2x80Gb RAID0 or 1x320Gb?

[mroz]

So, are we agreed that increased chance of disk failure is not a big determining factor in choosing whether to go RAID0?

I couldn't confidently go that far without getting my own hands on the data & making a lot more effort to revise my stats knowledge, but I'd certainly argue that 'failure rate is proportional to number of drives' overestimates the risks, probably significantly. Equally, the risk of failure of a group will still be higher than of a single drive, but the additional risk may be acceptable given the performance gains & the reduced size of this additional risk. This is all just my opinion based on gut feelings & a moderate grasp of the report previously quoted.

[mroz]

A more interesting question then is whether, if you have two 500Gb drives to play with, you are better going with RAID or a split usage policy - purely from the PoV of performance.

Well, as far as performance goes, RAID0 will perform much better in sustained read/write operations, for random I/O the performance improvement will be nominal, as that is dependant on access/seak time (the time it takes for the read/write head to get into position and the appropriate sector comes into alignment and start I/O), rather than raw throughput.

There is a negative to offset, if you go with the RAID0 solution. Without RAID, copies from one drive to the other will be faster. As soon as you RAID the drives you lose this in that any copy from the array back to the array will result in increased seeking. I'd be interested to know how the following three cases perform:

Copy from RAID0 to physically separate RAID0
Copy from one drive to another
Copy from RAID0 back to same RAID0

[TooNice]

A common misconception.

If sequential transfer rate is important to the user, then it is no lie, nor misconception that some 7200 drives can match the Raptor. The problem is when people take sequential transfer rate as an absolute measure of performance. I do not doubt that there are people who legitimately find good use for RAID-0 in applications that requires sustainable transfer rate above 100MB/sec, but I do suspect there are a fair few who look at the HD Tach/Sandra graphs and suddenly 'feel' that their games loading time is being halved.

Well, it isn't like I've not given RAID-0 a shot. I went from SCSI -> 7200 IDE -> 2G Raptor -> 2G Raptor in RAID-0 -> 3G Raptor. There are a few things that the two 2G Raptor in Raid-0 does better than the 3rd Gen Raptor (Windows booting for one), but I definitely find the later better overall for my use.

[aidanjt]

The raptor excels in random seeks as much as sustained throughput. The problem becomes when you pack a load of junk on a raptor and fill it up, then the head has to do a load of thrashing and NCQ has a hard time keeping up, 500GB disks on the other hand have plenty of space for NTFS to throw junk around in it's normal fragmented manner. With a sane filesystem Raptors absolutely trash ordinary 7200rpm disks in every regard (other than capacity).

[TooNice]

Excels, sure. Incomparable when it comes to seek time, definitely (SCSI/SSD excluded). And I certainly wouldn't be using one if I did not believe that it is faster for the applications I run (though quite frankly, I find the term 'trash' rather suspect). Fact remains that even after a fresh install, transfer rate has always been shy of 90MB/sec. My experience does not seem to be widely different from various reviews, who usually find outer zone performance to be in the mid-80s MB/sec. At least one 7200 drive edges the Raptor in that aspect, and the max/min average aspect.

[oralpain]

A common misconception.

thats not strictly true.

A Raptor will muller anything. EXCEPT...cos they're so small, as they fill up they get slower.

But kept as a clean boot drive...they're very fast indeed.

But expect to be able to use it for XP, plus apps, plus all your games...they get slow

As someone who has owned and used multiple raptors and who has replaced them with segate ST250410AS 7,200 rpm drives, I can say, from experience, that overall, these drivers are faster than raptors in many cases.

In general use the segate is on par, while working with massive ammounts of small files it's a bit slower, and in sequential transfer rates these segates are faster (I get more than 90MB/s in the middle of the drive, outer zone is about 105 MB/s).

There have been some rather noticeable changes in 7,200 rpm drives over the last few months, and I can scarcely recomend a raptor to anyone, unless 5% more performance is worth 300% more money per GB, or those who are building lower-mid end servers, where the gap is noticeable (and in favor of the raptor).

ADFD raptors are still pretty good (but do not beat the fastest 7,200 rpm drives in many areas), but the older GD models are beat in nearly every conceivable app by nearly every perpendicular recording drive.

[charleski]

And the whole raid0 = failure_rate * n_disks assumption has already been beaten to death and dismissed on these forums. 1 disk has an equal chance of bricking itself as 1 of a 2 disk setup.

Nope, I don't see the argument backing up this idea.

I read through the paper you linked. It's certainly very interesting, but it's analysing striped RAID arrays. The authors refer to their main concern several times in the paper - the problem of a second disk in a striped array failing while the array is being rebuilt. Their data analysis is targeted towards providing an estimate of the chances of failure of a 'fault-tolerant' RAID array. In such a situation the question of the independence of failure rates is very important, and the authors show that these rates demonstrate a significant degree of correlation.

