Whither "Hybrid" Disk Drives?

Whither "Hybrid" Disk Drives?

Post by Jon Forres » Thu, 05 Oct 2006 03:05:19


Microsoft is making a big deal about a new kind of disk
drive that combines a regular disk with some amount
of flash memory in the same package. (See
http://www.yqcomputer.com/ ,1697,1901955,00.asp)
for a presentation of this idea.

Except for speeding up booting and resuming,
I don't see why this is better than just adding
more system RAM memory, especially in OSs like
Windows and *nix where memory not needed for
processes is used as a disk cache.

The article mentions reducing the amount of time
a disk needs to spin, in order to reduce power conception,
but couldn't this be done just as easily by keeping data
in the OS cache? (I do recognize that flash memory
keeps its contents when the power goes off, but this
doesn't appear to me the major selling point of
hybrid disks in the articles I've seen).

The key issue I see is whether the OS or the disk controller
has a better conception of what to cache. Plus, what
happens, in worse case, if their conceptions conflict?
It's easy to imagine a case where the same data is
cached in both the OS cache and the on-disk cache.
Maybe a smart disk driver can prevent this from happening
but this whole approach seems anti-intuitive to poor me.

Where's the beef?

Jon Forrest
 
 
 

Whither "Hybrid" Disk Drives?

Post by Casper H.S » Thu, 05 Oct 2006 03:32:37

Jon Forrest < XXXX@XXXXX.COM > writes:


The main advantage is that a disk-flash cache can cache
writes with impunity; a OS cache cannot.

I'm not sure how this will help with faster booting unless you
fix a piece of the cache as "boot memory" which will make
it less effective in saving power.

(Considering that disks can easily deliver 60MB/s faster booting
can easily be achieved by making sure that the boot memory is
a sequential blob on disk so the amount needed to boot can
be delivered in a handful of seconds)

Casper
--
Expressed in this posting are my opinions. They are in no way related
to opinions held by my employer, Sun Microsystems.
Statements on Sun products included here are not gospel and may
be fiction rather than truth.

 
 
 

Whither "Hybrid" Disk Drives?

Post by Ketil Mald » Thu, 05 Oct 2006 15:42:26

Jon Forrest < XXXX@XXXXX.COM > writes:


I guess MS has an OS they want to sell. That aside:

Jon Forrest < XXXX@XXXXX.COM > writes:



As Casper points out, flash isn't much better - and perhaps worse? -
than the disk for transfer. Current OS'es seem to use a lot of time
to resume, (populating 1Gb should only take 20 seconds or so, but
seems to take a lot longer in practice), which I presume is due to the
stored state being fragmented on the disk - since flash eliminates
seek times, perhaps the idea is just to avoid fixing the OS in this
regard?


And/or more regular on-disk cache?

As flash is becoming cheaper than RAM, and you don't need outrageous
bandwidth anyway, perhaps this is just a way to increase on-disk cache
cheaply?


If we're talking about laptops, they already have a battery, too.
Perhaps some applications really need to write synchronously to disk
(mount without noatime), and this is a way to fix that?


Another option is to add a separate nvram disk to the system. Flash
these days seems to be cheaper than ... well, just about anything
else, really. How could an OS make efficient use of a GB or two of
'system flash'?

-k
--
If I haven't seen further, it is by standing in the footprints of giants
 
 
 

Whither "Hybrid" Disk Drives?

Post by Stephen Fu » Fri, 06 Oct 2006 00:56:54


Think laptops, where battery life is a significant selling point. Having
the disk spin less means less drain on the battery, hence longer battery
life. That is the main purpose of the feature.

--
- Stephen Fuld
e-mail address disguised to prevent spam
 
 
 

Whither "Hybrid" Disk Drives?

Post by Stephen Sp » Fri, 06 Oct 2006 02:56:52


Sequential read and writes are slower, but the zero seek time gives
flash a phenomenal advantage for non-sequential reads.


Restoring a hiberating PC is indeed faster with flash, but the real win
comes in normal boot times. I just read a review the other day of a
36GB ATA flash drive, and the system booted in under half the time that
it did with a conventional disk. That's big.

The flash drive also stomped the best rotating drive on some workloads
(e.g. web server), and it beat the merely "average" drives on nearly
every test.


The idea behind this is that flash can be an intermediate cache layer
between RAM and the HD. They're even trying to push USB flash drives as
cache devices.

Personally, I'd invest in more RAM before I'd add a flash cache, but
perhaps that's just general resistance to any "great idea" MS comes up
with.

