The Radified Guide to Booting from a SCSI Drive implies that SCSI will deliver better performance than SATA.  However, when looking at the WD Raptor in a single-user (not server) environment, it outperforms a wide variety of SCSI (SAS) drives – even those running at 15K RPM (see  Single-User Performance).  Additionally, the speed of the interface (i.e., whether SATA or SCSI in Gb/s) isn't a limiting factor in hard disk drive performance, since the bandwidth is so much higher than what a SATA or SCSI drive can deliver anyway.

True, the WD Raptor is high-end SATA drive, but doesn't this comparison suggest that framing the problem in terms of the interface ("SATA versus SCSI") is misleading?

All comments are welcome.

I think the "SCSI is faster" mentality developed back in the ATA 33 days when the speed of the interface WAS the limiting factor.  SCSI can still outperform SATA in high end server RAID configurations.  On a desktop, I think it would be difficult to see the same performance. 

To say the least, I don't think that SCSI will give nearly enough performance increase to justify the increased cost.  SCSI controllers and drives are expensive.  You could get 3 WD Raptor drives and put them in RAID 5 for the same price as a good SCSI controller and drive.  Just thinking about the performance of RAIDed Raptors gets me excited.

hi pleo.

most scsi drives are designed for server use. in other words they're designed to optimize performance in of randomly accessed data, which is how a server basically, works .. gathering data here and there, from all parts of a disk, and writing data (uploaded data), but less of that.

a single-user, on the other hand, retrieves data more sequentially.

i'm simplifying, but my point is that disk manufacturers can (& do) write *firmware* to optimize these different patters of disk access, which is why ide/sata drives are able to perform closer to scsi drives in certain benchmarks than the number you mentions (seek/access) would otherwise indicate.

i can say that upgrading from a 7200-rpm ide (not sata) to 10-K rpm scsi was a "wow" thing.

I have used the raptor-based machines of two friends, which *did* seem zippy, but lacked the same level of responsiveness i observed on 15k-rpm scsi based machine.

the coolest thing about scsi (i'm talking about 15k-rpm scsi here) is its *responsiveness*. if you haven't played with a 15k scsi based rig, it's hard to describe, but you actually feeler synchronized with the machine, whereas with ide/sata, you feel a certain clunkiness.

i also feel that benchmarks can't adequately capture this sense of responsiveness i'm talking about.

is it worth it? is scsi worth the extra cost? that's the *real* question. for most, probably not, but it depends on what the user is doing.

i regularly get mail from folks asking if they think scsi would be worth it for them. if i remember, i'll post such a mail at the end of this post, (maybe *after* it, so I don't hit the 9999 character limit).

since my days of scsi, i have moved on from pure performance to *reliability* .. and there's where scsi drives shine. they are made to higher standards (designed for hard 24x7 enterprise use).

i think if somebody has the money, they won't be disappointed with the performance (and reliability) they experience with scsi, but most aren't going to go that way.

the cost of the adapter must be factored in, but remember, you can take the adapter with you to you next rig/motherboard, so the number don't correlate as you might expect.

also, there is one drive that now has its firmware tweaked for desktop (not server) use, and that drive is the one to get. i forget off the top of my head ... fugitsu mag, maybe?

i used to have lots of money to play with, and scsi was fun. now that i have lawyer's bills, i don't buy scsi anymore. but many people have buckets of money (like i used to).

scsi drives spin faster (lower rotational "latency") and they have big-ass actuators .. to move the read/write heads around faster.

just to recap (cuz i said a lot), i think the key points are that:

1. the low seek/ access times of (15k-rpm) scsi drives scsi drives provide a level of *responsiveness* that you can't get with 10k-rpm sata drives .. no matter *what* (artificial) benchmarks say. you feel one-with-the-machine.

2. scsi drives, which are designed for 24x7 enterprise, are built to higher standards than their sata counterparts, and are therefore more reliable.

3. the average joe computer-user isn't going to find the cost diff "worth it" for scsi. if you read my scsi guide, you know the main point i make is "scsi is not for everyone". with 10k-rpm sata, i'd now say that "scsi isn't for most people".

4. the firmware of scsi drives (all but one) is programmed differently than that for a sata drive.

5. the more a user does with his system and the more intensive things he does with it, i feel, the more he will appreciate "the feel" of a system based on a 15k-rpm scsi drive.

in the old days, when the scsi/ide wars were raging, i'd find it funny that someone who had never used a scs-based system would tell someone else, that scsi wasn't worth it FOR THEM.

Rad, I believe I may be able to reconcile, on the one hand, your observation that the WD Raptor doesn't have quite the same responsiveness as SCSI drives (Reply #4); and, on the other hand, the published benchmarks showing that the Raptor exhibits faster real-world performance (Reply #0).

As compared to the first-generation, the second-generation Raptor drives are 14% faster; and, the third-generation units are 17% faster than the second.  Thus, if you were experiencing a first-generation model of the Raptor, you might have concluded that they are not as fast as SCSI, while at the same time that the real-world benchmarks of a third-generation Raptor presently show it to be faster (in a single-user environment) than SCSI.