But the problems with RAID0 remain the same, no matter how the disk failure pattern is distributed. In fact, the question of dependencies in failure rates is irrelevant to this particular question. A RAID0 array fails if just one disk fails. Multiplying the number of disks multiplies the probability of failure. This is just as true of a system with inter-dependencies.

I can't find a thorough analysis of real-world RAID0 failure rates, do you know of any? I doubt that there's much more than anecdotal data available, as no systems admin would use RAID0 on a large scale.

[aidanjt]

No sys admin would use single disk setups on a large scale, because there's no tolerance against failure either. The number of disks in a volume gives no probability of failure and there's no evidence to prove it as the paper pointed out quite clearly. Every hard drive is unique, but manufacturing flaws occur in batches, blanket statisticiation of HDD failure is completely wrong as is often the case with lousy short sighted statistics, when one ignores important variables and replaces them with a singular random probability factor you introduce errors into your equations.

What the paper clearly makes out is that the chance of cascade failure of a bad batch of disks is a highly likely event, not a random occurance. If you get a bad batch, one disk is just as likely to fail as 1 of 2 disks in the batch array (inevitably both disk will fail within a fairly short period of time given that they are subject to similar environmental factors).

Statistical probability always has, and always will be, the worst means to rationalise things. Random by it's very nature, and is akin to saying "I don't know, but here's an answer anyway". It's the same gripe Einstein had about quantum mechanics, he refused to believe in a universe that is based on entirely random events, and rightly so.

[this_is_gav]

ADFD raptors are still pretty good (but do not beat the fastest 7,200 rpm drives in many areas), but the older GD models are beat in nearly every conceivable app by nearly every perpendicular recording drive.

You're comparing a drive from 2004 to one from 2007? Really?

Still, my 36gb Raptors GDs (one from 2005 and a refurb from 2006) still beat my 7200.10s before both 7200.10s died. Rather put me off perpendicular recording to be honest...

[charleski]

The number of disks in a volume gives no probability of failure and there's no evidence to prove it as the paper pointed out quite clearly.

Well, what the paper demonstrated was that the factors causing HD failure apply across disks. It's still true that the more disks you have, the more likely it is that at least one of them will fail in an alloted time-span.

If you get a bad batch, one disk is just as likely to fail as 1 of 2 disks in the batch array (inevitably both disk will fail within a fairly short period of time given that they are subject to similar environmental factors).

I think this is where the misconception has arisen. The paper shows that there's a significant correlation between disk failures when measured over weeks or months. This is not surprising, given issues of batch manufacturing and (just as importantly) environmental factors. But, in systems composed of identical disks that rely on their perfect operation, one with two disks will still have twice the probability of failure of one of the two disks over a given time-span than one that only relies on one disk. It could be disk A that fails, or it could be disk B, but it doesn't matter which of them is the one to fail. The only claim the paper makes concerning the failure rate of individual disks (which is the problem that RAID0 faces) is that real-world failure rates are higher than those specified by the manufacturers. This result applies equally to all disk configurations, but the summation of probability that causes RAID0 arrays to fail more often than single disks still applies. Nothing in the paper challenges that idea, and indeed the authors don't even address it, because they're concerned with systems that rely on the sustained operation of multiple disks.

Statistical probability always has, and always will be, the worst means to rationalise things.

No, the problem is that too many people get lost in the numbers when dealing with statistics. It's always crucial to be absolutely sure you know what question you're asking and exactly how the numbers that you're dealing with relate to that question. If you keep those points clearly in mind then statistics provides very powerful tools to understand the data. This is why it's so important to work out exactly what questions this paper is asking. Their conclusion is quite clear, and very useful, that, even in 'fault-tolerant' RAID systems, the failure of one disk indicates a greater chance of failure in other discs from the same array and the entire set should be replaced, as you mentioned earlier.

It's the same gripe Einstein had about quantum mechanics, he refused to believe in a universe that is based on entirely random events, and rightly so.

And Einstein was wrong. Quantum mechanics is the single most successful scientific theory (in terms of predicting experimental results before the experiment is carried out) ever created.

[djg199]

I had a similar dilemma when looking at my newbuild and after reading this article http://www.tomshardware.com/2006/02/...formance_lead/ decided to go the Raptor route. This is mainly because I had a RAID 0 failure in the past. It may not be everyones cup of tea though.

[oralpain]

You're comparing a drive from 2004 to one from 2007? Really?

Still, my 36gb Raptors GDs (one from 2005 and a refurb from 2006) still beat my 7200.10s before both 7200.10s died. Rather put me off perpendicular recording to be honest...

I mentioned the newer raptors as well.

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).

