3.5" SATA->USB drive enclosure with good cooling?

[Bill Todd]

It doesn't help the latency for a completely serial workload, but it
lets you do seeks in parallel when the workload allows it.

You seem to have a better handle on this subject than most of the people
who are presuming to advise you on it.

1. Conventional PATA/SATA drives are indeed less reliable for 24/7
server-style workloads than higher-end drives (which indeed do use more
robust internals) - today, as well as in the past. Microsoft and
Seagate representatives held a joint presentation at last year's WinHEC
describing stress tests on three groups of 300 desktop drives which
found that failure rates went up significantly when the drives were used
24/7 rather than at their rated duty cycles (typically 8 hrs/day).
After 1000 hours of continuous operation the failure rate using a
typical desktop workload (lots of idle time) was 1.7% (about 10x what
one would expect after 1000 hours for a desktop drive rated at 600,000
hours MTBF in 'normal' use, and with especially considerate handling
Copan Systems found that 18,000 disks used in its 'MAID' products over a
two-year period starting in April, 2004, had a real-world failure rate
nearly 5x lower still, corresponding to nearly a 3,000,000 hour MTBF).
The failure rate after 1000 hours under constant streaming access (i.e.,
no significant seeks) was more than twice as high (4%), and the failure
rate after 1000 hours for a constant seek-intensive workload was higher
still (6.3% and still rising steeply, reaching 7.4% after 1200 hours,
while the other two workload failure curves were flattening out). It's
too bad they didn't test Raptors (well, Seagate *was* the joint
presenter...) - I've wondered how well they lived up to WD's marketing
in that area, which certainly suggests that they're full-fledged
high-end-FC/SCSI replacements in terms of reliability. SATA failure
rates for somewhat less demanding (but still 24/7 and fairly heavy)
workloads in Microsoft's TerraServer application were stated a year ago
to be 7.2% annually (about 5x the predicted failure rate for a
600,000-hour MTBF disk in its specified 'desktop' duty cycle). The
Copan paper ("Disk Failure Rates and Implications of Enhanced MAID
Storage Systems") is a good general starting point for this topic, and
contains the interesting tidbits that having more platters (because of
increased loading on the motor and head actuators?) and more power-on
hours (apparently even if the disk is spun down) increases failure
rates. 'Near-line enterprise' drives are specced for 24/7 operation but
not for the kind of intense workloads that the high-end enterprise
drives are: it would be nice to see some head-to-head stress testing
comparing them to conventional desktop SATA drives under both light and
heavy workloads, since they cost very little more (though conventional
drives may be more deeply discounted during sales). And as you already
know, keeping the disk cool is a significant win (I vaguely remember
reading somewhere that each 10 degree C. rise in operating temperature
doubled the failure rate).

2. So yes, your inclination toward using a Raptor purely for
reliability is entirely reasonable - though if you can tolerate a disk
failure during your usage then just figuring you'd replace a
conventional SATA drive if it failed (or even if two of them failed)
might be more economical.

3. The assertion that USB access significantly increases latency
appears to be false - at least from a quick look at some tests at Tom's
Hardware, where the listed external USB drive access times are
comparable to internal drive access times. This makes sense, since
three's no overhead in the purely hardware/software USB processing
that's remotely comparable to the a disk's seek and rotational latency.
While as you claimed the Raptor's higher rotational speed would
decreases average access times (by about 1 ms.), its much faster average
seek times would likely be more important to random-access performance
(all as Folkert observed: he may seldom be civil, but that doesn't mean
he's always wrong).

