RE MTBFs...
Hi throAU,
I agree with most everything you say (see the P.S. for an exception), especially the part about MTBF numbers. Anecdotal evidence seems to suggest that MTBF numbers do not apply to actual commercially bought drives. I've always wondered about this.
Of course, R/W errors on HDDs happen all the time, with most of these errors being recoverable. HDDs have "spare" blocks on them to handle just such errors, and these "spare" blocks are probably the reason for such long MTBF numbers. I wonder if the MTBF values are not calculated based upon the number of "spare" blocks and the average rates of recoverable R/W errors? While an actual catastrophic failure of a HDD occurs in practice not when all of the "spare" blocks are used up but rather when a specific block fails or a vibration resonance occurs that sets the flying head crashing into the rotating disk, or some such event that is not factored into the MTBF values. From anecdotal evidence, I'd have to say that MTBF numbers do not apply to actual usage in physical RAID systems.
Maybe I am one of your "edge cases", as you so eloquently put it. But I do use the added R/W speeds and it allows me to accomplish more work faster because of what I do and how I use the RAID 0. Personally, I don't think of my usage as "edge", rather just number crunching on large data sets. I kind of resent your implications (actually rather strong) that I personally don't need the faster speeds, how would you know? If I didn't need it, why would I go to the trouble and expense of building it? The disk R/W speeds were the bottleneck for my computations before, now with a RAID 0 SSD the bottleneck is elsewhere. But I don't want to get into that here...
And my contention still stands: double a small number is still a small number -- just because your risk for failure is double should not necessarily mean that you should not pursue the technology... If you would not pursue RAID 0 SSD because of double the risk of failure, then I would question whether you should pursue standard SSD in the first place.
Let me give you an example illustrating my thinking about this scenario...
Let's say that in an average number-crunching computer lifetime, say 5 years albeit this is probably on the long side, of 100 such machines having SSDs, that 1 will fail, then using a RAID 0 of 2 such SSDs on those 100 machines will increase the risk to 2 failed systems out of the 100 machines during their lifetimes. Now let's say that using the RAID 0 increases the throughput by a factor of 1.2 (2 is maximal, of course, but most codes won't run twice as fast even when the R/W speeds are twice as fast --- but put whatever reasonable number you want here, say the factor of 1.2 that I suggest - this factor depends upon exactly what code you are running). If we assume an increase throughput by a factor of 1.2 for the RAID 0 over the SSD, then during the 5 years of use we gain a full year of results (6 years of computations compressed into 5 years of time by using the RAID 0 versus the standard SSD). Personally, the guaranteed 1 year of time savings is well worth my double risk of RAID 0 failure. In other words, I would gladly trade the 2 in 100 chance of failure for the 1 year of my time, especially since it is relatively easy and no time cost for me to protect against the RAID 0 failure --- in fact, I already protect against the 1 in 100 risk of failure for the single SSD drive, so protecting against the 2 in 100 risk for the RAID 0 SSDs is not more expensive for me, neither in time nor in money. Basically, I appreciate the guaranteed gain of 1 year of my time versus the very small chance that I will have a failed drive. I think that my chances are good that I will be one of the 98 individuals that don't have their RAID 0 SSDs fail (probability is 98 out of 100), but I will have 100 out of 100 chances of gaining 1 year of computation time. One year gain in computation time can mean the difference between success and failure, the difference between publishing first and not publishing at all.
Of course, put whatever reasonable numbers you want into the above vignette, say you believe that 5 or even 10 out of 100 machine will have failing SSDs, then the number of failing RAID 0 SSDs would increase to 10 to 20 out of 100. Still worth the risk to me because of the year I gain in computation time, and I do already protect against SSD failure with multiple backup systems. Basically, as long as the RAID 0 SSDs last long enough to complete a single computation so that I actually obtain publishable results, then I would argue that the risk of failure of RAID 0 is overridden by the gain in time.
Regards,
Switon
Edit: P.S. Your comment that "murphy is a prick..." I believe does not apply here, since Murphy's Law is a compounding effect (geometric or exponential growth) while the increase in failure risk is a linear growth thing. If the RAID 0 failure was indeed exponential and not linear in the number of drives (but we both agree it is linear), then we'd have to be more careful in our analysis and your contention might just win out in the final analysis. But being linear means that my above scenario applies and not Murphy's Law. Murphy's Law is often invoked in cases where it is not applicable; where it is not a valid model for the situation at hand -- I believe that this is the case here, Murphy's Law is not applicable.
