/sec STR now.

[Cerb]

noone has stated the obvious yet - unless you're only using onboard IDE/SATA ports, any solution will also very quickly be bottlenecked by PCI. hardware vs software on speed alone is pretty much a non-issue unless you're only using a few drives or are using PCI-X controllers. PCIe works too but good luck finding anything but single and 16 lane slots on a consumer board.

remember the outer part of any modern, large HD can deliver >60MB/sec STR now.

If you are doing software RAID, why wouldn't you use the onboard SATAs? You can get 4x SATA mobos for at and under $80, now. It's really not obvious, and in a very, very, good way .

In fact, the OP's system (that mobo appears to have ICH7R) is new enough, it appears, even for half-way decent RAID 5:
http://www.gamepc.com/labs/view_content ... e%5Ftest=1
Burst rates are almost silly, really. The NF4 is somewhat dissappointing, though--that nice ICH7R doesn't come cheap.

[widman]

RAID 0 failed me before. There is no way I will install RAID 0 again. It's completely waste of time and money. With little overall performance gain, you sacrifice bunch of data or OS config.

The problem in RAID 0 was not HDD failure as theorical. But more on logical binding. It's very easy to break RAID 0 binding, and once it's break no way you can recover data/system. There are many situation that can break RAID 0/1 config as general. But breaking on RAID 1 is non-destructive.

Here a list of my experience of breaking RAID 0 or 1 in Adaptec HighPoint RAID, Promise and built in RAID motherboard.
- one of hdd spoilt/bad sector (Promise)
- lag one of the hdd make system hang, restart and corrupted (Promise)
- bad controller (adaptec HighPoint)
- rearrange hdd and forgot which cable connect to which hdd. (Promise)
- forgot to plug power of one of hdd after changing PSU. (Promise)
- flashing mobo BIOS. It reset the config and break the binding. (built in)
- Overclock CPU and need to reset motherboard, it reset the RAID config as well (built in)
- no symptom , it break by itself after shutdown, buggy raid controller. (built-in)

[vertigo]

Why did you try it in the first place? Some people see it coming and yet they do nothing...

[mattthemuppet]

i think the probality of one person being wrong about probabilities =

(n + q + z)/r,

where n = the IQ of the proposer, q = the age they dropped out of school, z = the no. of humanities courses they studied and r = the no. of people talking about probabilities.

I think that's pretty clear