4. You were correct (and Folkert was wrong) in your suggestion that
using multiple slower drives can be a cost-effective way to improve
throughput (even random-access throughput) for parallel workloads. The
secret is not to use traditional small stripe-segment sizes. Storing
only 64 KB (or even less) per disk before moving on to the next *never*
made much sense: with today's disks, 1 - 4 MB segments allow you to
maximize streaming throughput (it takes a good-sized client buffer to
inhale multiple multi-MB segments in parallel, but RAM's cheap these
days) while still placing a reasonable upper limit on smaller-access
latency (so the worst case may rise from 20 ms. to 80 ms. - BFD, since
for the kind of parallel workload you described this will be balanced by
reduced queuing delays; in fact, these stripe-segment sizes ensure 70% -
90% disk utilization for *any* workload, unlike smaller stripe segments
which can reduce utilization by a factor of N = the number of disks that
s modest access spans).

- bill

 

[David Flory]

This is the first post I've actually saved to my PDF archives. Very
informative! Thanks. I hadn't heard about those tests.

Dave

 

[Rod Speed]

Paul Rubin wrote

You seem to have a better handle on this subject than most of the people who are presuming to
advise you on it.

We'll see...

1. Conventional PATA/SATA drives are indeed less reliable for 24/7 server-style workloads than
higher-end drives (which indeed do use more robust internals) - today, as well as in the past.

Easy to claim, hell of a lot harder to substantiate that claim.

Microsoft and Seagate representatives held a joint presentation at last year's WinHEC describing
stress tests on three groups of 300 desktop drives which found that failure rates went up
significantly when the drives were used 24/7 rather than at their rated duty cycles (typically 8
hrs/day).

What matters is whether they were adequately cooled.

After 1000 hours of continuous operation the failure rate using a typical desktop workload (lots
of idle time) was 1.7% (about
10x what one would expect after 1000 hours for a desktop drive rated at 600,000 hours MTBF in
'normal' use, and with especially
considerate handling Copan Systems found that 18,000 disks used in
its 'MAID' products over a two-year period starting in April, 2004,
had a real-world failure rate nearly 5x lower still, corresponding to
nearly a 3,000,000 hour MTBF). The failure rate after 1000 hours
under constant streaming access (i.e., no significant seeks) was more than twice as high (4%), and
the failure rate after 1000 hours for a constant seek-intensive workload was higher still (6.3%
and still rising steeply, reaching 7.4% after 1200 hours,

Thats nothing like 'rising steeply'

while the other two workload failure curves were flattening out). It's too bad they didn't test
Raptors (well, Seagate *was* the joint presenter...) -

I've wondered how well they lived up to WD's marketing in that area,
which certainly suggests that they're full-fledged high-end-FC/SCSI
replacements in terms of reliability. SATA failure rates for somewhat less demanding (but still
24/7 and fairly heavy) workloads in Microsoft's TerraServer application were stated a year ago to
be 7.2% annually

Still bugger all difference in those numbers and it could easily
be just the model detail, not the 'enterprise drive' question.

(about 5x the predicted failure rate for a 600,000-hour MTBF disk in its specified 'desktop' duty
cycle). The Copan paper
("Disk Failure Rates and Implications of Enhanced MAID Storage
Systems") is a good general starting point for this topic, and
contains the interesting tidbits that having more platters (because
of increased loading on the motor and head actuators?) and more
power-on hours (apparently even if the disk is spun down) increases
failure rates.

It'd be a hell of a lot more surprising if it didnt.

'Near-line enterprise' drives are specced for 24/7 operation but not for the kind of intense
workloads that the high-end enterprise drives are: it would be nice to see some head-to-head
stress testing comparing them to conventional desktop SATA drives under both light and heavy
workloads, since they cost very little more (though conventional drives may be more deeply
discounted during sales). And as you already know, keeping the disk cool is a significant win (I
vaguely remember reading somewhere that each 10 degree C. rise in operating temperature doubled
the failure rate).

Thats always been a number plucked out of someone's arse.

2. So yes, your inclination toward using a Raptor purely for reliability is entirely reasonable -

Have fun explaining the storagereview stats.

though if you can tolerate a disk failure during your usage then just figuring you'd replace a
conventional SATA drive if it failed (or even if two of them failed) might be more economical.

