Striping data across platters in a single hard disk

[tony]

Why don't hard drive manufacturers figure out a way to
stripe data across the multiple platters in a hard drive to
get greater throughput? It seems to me that if you have
multiple heads and platters you should be able to multiply
the throughput to some extent based upon the number
of heads/platters: a 2 platter drive should get almost twice
the performance of a single platter drive. Why doesn't it
work that way/how does it work?

Tony

 

[Keith]

Why don't hard drive manufacturers figure out a way to
stripe data across the multiple platters in a hard drive to
get greater throughput? It seems to me that if you have
multiple heads and platters you should be able to multiply
the throughput to some extent based upon the number
of heads/platters: a 2 platter drive should get almost twice
the performance of a single platter drive. Why doesn't it
work that way/how does it work?

Only one head controls the head servo positioning at a time. Only
one head is operational at a time. Note that an adjacent track
seek takes about the same time as a head to head switch.

 

[GSV Three Minds in a Can]

Only one head controls the head servo positioning at a time. Only
one head is operational at a time. Note that an adjacent track
seek takes about the same time as a head to head switch.

Or more basically, 'just because head0 is above Track0 on platter0, you
cannot assume that head1 is above Track0 on plattert1' .. maybe it was
when you wrote, but it might not be when you read some time later.

It might be slightly less risky having two 'side by side' heads on the
same platter (remember 'drums', anyone? (Anyone really ancient I mean)),
but even that probably won't work, given thermal effects.

 

[The little lost angel]

Only one head controls the head servo positioning at a time. Only
one head is operational at a time. Note that an adjacent track
seek takes about the same time as a head to head switch.

I have always thought a drive with 4 heads would read/write 1 chunk of
data in parallel of 4 pieces at the same time. But looking around due
to the OP's question, I think I've been mistaken. So like the OP is
wondering, why don't they do that?

i.e. system request write 1 sector 512bytes. If the drive has 3 drive
heads (2 platter), each will write 171bytes plus padding (for last
head)/control/ecc. If 5 heads (3 platters), then each 103bytes + some
padding/control/ecc and so on.

It would be equivalent of having an inbuilt Raid 0 no?

 

[Keith]

I have always thought a drive with 4 heads would read/write 1 chunk of
data in parallel of 4 pieces at the same time. But looking around due
to the OP's question, I think I've been mistaken. So like the OP is
wondering, why don't they do that?

GSV has the answer. There is no guarantee all heads are on the
track. Only the active head is used to position the head actuator
over the track. A switch to another head involves seeking and
recentering that head over its track.

i.e. system request write 1 sector 512bytes. If the drive has 3 drive
heads (2 platter), each will write 171bytes plus padding (for last
head)/control/ecc. If 5 heads (3 platters), then each 103bytes + some
padding/control/ecc and so on.

Why the missing heads?

It would be equivalent of having an inbuilt Raid 0 no?

Equivalent, yes. Workable, no. ;-)

 

[daytripper]

On Thu, 5 Jan 2006 16:57:57 +0000, GSV Three Minds in a Can

It might be slightly less risky having two 'side by side' heads on the
same platter (remember 'drums', anyone? (Anyone really ancient I mean)),

Yo! The 2305 Model 1 was rolling out when I was working on my first design
job. A box the size of an upright freezer to hold 5MB (uncooked, at that)...

/daytripper (verging on decrepit, apparently ;-)

 

[tony]

Or more basically, 'just because head0 is above Track0 on platter0, you cannot assume that head1 is above Track0 on
plattert1' .. maybe it was when you wrote, but it might not be when you read some time later.

Couldn't something be enforced though (surely it can because it can be
done with RAID) to make a multi-platter drive do this? I mean not within
the constraints of the way the firmware/formatting/heads/electronics are
now, but in a brand new way that would make it work. IOW, what would
it take to make it work? Having the fast bus there and no way to take advantage
of it other than expensive RAID (double the cost of hard disks at least, but
do get reliability though) is a shame.

