The best hard drives??

[Warner Losh]

On Wed, Nov 18, 2020 at 6:24 PM Bill Gunshannon via cctalk <
cctalk at classiccmp.org> wrote:

> On 11/18/20 8:16 PM, Warner Losh via cctalk wrote:
> > On Wed, Nov 18, 2020 at 6:00 PM Paul Koning via cctalk <
> > cctalk at classiccmp.org> wrote:
> >
> >>
> >>
> >>> On Nov 18, 2020, at 5:56 PM, Chuck Guzis via cctalk <
> >> cctalk at classiccmp.org> wrote:
> >>>
> >>> Tangential to this, I've long wondered about some things relating to
> >>> SSDs.  Are there any solid figures on their retention period after
> years
> >>> of being unpowered?
> >>>
> >>> The reason I ask is that I've long been in the habit of simply shelving
> >>> an old hard drive when I upgrade or replace a system.  I've got hard
> >>> drives that still work that hail back to the days of OS/2 1.1; some
> >>> larger ones go back to the 1970s.
> >>
> >> You should be able to find the answer in the drive specs.
> >>
> >> As I understand it, there are two rather different ranges of answer
> >> depending on whether you're looking at an enterprise class drive, which
> is
> >> optimized for high speed and large total amount of data written, vs. a
> >> consumer drive.  The power-off retention spec is much shorter for the
> >> enterprise drives.  I forgot the numbers; I vaguely remember it being
> less
> >> than a year.
> >>
> >
> > For SSD devices, based on NAND Flash, the specs for retention are 90 days
> > for enterprise drives and 1 year for consumer drives, both at 20C. The
> > difference allows enterprise drives to trade retention for increased
> write
> > rate.
> >
> >
> >> If the drive has power it will do something analogous to DRAM refresh to
> >> keep the bits in good shape.  But it seems that the HDD rule that you
> can
> >> just set a drive on the shelf for a decade (ditto with other magnetic
> >> media) does not necessarily carry over to SSD.
> >>
> >
> > Yes. NAND is just a bunch of small capacitors that decay over time. The
> bit
> > error rate increases following the arrhenius law. The ECC that goes along
> > with NAND is paired to allow NAND that's almost worn out to still retain
> > data for {3 months/1 year} given its expected bit error rate when it's
> > almost worn out when programmed, coupled with the expected decay during
> the
> > specified retention time.
> >
> > Also note I said "at 20C." The acceleration effect can be quite
> pronounced
> > should the data center suffer some catastrophic event that leaves it
> > without power in a super hot environment for weeks or months. At ~70C the
> > acceleration factor can be as high as 30-90x, which can render enterprise
> > drives not reliable after a few days baking at high temperatures.
> >
> > Brand new NAND, on the other hand, typically has retention capabilities
> > measured in years or tens of years. It's the wear and tear of use that
> > makes it less reliable, often much less reliable. And the multi level per
> > cell technologies are much worse than the single level per cell. It's one
> > reason that the smaller number of bits per cell NAND tends to last longer
> > than larger bits per cell, all other things being equal. The smaller
> > process sizes also were less reliable since they could store fewer
> > electrons (sometimes as few as a dozen or two per state). 3D NAND was so
> > much better because it could grow vertically, allowing NAND manufacturers
> > to return to larger process sizes and still increase density, also giving
> > better endurance for a time...
> >
>
> An interesting write-up. Brings up a question on a slightly related
> item.  Do Compact Flash and SD have the same short life when not
> powered?  What things like Flash Memory used to hold firmware on
> other kinds of chips.
>

Yes. They do. It's all the same NAND. However, as I said, freshly made NAND
tends to have very long retention times, so for those use cases, the
application is fine. Unless you are doing a lot of writing to the drive
where the firmware is on, you'll see good results. And a lot usually means
on the order or rewriting the drive every day. Almost all CF (larger than
around 16MB) and SD cards (larger than about 32MB) have wear leveling as
well, which periodically moves the cold OS data around to even out the P/E
cycles the erase blocks across the device. Also, most of the time, the
firmware is in devices that's powered on, so it will move the data should
it decay too much, even when there's not a lot of traffic to the drive
(especially, that's when the FTL loves to do its housekeeping).

'writing often' for NAND in this context is usually measured in 'several
times per day', though with QLC drives, this can be as little as 0.3 or
0.1. The drive writes per day (DWPD) is a spec sheet item these days, and
different levels of drive have differing values clustering around 0.3, 1, 3
and higher...

Warner