3. The assertion that USB access significantly increases latency appears to be false

You need to get your eyes tested then.

- at least from a quick look at some tests at Tom's Hardware, where the listed external USB drive
access times are comparable to internal drive access times.

Pity about the USB protocol overhead that stands out like dogs balls.

This makes sense, since three's no overhead in the purely hardware/software USB processing that's
remotely comparable to the a disk's seek and rotational latency.

Have fun explaining the massive difference with
the same drive internal and connected with USB2.

While as you claimed the Raptor's higher rotational speed would decreases average access times (by
about 1 ms.), its much faster average seek times would likely be more important to random-access
performance

Have fun explaining the massive difference with
the same drive internal and connected with USB2.

 

[Bill Todd]

We'll see...

Indeed we will.

Easy to claim, hell of a lot harder to substantiate that claim.

Perhaps you need remedial reading lessons: I already provided two
independent substantiations (in addition to those in the paper that Paul
originally cited - which your response to him indicated that you
couldn't understand either, so at least you're consistent).

You, by contrast, have provided absolutely squat in the realm of
substantiation. Perhaps you're really God's Gift to the Storage
Industry and I just never got the word, but until I do I'll be inclined
to consider you just another incompetent Usenet blowhard.

What matters is whether they were adequately cooled.

Based on the level of ignorance that you've already displayed you
probably don't know who Jim Gray (something of a living legend in the
industry) is - but for those who *do* know who he is I'll just mention
that he was involved in both of the evaluations that I cited, and that
therefore the probability that the systems were not properly set up (and
cooled) is very likely indistinguishable from zero.

Thats nothing like 'rising steeply'

That, of course, is purely a matter of how you scale the graph, at least
in any absolute sense. However, my comparison was relative to the other
two failure rates (as you'd have seen had you bothered to follow up the
reference instead of just continuing to bloviate incompetently).

Still bugger all difference in those numbers

Ah - I see that you are as incompetent mathematically as you are
linguistically. But since I've got better things to do than attempt to
educate someone who I strongly suspect is relatively ineducable, I'll
just leave it to the rest of the group to decide for themselves how
relevant the differences may be (having already presented them clearly
once).

and it could easily

be just the model detail, not the 'enterprise drive' question.


It'd be a hell of a lot more surprising if it didnt.

Not really: just making some parts larger without increasing
*complexity* much could well have relatively minor effects on
reliability (though it could cause additional heat to be generated, a
subject apparently near and dear to your heart, and thus have additional
indirect impact - had you understood the paper that Paul cited, you'd
have seen that observation there), and absent any mechanical
considerations (since the disk is already spun down) electronics are
pretty reliable running under steady-state idle conditions and at least
somewhat sensitive to being powered up and down promiscuously. But my
main point was that the Copan paper found the two differences actually
worth mentioning rather than merely measurable (of course, that could
have been for marketing purposes: I didn't check at the time to see
just how 'worth mentioning' the numbers actually were, since my
observation was purely incidental to the subject at hand, but in
revisiting the paper now it looks like the power-on hour information was
taken from the Seagate paper, and while the difference was indeed
substantial it may well have related to spinning rather than spun-down
powered-on drives - perhaps 'spun down' is what Copan actually means
when it says 'powered down').

Thats always been a number plucked out of someone's arse.

Wrong again, I'm afraid: while my memory was faulty (it's actually a
*15* degree C. rise that doubles the failure rate), that information
comes directly from arguably the world's premier manufacturer of disk
drives, who's surely in an immeasurably better position to know (and in
fact in the pursuit of its business be very intimately concerned with)
such things than you are. Once again, if you'd actually understood the
paper that Paul cited, you'd have seen that there.

Have fun explaining the storagereview stats.

