2x80Gb RAID0 or 1x320Gb?

[this_is_gav]

Don't see why the failure of a pair of drives would put you off perpendicular recording in general.

What exactly did your GD raptors beat the 7200.10s in? I have trouble beleiving they would be much faster in any real world task. I had a pair of 36.7GB raptor GDs in a raid 0 and they don't match my single revised 7200.10 (the revised models with the larger plater size are noticeably faster than the older 7200.10s).

VirtualDub, encoding a bunch of uncompressed video files from my school's panto. I was just using spare drives I had lying around as my main ones were full.

Don't see why the failure of a pair of drives would put you off perpendicular recording in general.

OK, fair enough, it put me off Seagate's perpendicular recording drives. A 100% failure rate. Sweet.

[charleski]

You want to argue the time to first failure for one disk in a group of n is significantly sooner than than the time to failure for a single disk.

It is conceivable this could be the case /&/ disk failures could cluster, but the most plausible explanation for that would be if disk failures aren't independent.

Absolutely right, and that's the whole point of the paper aidanjt cited, though I don't think it has a bearing on RAID0 systems. Disk failure depends on 1)manufacturing issues, which will affect all disks in the same batch and 2)environmental issues, which will affect all disks housed in the same conditions. If you create a RAID array from a set of disks which are all bought at the same time (and thus probably from the same batch) and are housed in the same enclosure, then the failure of one disk very definitely does indicate a much greater risk of the rest of the disks in the same group failing over the next 4 months.

Here's a simple example to illustrate the idea (I'm not claiming the details apply directly to the disk case). Imagine a bomb designed to detonate randomly according to a Poisson process

Yep, that's the conventional wisdom. Aidanjt was quite right to point out that it has been shown that clusters of disks do not exhibit statistical independence and do not follow an exponential distribution in terms of failure. (Take a look at the autocorrelation function described in Fig 7, a Poisson process should have an autocorrelation of 0.)

But this only applies to batches of disks, and thus doesn't affect a situation in which the failure of one disk is sufficient to bring the system to a halt. A RAID0 system only needs one disk to fail.

I think the problem in this thread is partly that none of the posters have a great deal of objective data on disk failure patterns

Actually, I think the paper that aidanjt cited is very good, it covers experimental data from thousands of disks used in real-world conditions in HPC setups. I think their data conclusively rejects the null hypothesis of drive failure independence, and I think the rationale for that is completely sensible.

But I don't think this data affects any system in which the failure of a single drive is catastrophic, and that includes RAID0 systems. That is the issue in this thread .

I never really liked statistics & I don't want to think about this any more.

Don't be scared of stats! Statistics is a very powerful tool, as long as you make absolutely sure you know what the numbers mean. Always go back to the basics and check that it all makes sense, and then check again. When I did my PhD there was another candidate in my lab who claimed his results had a significance of p<0.0000001 - I felt sorry for the professor, though he did finally manage to clean up his analysis into something decent.

[mroz]

Actually, I think the paper that aidanjt cited is very good, it covers experimental data from thousands of disks used in real-world conditions in HPC setups. I think their data conclusively rejects the null hypothesis of drive failure independence, and I think the rationale for that is completely sensible.

But I don't think this data affects any system in which the failure of a single drive is catastrophic, and that includes RAID0 systems. That is the issue in this thread .

If you accept drive falures aren't independent I don't understand what your argument is to support the case that the expected first failure in a RAID0 system is an n'th of the time of expected faiure for a single drive.

Don't be scared of stats! Statistics is a very powerful tool, as long as you make absolutely sure you know what the numbers mean. Always go back to the basics and check that it all makes sense, and then check again. When I did my PhD there was another candidate in my lab who claimed his results had a significance of p<0.0000001 - I felt sorry for the professor, though he did finally manage to clean up his analysis into something decent.

Yeah, I can imagine. I did my degree in maths, but was more into pure than applicable. I'm not afraid of stats - I see them more as a jake blues than a jason voorhees

[charleski]

If you accept drive falures aren't independent I don't understand what your argument is to support the case that the expected first failure in a RAID0 system is an n'th of the time of expected faiure for a single drive.

The time-to-failure of a RAID0 system is the time-to-failure of a single drive.

Sorry, but I can't see what the problem is here. Correlations between drive failures are irrelevant in this case because we're talking about single drive failures. The expectation function of an arbitrary single drive failure increases linearly with the number of drives subject to failure. That really is as basic as tossing coins and building a truth table.

I've tried to lay it out pretty clearly, and there is a big difference between the failure of a random single drive and the failure of multiple drives in the same group.The statistics of individual, arbitrary events is very different to that of grouped events. A failure of A or B is twice as likely as a failure of A alone - there are 2 things that can go wrong, so there are 2 times the chances of failure. A failure of A and B is an entirely different matter, and is covered by the paper cited, but does not have any relevance to RAID0 systems.

As I said, always go back to the basics, always make absolutely sure you know what the numbers mean. I think some confusion has crept in here over the logical difference between or and and - these are very different combinations and have very different outcomes.