S

--
Stephen Sprunk "God does not play dice." --Albert Einstein
CCIE #3723 "God is an inveterate gambler, and He throws the
K5SSS dice at every possible opportunity." --Stephen Hawking



--
Posted via a free Usenet account from http://www.yqcomputer.com/
 
 
 

Whither "Hybrid" Disk Drives?

Post by Jon Forres » Fri, 06 Oct 2006 03:01:19


I thought I had mentioned this in my original posting

Again, if the goal is to reduce the amount of time a disk spins,
what difference does it make if the cache is on the disk
as opposed to on the motherboard? In either case, I/Os will still
be avoided.

Jon
 
 
 

Whither "Hybrid" Disk Drives?

Post by Rick Jone » Fri, 06 Oct 2006 03:47:45


With flash I suppose the flush can be "whenever" whereas if one wants
to power-down the host, cache in RAM would need to be flushed. Also,
which draws more power - RAM in the host or flash in the pan/disc?

rick jones
--
Process shall set you free from the need for rational thought.
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to post, OR email to rick.jones2 in hp.com but NOT BOTH...
 
 
 

Whither "Hybrid" Disk Drives?

Post by Stephen Fu » Fri, 06 Oct 2006 04:48:04


If you are talking RAM on the motherboard, then the difference is the
non-volatility. This is important when powering up for a reboot. Any data
stored in the RAM on the motherboard would have been lost and must be
reloaded so you must power up the disk to do that. If you are talking flash
on the motherboard as opposed to the disk, there is the issue of not having
the OS spend cycles manageing the cache, and I would guess that Seagate went
to Microsoft and said "we could do this, let's work together" but no laptop
motherboard manufacturer did so. Business. :-)

--
- Stephen Fuld
e-mail address disguised to prevent spam
 
 
 

Whither "Hybrid" Disk Drives?

Post by Tim Bradsh » Fri, 06 Oct 2006 05:48:00

On 2006-10-04 19:47:45 +0100, Rick Jones < XXXX@XXXXX.COM > said:


Well, flash has (or can have, since it's nonvolatile) zero power
requirements if it is not being read or written...

--tim
 
 
 

Whither "Hybrid" Disk Drives?

Post by Bill Tod » Fri, 06 Oct 2006 06:51:21


As has been observed by others, there's no reason why such bulk-transfer
activities should take longer from disk than from flash - as long as
they've been optimized to execute from the disk in streaming sequential
fashion (which unfortunately does not usually seem to have been done in
current Windows environments - and perhaps others as well, FAIK - so the
flash becomes a potentially useful feature only due to the uncorrected
deficiencies of the OS).


But it's *related to* the major selling point, which is that spinning up
the disk becomes a rare occurrence, hence saves significant power on a
laptop.

Laptop activities tend to save data to the disk fairly often, and the
fact that the drive may have its write-back cache enabled does not keep
it from spinning up each time (if it had been spun down - and with the
large amounts of system RAM laptops sport these days it likely otherwise
could have been), because the write-back cache is destaged fairly
immediately (unless the disk is otherwise busy) under the assumption
that if an application deliberately wrote to it it probably had good
reason for wanting that data to be persistent (otherwise, it would still
be sitting in the system's RAM write-back cache area).

Place a NV flash cache in front of the disk and it need be destaged -
well, never, unless its dirty contents overflows the area alloted for it
(and the kinds of things that tend to be saved to disk on a laptop tend
to be smallish, so that shouldn't happen often).

My main reservations about such behavior would center around whether
flash memory has reliability - especially when frequently updated in
small increments - comparable to (or better than) data on the disk
platters: if not, such write-back policies would compromise mean time
to data loss, but the average laptop user may not worry much about such
issues (though if the drive manufacturers want to convince enterprise
users that this is a great way to obtain cheap NV write-back cache in
their environments, they will have to be *very* convincing on this point).

- bill
 
 
 

Whither "Hybrid" Disk Drives?

Post by Ketil Mald » Sat, 07 Oct 2006 20:35:17

"Stephen Sprunk" < XXXX@XXXXX.COM > writes:


An iPAQ with Linux can apparently suspend to RAM and resume in 10ms,
and the OLTP laptop hopes to get close to that, if I interpret the
article correctly.

I guess the context here was suspend to disk, so perhaps it isn't
terribly relevant, but I thought I'd mention it anyway. The URL is

http://www.yqcomputer.com/

-k
--
If I haven't seen further, it is by standing in the footprints of giants
 
 
 

Whither "Hybrid" Disk Drives?

Post by Casper H.S » Sat, 07 Oct 2006 21:09:19

Ketil Malde <ketil+ XXXX@XXXXX.COM > writes:


Big deal; suspend to RAM is easy and requires powered RAM.