You mean the ones that show the new Raptor to be, by a considerable
margin, the most reliable evaluated disk that WD sells (with its
predecessor close behind in second place), more reliable than all but 3
specific members of the Seagate family (two of which are Cheetahs), and
more reliable than any evaluated drive from any other manufacturer?
Since one might suspect that Raptors are used in more demanding
environments than the average SATA drive, those numbers become even more
impressive.

Do you just throw out such misleading challenges under the assumption
that the people you're talking to are as lazy about checking such
allegations as you are?

You need to get your eyes tested then.

I'm afraid that that won't be possible until you provide something more
substantial than your own personal drivel to test them on.

Pity about the USB protocol overhead that stands out like dogs balls.

Feel free to provide something at least remotely resembling quantitative
evidence of that (since I already provided some to the contrary: I
hardly consider Tom's Hardware a definitive source, but it was readily
accessible and is still a hell of a lot more substantial than anything
you've offered up).

Have fun explaining the massive difference with
the same drive internal and connected with USB2.

Since that's the comparison I made at Tom's site, and since I found
nothing remotely resembling that situation, there's really nothing *to*
explain (except perhaps the rather impressive level of incompetence that
you're displaying here, but I'm not sufficiently interested to delve
into the reasons for that).

Have fun explaining the massive difference with
the same drive internal and connected with USB2.

You seem to be repeating yourself - do try to get a grip.

- bill

 

[Rod Speed]

Rod Speed wrote

Indeed we will.

Indeed we have.

Perhaps you need remedial reading lessons:

No perhaps about your problem.

I already provided two independent substantiations

No you didnt on the GENERAL claim.

(in addition to those in the paper that Paul originally cited - which your response to him
indicated that you
couldn't understand either, so at least you're consistent).

Never ever could bullshit its way out of a wet paper bag.

You, by contrast, have provided absolutely squat in the realm of substantiation.

YOU made that stupid pig ignorant claim.

YOU get to do the substantiation.

THATS how it works.

Perhaps you're really God's Gift to the Storage Industry and I just never got the word, but until
I do I'll be inclined to consider you just another incompetent Usenet blowhard.

Never ever could bullshit its way out of a wet paper bag.

Hardly surprising that three times the use sees more failures, ****wit.

Based on the level of ignorance that you've already displayed you

Never ever could bullshit its way out of a wet paper bag.

probably don't know who Jim Gray (something of a living legend in the industry) is

Rigorous science/engineering might just be about EVIDENCE, ****wit.

- but for those who *do* know who he is I'll just mention that he was involved in both of the
evaluations that I cited, and that therefore the probability that the systems were not properly
set up (and cooled) is very likely indistinguishable from zero.

What matters is the temperature data actually cited, ****wit.

All that shows is the irrelevance of MTBFs to the real world, ****wit.

Or that is due to a difference in the models actually used.

That, of course, is purely a matter of how you scale the graph, at least in any absolute sense.

Pathetic, really.

However, my comparison was relative to the other two failure rates
Liar.

(as you'd have seen had you bothered to follow up the reference instead of just continuing to
bloviate incompetently).

Never ever could bullshit its way out of a wet paper bag.

Ah - I see that you are as incompetent mathematically as you are linguistically.

Never ever could bullshit its way out of a wet paper bag.

But since I've got better things to do than attempt to educate someone who I strongly suspect is
relatively ineducable, I'll just leave it to the rest of the group to decide for themselves how
relevant the differences may be (having already presented them clearly once).

Never ever could bullshit its way out of a wet paper bag.

Cant ignore that, it aint gunna go away, ****wit.

All that shows is the irrelevance of MTBFs to the real world, ****wit.

Not really:

Fraid so.

just making some parts larger without increasing *complexity* much could well have relatively
minor effects on reliability

Thanks for that completely superfluous proof that
you couldnt bullshit your way out of a wet paper
bag. Or manage even the most basic stuff either.

(though it could cause additional heat to be generated,

And require more from the rotation motor, etc etc etc.

a subject apparently near and dear to your heart,

Never ever could bullshit its way out of a wet paper bag.

and thus have additional indirect impact

Funny that.