[aidanjt]

Their conclusion is quite clear, and very useful, that, even in 'fault-tolerant' RAID systems, the failure of one disk indicates a greater chance of failure in other discs from the same array and the entire set should be replaced, as you mentioned earlier.

In one hand you're saying the chance of failure is a multiple factor to n disks, in the other you're agreeing that failures cascade. You're contradicting yourself. By your priori logic, after a failure of a disk in a given batch probability should recede. We know that isn't the case, in fact the risk of failure increases to a near certainty within 3 months. Which brings me back to, back of envelope anecdotes are worthless, and worse, misleading.

And Einstein was wrong. Quantum mechanics is the single most successful scientific theory (in terms of predicting experimental results before the experiment is carried out) ever created.

No he wasn't, you don't understand Quantum mechanical principles, in Quantum mechanics, you *cant* predict anything, because Quantum systems are inherently unpredictable, that's why String/M theory is being worked on. General and special relativity on the other hand, stills stand correct to date.

[TooNice]

[...]
In general use the segate is on par, while working with massive ammounts of small files it's a bit slower, and in sequential transfer rates these segates are faster (I get more than 90MB/s in the middle of the drive, outer zone is about 105 MB/s).
[...]
ADFD raptors are still pretty good (but do not beat the fastest 7,200 rpm drives in many areas), but the older GD models are beat in nearly every conceivable app by nearly every perpendicular recording drive.

Transfer rate is nice and all, but transfer rate is not the end it all.

Case to the point, the 2nd Gen Raptor is showing it's age for desktop applications (did so even prior perpendicular drives to be honest), is no record holder in transfer rate, but you'll have to exclude pretty most server applications as part of the "nearly" in the 'nearly every conceivable app'.

The ADFD is slower than the AAKS on average transfer rate, faster in in desktop applications. Since I do not have the 250GB platter 7200.10 I can not comment on its performance. But I'd like to see much the aural density bring outside transfer rate. Would like to see an SR review.

[charleski]

In one hand you're saying the chance of failure is a multiple factor to n disks, in the other you're agreeing that failures cascade. You're contradicting yourself.

That's not a contradiction. The factors causing HD failure may be common to multiple disks so that if one disk fails you know there's a higher chance of other disks in the set failing over the next 20 weeks. But this is getting away from the central point. The 'if' is important and you can only adjust your expectations after one disk has already failed. For a RAID0 array, that's too late.

By your priori logic, after a failure of a disk in a given batch probability should recede.

Well, I didn't say that, but the conventional wisdom is that HD failures are independent, and thus the failure of one disk should have no impact on the expectation of another disk failing (this is the exponential distribution the authors talk about). That is not the case. Again, this is only of importance in determining the probability of multiple disk failures.

in Quantum mechanics, you *cant* predict anything, because Quantum systems are inherently unpredictable, that's why String/M theory is being worked on. General and special relativity on the other hand, stills stand correct to date.

Quantum mechanics is not deterministic, and so does not predict individual events. If you want an example of the excellent quality of the predictions from quantum mechanics, this is a good example. Special and general theory are very good predictors as well, but operate on an entirely different scale. String theory fails to make any testable predictions at all. This is all very OT though .

[mroz]

That's not a contradiction.

Per se agreed.

The factors causing HD failure may be common to multiple disks so that if one disk fails you know there's a higher chance of other disks in the set failing over the next 20 weeks...

...but the conventional wisdom is that HD failures are independent,

I think this doesn't make sense.

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. Say if there was a causitive link between failures resulting in a cascade (I'm not sure this is strictly required but think it might be).

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, iow equally likely to go off in any time intervals of equal duration. The time to detonation will have an exponential distribution. Now put n of these in a room together. The time to first detonation will also have an exponential distribution but with a mean of one n'th the single bomb case. However we'll most certainly see a very tightly clustered distribution of detonations due to the minor causitive process of the first detonation triggering the rest & in so doing destroying their indpendence to that point.

Now back to the disks. If they are independent the mean failure time of the group will be the mean failure time of a single disk & the time to first failure will be bounded below by this mean failure time less the spread of the group. Hence in this case, tight clustering of failures implies the time to first failure will be approximately the same as the mean time to failure of a single disk.

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 (as the previously linked report states, not much data has been published) but mostly that they all share an inadequate grasp of statistics & probability theory, myself included. Apologies if I've got this wrong, but that's the impression I get. Stats is a bit odd as most everyday maths is intuitive, but stats can throw up some results which are counter intuitive & consequently cause problems when those trying to manipulate them have a shaky grasp, make assumptions & rely on analogies in place of reasoning, as opposed to logic & use of analogy purely as an illustrative aid.

My head hurts now. I never really liked statistics & I don't want to think about this any more. I have a shiny new box to play with & stress so I'm off to do that now. Hopefully I'll forget about this thread & all will be well again