Suspend to disk is harder as you effectively have to recover from
a hardware power-cycle.

Casper
--
Expressed in this posting are my opinions. They are in no way related
to opinions held by my employer, Sun Microsystems.
Statements on Sun products included here are not gospel and may
be fiction rather than truth.
 
 
 

Whither "Hybrid" Disk Drives?

Post by JJ » Sun, 08 Oct 2006 08:06:07

Stephen Sprunk wrote:

Atleast until recently, Flash has been sort of a throwaway thing for
keychains ie 4x256MB for $20 or so, but the sizes are now ramping up so
fast that speed or lack of became obvious, it had to be fixed. 1MB/s to
10MB/s was good but we want more, to the limit of the connection
please!.

I was told recently by one vendor of Flash Drives and USB Flash sticks
(Apacer) that all devices will be hitting 30MB/sec pretty soon, and are
available already (I haven't seen those yet). They had them in the
booth, everyone is using Samsung too. So far we have seen the Toms?
article on the 30GB 30MB/s sub 2.5" drive but the same is likely to
happen with pen drives just much smaller cheaper. Personally I can't
wait till 30MB/s is the norm then maybe run mostly off that, leave the
spinning media for media files and backing the Flash drive.


The ability to reload quickly seems to be poorly done in most current
OS software and I would hope one day that Flash will take away all
excuses the software guys put up for slow boot times. SInce most of the
DRAM doesn't have to be reloaded, it should only take a few secs max to
reload working OS. Bloatware like Firefox seems to use way more ram
than this OS here.


It seems for the longest time since the 80s NV storage, eprom, eeprom
and early Flash was way more expensive than DRAM per bit and that
lasted till the mass arrival of digital cameras and MP3 players which
demanded flash card or tiny spinning media. That allows flash to be on
a much faster density growth track than DRAM even from the same vendor.
I get the impression that 16Gbit flash chips are in same $ ballpark as
1GB DRAM parts so flash has won for now, flash prices plumet while DRAM
appears to head back up again.

John Jakson

 
 
 

Whither "Hybrid" Disk Drives?

Post by anto » Sun, 08 Oct 2006 16:40:43

Bill Todd < XXXX@XXXXX.COM > writes:

I would expect that the disk has to spin up more often for reading
stuff that is not yet in RAM, and then the writes can be destaged, so
spinning up for writing should be extremely rare.


Any reason to think that this is not the case? If you are thinking
about parts of the flash failing to be writable because of too many
writes, that should be recognizable by reading after writing, and the
part would be marked as unusable. Also, I can think of a number of
reasons that make such events rather rare.

Or are you thinking about the flash memory losing the contents
afterwards; that would be a problem.

- anton
--
M. Anton Ertl Some things have to be seen to be believed
XXXX@XXXXX.COM Most things have to be believed to be seen
http://www.yqcomputer.com/
 
 
 

Whither "Hybrid" Disk Drives?

Post by Bill Tod » Sun, 08 Oct 2006 22:49:39


I question that, with laptop RAM complements rapidly approaching 1 GB
theses days: typical use should result in declining disk-read rates
over time as less and less cannot be found in the system RAM.

Then again, I tend to save changes on a laptop as frequently as I do on
my desktop - in perhaps both cases an increasingly obsolete habit, given
my desktop UPS and the laptop's battery plus the improving stability of
Windows (not that my family benefits much from the last, since we still
use Win98 for applications that support it: it's fast enough, stable
enough, and behind the router and local firewalls secure enough that
there's just no compelling reason to move).


Only if the disk is willing to wait a while to destage dirty data in its
write cache - which it really shouldn't be, since that decision is the
job of the OS-level write-back caching mechanism.

I.e., once the OS thinks it's time to write data to disk (e.g., because
a user explicitly 'saved' it), the disk should get it onto its platters
as soon as is consistent with short-term performance optimization goals
(such as reordering it or, in the case of drives that don't support
large sequential transfers, aggregating it) rather than second-guess
higher-level software as to how important making it persistent may be.

so

Only if typical use patterns differ significantly from what I've assumed
above - which they well may, but I'd want to see actual evidence of that.


My impression is that the probability of data loss from UPS-backed
conventional ECC RAM may be somewhat higher than that of data loss from
a disk (at least if one ignores whole-disk failure, which is appropriate
in this situation), but I could be mistaken or if not flash might differ
in that respect.

If you are thinking

That was what I was referring to in the 'updated in small increments'
portion.


It *should* be, but is this actually done?

and the

They need to be 'rather rare' in comparison to other modes of disk data
loss - which themselves are quite rare.


I'm not at all familiar with flash failure modes, which is why I phrased
my reservations conditionally.

- bill