It might be slightly less risky having two 'side by side' heads on the same platter (remember 'drums', anyone? (Anyone
really ancient I mean)), but even that probably won't work, given thermal effects.

Tony

 

[Keith]

Couldn't something be enforced though (surely it can because it can be
done with RAID) to make a multi-platter drive do this? I mean not within
the constraints of the way the firmware/formatting/heads/electronics are
now, but in a brand new way that would make it work. IOW, what would
it take to make it work? Having the fast bus there and no way to take advantage
of it other than expensive RAID (double the cost of hard disks at least, but
do get reliability though) is a shame.

It's got nothing to do with the electronics. It's a physical
problem. The tracks spacing is smaller than the differences
between the platters (and that difference is a function of
temperature). I suppose you could independently control each head,
but that would mean an actuator per head. At that point why not
simply use separate drives.

 

[George Macdonald]

Why don't hard drive manufacturers figure out a way to
stripe data across the multiple platters in a hard drive to
get greater throughput? It seems to me that if you have
multiple heads and platters you should be able to multiply
the throughput to some extent based upon the number
of heads/platters: a 2 platter drive should get almost twice
the performance of a single platter drive. Why doesn't it
work that way/how does it work?

Well they used to do that when a single platter only held a few MBs (KBs if
you go back far enough) and spun at 3600rpm - don't recall if they went
slower but likely. With 160GB on a single platter now, you only need
multiple platters for huge drives or if you make a change in form factor.

It could also be that the more heads you have, the higher the chance of
mechanical failure - IOW aligning say 8 heads with 4 platters is very
difficult(?). The price of the on-board controllers is also orders of
magnitude lower than back in the old days so ganging a bunch separate
drives together in some kind of RAID formation is easy, cheap and
practical.

 

[The little lost angel]

GSV has the answer. There is no guarantee all heads are on the
track. Only the active head is used to position the head actuator
over the track. A switch to another head involves seeking and
recentering that head over its track.

Hmm, why does that matter? If all the heads/arms are fixed on the same
axle, thus even if they aren't above the same exact spot on their own
platter, wouldn't they always be above the same respective position?
Then as long as they are always in the same position relative to their
platter, the write/read would always be to the same spot as far as
that head/platter is concerned, right?

Unless you're implying that there are pre-etched spiral tracks like
CDROM on the HDD as well?

Why the missing heads?

1 side of the platter and 1 head is used for storing/reading tracking
information. So 2 platters only have 3 for data, if I'm not mistaken
by what I read.

 

[daytripper]

Hmm, why does that matter? If all the heads/arms are fixed on the same
axle, thus even if they aren't above the same exact spot on their own
platter, wouldn't they always be above the same respective position?
Then as long as they are always in the same position relative to their
platter, the write/read would always be to the same spot as far as
that head/platter is concerned, right?

Not when you throw thermal effects at the problem, unless you could guarantee
dead-matched platter geometry across operating temperature and through
temperature changes.

Unless you're implying that there are pre-etched spiral tracks like
CDROM on the HDD as well?

Well....not etched, of course....but the tracks *do* exist and never shift
location once they are created at the factory.

1 side of the platter and 1 head is used for storing/reading tracking
information. So 2 platters only have 3 for data, if I'm not mistaken
by what I read.

I think that paradigm died long ago with the advent of embedded servo data...

/daytripper

 

[craigm]

Hmm, why does that matter? If all the heads/arms are fixed on the same
axle, thus even if they aren't above the same exact spot on their own
platter, wouldn't they always be above the same respective position?

No, any change in temperature causes them to change position relative to
each other. Twenty years ago they would have been close enough, but
today track densities are much higher. Then drives had about 600 tracks
per inch, today you may see 60000+ tracks per inch.

Then as long as they are always in the same position relative to their
platter, the write/read would always be to the same spot as far as
that head/platter is concerned, right?

But the problem is that your assumption that they would be in the same
relative position doesn't work.

Unless you're implying that there are pre-etched spiral tracks like
CDROM on the HDD as well?
No.