[mroz]

The time-to-failure of a RAID0 system is the time-to-failure of a single drive.

Obviously.

Sorry, but I can't see what the problem is here. Correlations between drive failures are irrelevant in this case because we're talking about single drive failures. The expectation function of an arbitrary single drive failure increases linearly with the number of drives subject to failure.

No, it doesn't. It increases approximately linearly for small probabilities, but that is only trivially applicable if failures are independent & you agreed they aren't - at least I thought you had.

I think we'll just have to each agree to assume the other is wrong.

A failure of A or B is twice as likely as a failure of A alone - there are 2 things that can go wrong, so there are 2 times the chances of failure.

Aside from being somewhat imprecisely stated, that's obviously wrong - what if the chance of failure of each of A and B is 0.75? You're claiming A or B fails with probability 1.5 ?!

Never mind - I've got to go check on Deep Twit now.

[charleski]

The expectation function of an arbitrary single drive failure increases linearly with the number of drives subject to failure.

No, it doesn't. It increases approximately linearly for small probabilities, but that is only trivially applicable if failures are independent

No, you're confusing the issue again. Yet again you're confusing multiple failures with single failures. They are not the same thing, and this is what you're missing. The question of independence is irrelevant when we are only interested in the chance of a single event occurring.

Aside from being somewhat imprecisely stated, that's obviously wrong - what if the chance of failure of each of A and B is 0.75? You're claiming A or B fails with probability 1.5 ?!

MTTF (group)≈MTTF(disk)/(no.of disks)
Here's a simpler example which may help: If the chance of a single drive failure is 50%, then the chance of the system not failing is 50% with one drive, but only 25% with 2. If you have a large number of systems, the number of systems that fail with 2 disks will, on average, be twice that of the systems with 1 disk. If you find the numbers confusing, just draw up a truth table: (F=fail, P=pass)
System with 1 drive (each line=50% chance):
F
P
System with 2 drives (each line=25% chance):
F F
F P
P F
P P

[mroz]

If anyone else wishes to comment I'd be interested to see the resulting discussion. Failing that I've nothing more to add as I'd just be repeating myself.

[aidanjt]

Not really, I gave up repeating myself at the top of this page.

[kungpo]

Ha ha - interesting thread.

I'll stick my tuppence worth in now.

Disk failure is NOT a reason to choose RAID0. A disk can fail at anytime, whether you have a single disk or a pair. So....not protection from disk failure.

OK - so in a RAID0 array, the second disk has as much chance as failing as the first. So what? if one disk goes the whole lot is trashed anyway. So, if one or both disks fail, it makes little overall odds in the scheme of things.

So, back to the original question - is it worth it?

It's up to you? RAID0 IS faster than single disk assuming similar disks, as you have doubled your disk I/O.

Finally, two smaller disks can be cheaper than 1 larger disks. Especially true at larger capacities, where 2x 500GB is cheaper than 1TB.

And last words, no RAID configuration is a substitute for backups. Whether the disk(s) fail in the first day or after 3 years is no measure of protection.

[charleski]

Disk failure is NOT a reason to choose RAID0. A disk can fail at anytime, whether you have a single disk or a pair. So....not protection from disk failure.

OK - so in a RAID0 array, the second disk has as much chance as failing as the first. So what? if one disk goes the whole lot is trashed anyway. So, if one or both disks fail, it makes little overall odds in the scheme of things.

Take a look at the table I posted above. If only one disk needs to fail, then yes, it makes a difference. I really advise you to read over it.

I can understand why people find stats confusing, and I've seen people (who should know better!) screw up their stats a lot. Just go back to the basics every time and remember to stay real, because that is what stats is all about.

Disk failure IS a problem with RAID0, kunpo, in fact it's a problem with any form of RAID, but the degree of the problem differs.

[aidanjt]

No, the problem with stats is, it's useless in the real world and proper application of logic. And even more worthless where mechanics are involved.

[charleski]

No, the problem with stats is, it's useless in the real world and proper application of logic. And even more worthless where mechanics are involved.

Obviously you're not an engineer of any sort and had to apply theory to reality.

Sorry, but that's a pretty silly statement.

[mroz]

Stats are only useless, in fact worse than useless as with any science, when trusted without being understood. Especially when being confidentally promoted as fact when in actuality many mistakes are incorporated. I hope no one is relying on this thread for a source of facts. It should contain a big warning telling the unsure to steer clear.

[Clunk]

Three - last word freaks, in one thread, now that's quality

[mroz]

I don't know what you mean

[aidanjt]

Obviously you're not an engineer of any sort and had to apply theory to reality.

I'm the sort who doesn't over-generalise.

Sorry, but that's a pretty silly statement.

In your lone, solitary opinion.

Enlighten yourself:
http://en.wikipedia.org/wiki/Misuse_of_statistics

Three - last word freaks, in one thread, now that's quality

I'm just wondering when it'll reach the point of "Am not!", "Are too!", "Are three!".