Could this explain the discrepancy?

Returning to the reliability issue, it is worthwhile to note that the WD Raptor has a 1.2 million hour MTBF specification, while the Seagate Cheetah 15K.4 SCSI, for example, is rated at 1.4 million hours – i.e., they are essentially the same.

having the same waranty period doesn't mean the drives are equally reliable.

i'd wager they have to make good on those raptor warantys far more than on those for their cheetahs.

replacing a drive, even if under waranty, still suks.

there was a study published recently (a few months ago?) by google, who owns a zillion hard drives, and they basically said drives start dying agter only 2 years of operation .. i think the number was somewhere around 7-12% (from memory) .. each year after the 2nd.

surprised you didn't see that. interesting data.

but the big message was that the mtbf figures published by drive manufacturers are essentially bunk. wishful thinking. not real world, where we inadvertantly bang and kick our pc's, crach into the desk holding our pc's.. etc. i'm sure their test drives operate in vastly different environments (climate controlled, etc.).

have to be careful about using the word "faster" when it comes to hard drive benchmarks, which are artificial routines, designed by some benchmark creator .. which may or may not represent how you & I use our hard drives.

for example, when you use the % increase figures for raptors, does that mean they decreased their respective seek/access times for each generation, or simply boosted str's (sustained transfer rates), which is typically a product of data density.???

these numbers given are often subjective, and can be analyzed in numerous ways.

what are the respective seek/access times for each raptor generation?

the wiki on raptors does not give any seek/access specs, which is the key spec that translates into responsiveness.

none of the raptors-based systems i tried contained the clear-window drive, so i'm sure it wasn't the most recent version (which is cool).

in my most recent system design (last year), i selected a raptor for my boot/system drive:

http://radified.com/computer/computer_2006.htm

Rad, the subject of hard disk reliability is a very complex subject.  (I did see and read the Google research.)

Yes, a warranty period isn’t a perfect indicator of the reliability of the drive, but it should at least be positively and directionally correlated with reliability.

I do agree that MTBF isn’t a perfect measure of reliability.  (A good article on the subject is Mean time between failures.)  Do note that a MTBF of 137 years does not imply that on average a drive will run continuously and properly for over a century (a common misinterpretation of the statistic)!  However, as a metric to compare one hard disk drive (or manufacturer) against another, what is the alternative?

I see your objective in Reply #8, but the math isn’t quite right.  The problem is that the longevity estimate itself varies as a function of time.  For example, you can compute the average lifespan of a population of humans, but it is not the same as computing the projected lifespan of a person given that s/he is now 65 years old.  There is a Bayesian component to the problem.

There are literally an infinite number of ways to measure hard disk drive performance.  Thus, any benchmark study is a sample from that universe and as such is subject to sampling error.  However, again, it worth asking:  what is the alternative?  The benchmarks used by StorageReview appear to have a solid, real-world orientation.  For example, the SR Office DriveMark 2006 benchmark is based on the usage of a mix of commonplace applications, including Microsoft's Office (Word, Excel, Access, Outlook), Internet Explorer, Symantec Antivirus and Winzip.  That sure seems quite an appropriate metric.

In the real world, statistics like the warranty period, MTBF, and independent product reviews and benchmarks are quite valuable, from my perspective.  None of these are without limitations, of course; but, using them results in a reduction of uncertainty in the decision making process that would otherwise be absent.

This is a good discussion.  Please continue to add your perspective!  

Hey, Rad – good to hear your comments.

The problem with the Google research is that it doesn’t offer reliability numbers by drive manufacture or model – right?  Thus, it is useful at an overall level, but wouldn’t help in evaluating the reliability of one drive versus another.

I absolutely agree that the MTBF numbers are not equivalent to - nor were they ever intended to be - representative of the useful lifetime of a drive.  Many individuals misinterpret the statistic in that way.  Like the duration of the warranty, however, they are suggestive of (but do not conclusively demonstrate) the reliability of the drive.

I don’t know exactly how the SR Office DriveMark 2006 benchmark works, but my reading of the description is that the benchmark is a script that invokes the set of applications to perform predetermined tasks, and the total duration is timed.  Thus, I do believe that the actual applications (e.g., Excel, WinZip, etc.) are in fact launched and run.  It is possible – but highly unlikely – that Western Digital or Seagate cares so much about StorageReview that they would tailor their hardware to artificially inflate their benchmark scores.

A benchmark will not match the actual usage pattern of any individual user, of course.  That, however, is not the intent.  The goal is to provide a standardized metric that reasonably mirrors a real-world usage scenario, so as to permit comparisons among drives.  It is the ability to compare drives that is facilitated by benchmarks, not the ability to predict the performance that a specific user will experience (which is highly dependent on her or his unique usage of the equipment).

I don’t think that reliability and performance are at opposite ends of the same continuum.  You could have low/high reliability crossed with low/high performance in a 2 x 2 matrix, and probably find hard disk drives representative of each quadrant.

P.S.:      Glad to hear you earned an “A” in statistics!   That is an essential prerequisite to being a well-educated individual, in my opinion.