1 side of the platter and 1 head is used for storing/reading tracking
information. So 2 platters only have 3 for data, if I'm not mistaken
by what I read.

That is also 20 year old technology. Today all hard drives use embedded
servo where servo and data reside on the same track. This was necessary
to deal with the temperature shift of the heads. Also with few disks,
this is a much more efficient use of the disk surface.



To accommodate your concept each head would have to be positioned
independently. This would vastly add to the cost of a drive.

You would also need to replicate the entire read/write electronics chain
for each head. Again, a significant increase in cost.

Now, when you add in defect management, you can run into a situation
where you have to map out the same areas on each surface to keep the
data synchronized across heads. This would reduce the capacity of the
drive.

There are other technical issues involved that I won't bother to address.

The bottom line is: If you could make it work, the drive would be too
expensive for anyone in their right mind to buy. For those needing high
performance, look for a high performance RAID system. (I'm not talking
two drive raid 1 inside a PC either.)

Craigm

 

[The little lost angel]

Not when you throw thermal effects at the problem, unless you could guarantee
dead-matched platter geometry across operating temperature and through
temperature changes.


Well....not etched, of course....but the tracks *do* exist and never shift
location once they are created at the factory.

How is it not affected by the thermal effects then?

I think that paradigm died long ago with the advent of embedded servo data...

Hmm am I correct to understand that due to this embedded servo data,
there is no longer any side used for tracking? hence the tracks
mentioned above are not longer made?

If this is the case, how does the drive guarantee that after thermal
effects, the embedded position data is valid? Unless they store data
for every possible temperature point within the operating range?

Furthermore, if embedding this data will allow the head to seek
correctly, why not embed the data for each platter/head and allow a
potential multifold increase in performance?

Alternatively, since the distance created by thermal effects must be
quite microscopic if not nanoscopic, why not allow each head some
limited degree of adjustment so that while they are all moving on the
same main axle, each has the ability to adjust itself slightly to
compensate for the thermal issue and bring about a potential
multi-fold performance?

 

[Al Dykes]

But the location changes as the platter goes through temp changes, and
if you stack platters they won't all be the same temp at once.

 

[daytripper]

How is it not affected by the thermal effects then?

They are.

Hmm am I correct to understand that due to this embedded servo data,
there is no longer any side used for tracking? hence the tracks
mentioned above are not longer made?

That would be a Roger.

If this is the case, how does the drive guarantee that after thermal
effects, the embedded position data is valid? Unless they store data
for every possible temperature point within the operating range?

Huh? The drive uses the embedded servo fields to physically fine-tune the
position of the head to the track. Other metadata used for relating the
logical location and bounds of each sector.

Furthermore, if embedding this data will allow the head to seek
correctly, why not embed the data for each platter/head and allow a
potential multifold increase in performance?

FOCUS! Again, each track *does* have its own embedded servo data, because you
cannot guarantee a track will be at the exact cylindrical location from
platter to platter, making the Old School concept of a separate servo platter
useless. And unless you put each head for each track on it's own (ie: totally
independent) motion control, it is the thermal effects that precludes trying
to read all of the tracks constituting a cylinder across multiple platters,
because you can't assure that Cylinder X Track 0 is perfectly aligned with
Cylinder X Track n.

Alternatively, since the distance created by thermal effects must be
quite microscopic if not nanoscopic, why not allow each head some
limited degree of adjustment so that while they are all moving on the
same main axle, each has the ability to adjust itself slightly to
compensate for the thermal issue and bring about a potential
multi-fold performance?

Not so microscopic in impact: apparently you've never heard (or even just
heard about) hard drives performing periodic thermal calibration cycles.

At least you're showing some movement - like the quasi-independent adjustable
head positioners you're now allowing ;-) Unfortunately, you're starting down
the road to an exotic drive with much higher seek latency because you've
weighed down the positioner with all the extra mechanical bits to provide
independent adjustment for each individual head. Nice work ;-)

Finally....When there are solutions with *much* higher availability at
commodity price/capacity points based on raid sets - where you can lose a
drive without losing data - exotic, single-point-of-failure solutions are
going to be a wicked hard sell...Which is ultimately why there are precious
few to be found...