- had you understood the paper that Paul
cited, you'd have seen that observation there),

Its so obvious its superfluous, ****wit.

and absent any mechanical considerations (since the disk is already spun down)

It aint spun down all the time, ****wit.

electronics are pretty reliable running under steady-state idle conditions and at least somewhat
sensitive to being powered up and down promiscuously.
Waffle.

But my main point was that the Copan paper found the two differences actually worth mentioning
rather than merely measurable

I didnt comment on that, ****wit.

(of course, that could have been for marketing purposes: I didn't check at the time to see just
how 'worth mentioning' the
numbers actually were, since my observation was purely incidental to the subject at hand, but in
revisiting the paper now it looks like
the power-on hour information was taken from the Seagate paper, and while the difference was
indeed substantial it may well have related
to spinning rather than spun-down powered-on drives - perhaps 'spun
down' is what Copan actually means when it says 'powered down').

I didnt comment on that either, ****wit.

Wrong again, I'm afraid:
Nope.

while my memory was faulty (it's actually a *15* degree C. rise that doubles the failure rate),
that information comes directly from arguably the world's premier manufacturer of disk drives,

All that proves is where the arse resided, ****wit.

who's surely in an immeasurably better position to know (and
in fact in the pursuit of its business be very intimately concerned
with) such things than you are. Once again, if you'd actually
understood the paper that Paul cited, you'd have seen that there.

Never ever could bullshit its way out of a wet paper bag.

You mean the ones that show the new Raptor to be, by a considerable
margin, the most reliable evaluated disk that WD sells (with its
predecessor close behind in second place), more reliable than all but
3 specific members of the Seagate family (two of which are Cheetahs),
and more reliable than any evaluated drive from any other manufacturer?

Lying, again.

Since one might suspect that Raptors are used in more
demanding environments than the average SATA drive,

Stupid assumption with the storagereview stats.

those numbers become even more impressive.

Only in you pathetic little drug crazed pig ignorant fantasyland.

Do you just throw out such misleading challenges under the assumption that the people you're
talking to are as lazy about checking such allegations as you are?

Never ever could bullshit its way out of a wet paper bag.

I'm afraid that that won't be possible until you provide something
more substantial than your own personal drivel to test them on.

Never ever could bullshit its way out of a wet paper bag.

Feel free to provide something at least remotely resembling quantitative evidence of that (since I
already provided some to the contrary:

Lying, as always.

I hardly consider Tom's Hardware a definitive source, but it was readily accessible and is still a
hell of a lot more substantial than anything you've offered up).

YOU made that stupid pig ignorant claim.

YOU get to do the substantiation.

THATS how it works.

Since that's the comparison I made at Tom's site,

Pity about the ear to ear dogshit of yours involved when you did.

and since I found nothing remotely resembling that situation, there's really nothing *to* explain

Is that so ? No one has ever produced any benchmarks
that show that a USB2 connected drive performs substantially
worse than the same drive internally connected eh ?

(except perhaps the rather impressive level of incompetence that you're displaying here, but I'm
not sufficiently interested to delve into the reasons for that).

Never ever could bullshit its way out of a wet paper bag.

You seem to be repeating yourself

That sometimes helps with ****wits as stupid as you, ****wit.

- do try to get a grip.

Let go of yours before you end up completely blind.

 

[chrisv]

YOU made that stupid pig ignorant claim.

YOU get to do the substantiation.

THATS how it works.

Umm... He did provide substantiation, Ron^Hd, while you just threw a
tantrum.

You got your ass kicked, Ron^Hd, and I doubt your attempt to bullshit
your way out of your predicament fooled anyone.

 

[Rod Speed]

Some gutless ****wit desperately cowering behind
expect from a desperately cowering gutless ****wit that
has never ever managed to contribute a damned thing, ever.