Cheers

/daytripper

 

[daytripper]

But the location changes as the platter goes through temp changes, and
if you stack platters they won't all be the same temp at once.

Exactly.

 

[Alexander Grigoriev]

Having the heads aligned on the same cylinder was a BIG problem back in
1990-es. Thermal recalibration, remember? Now that the track density is
quite a few per micron, that's simply impossible.

 

[Alexander Grigoriev]

On Fri, 06 Jan 2006 04:05:15 GMT, [email protected]
(The
little lost angel) wrote:

Finally....When there are solutions with *much* higher availability at
commodity price/capacity points based on raid sets - where you can lose a
drive without losing data - exotic, single-point-of-failure solutions are
going to be a wicked hard sell...Which is ultimately why there are
precious
few to be found...

Cheers

/daytripper

You mean... there are any???

 

[Robert Redelmeier]

How is it not affected by the thermal effects then?

Because the head positioning mechanism compensates. Modern
drives use voice-coils to position heads, NOT floppy-drive
stepper motors.

Hmm am I correct to understand that due to this embedded servo
data, there is no longer any side used for tracking? hence the
tracks mentioned above are not longer made?

Yes. This saves one surface, which is valuable in half-height
drives (means less on some of the 8+ side full-height 5.25") It
also means much less rigid (lighter) head combs for faster seeks.

If this is the case, how does the drive guarantee that after
thermal effects, the embedded position data is valid? Unless
they store data for every possible temperature point within
the operating range?

Because the positioning data (track and sector number) is
effectively part of the track data. The voice-coil needs feedback
from the surface to position itself correctly. It can probably
hit the next track stochastically, but much further away will
need repositioning. Rather like you looking something up in a
telephone book.

Furthermore, if embedding this data will allow the head to
seek correctly, why not embed the data for each platter/head
and allow a potential multifold increase in performance?

Parallelizing heads could be done, but at considerable
cost, and only yield bandwidth gains that serious users
get via RAID anyways. It would not help (and likely hurt)
latency, which is at least as big a performance bottleneck.

Alternatively, since the distance created by thermal
effects must be quite microscopic if not nanoscopic, why

Nope. Thermal effects are _much_ larger than many people
imagine. 3.5" aluminum platters grow 44 um from 20-40'C.
That's about 440 tracks!

If you want to scare a mechanical engineer sometime, tell
her to multiply Young's Modulus of Elasticity by the linear
thermal expansion coefficient. The resulting stress per
degree _is_ scary: Many materials will yield below 200'C
temperature difference.

-- Robert

 

[tony]

Well they used to do that when a single platter only held a few MBs (KBs if
you go back far enough) and spun at 3600rpm - don't recall if they went
slower but likely. With 160GB on a single platter now, you only need
multiple platters for huge drives or if you make a change in form factor.

Or if it made doing the parallel head thing feasible (lower density makes the
positioning thing easier?). Trade some capacity for higher throughput.

It could also be that the more heads you have, the higher the chance of
mechanical failure - IOW aligning say 8 heads with 4 platters is very
difficult(?). The price of the on-board controllers is also orders of
magnitude lower than back in the old days so ganging a bunch separate
drives together in some kind of RAID formation is easy, cheap and
practical.

It would take a lot of drives though to saturate a SATA II bus.

I was at the Samsung site a few days ago and noticed that they are
working on hybrid drives (flash mem + rotating media), but the mem
part of it is only 1 GB so it seems more like a cache of sorts rather
than general storage. I can't imagine that the hybrid drives will be
cheap. If a single "striped" drive could be had for the price of a hybrid,
that may be the acceptable price point to shoot for. Maybe getting multiple
platter drives to produce higher throughput is wasted effort given the following
technology:

I read something about perpendicular magnetic platters. How about
getting that to work in some kind of "parallel" fashion to increase throughput?

Tony