On Wed, Apr 30, 2008 at 01:15:56PM +0100, Dave Ewart wrote:
> > SSD is currently a myth.  They aren't better faster greater etc.
> > Maybe the next generation...
>
> The advantages of SSD are that it is quieter, uses less power and
> generates less heat than a conventional drive.  I don't believe anyone
> promised that the drives would be *faster*, necessarily.  If you want
> quiet and low power, and have no need of large capacity, then SSD is
> fine...

This statement is false.  There exist SSD drives that are not a myth,
and are much much faster than spinning disks.  For example, I have a
few enterprise-grade SSD drives in my lab that have 73GB capacity, can
read (both random and sequential) at about 200 MBytes/s sustained, can
write at about half that speed, and can do about 50 thousand small random
IOs per second.  The comparison for a typical SATA disk about about
90MB/s sustained for read and write, and about 100 small random IOs per
second; so depending on the workload the SSD is anywhere from slightly
faster than the spinning disk (all sequential writes) to 500x faster
than the spinning disk (small random reads).

On Wed, 30 Apr 2008 09:09:30 -0500, Marco Peereboom wrote:
> Sure but you forget to mention the MTBU.  Currently SSD is lower than
> SATA drives.

That statement is also false.  The write endurance of modern, high-
performance SSDs is amazingly high.  I calculated that for an enterprise
SSD, you can write continuously (the interface is saturated with about
100 MBytes/s of writes 24x7x365) for over a year before the flash chips
wear out.  With a sensible workload (duty cycle is not 100%, and the bulk
of the IOs are reads and not writes), the endurance will be much larger,
typically longer than the economic lifespan of the drive (which is about
5 years).  Note that the write endurance of an SSD that uses spare capacity
and wear leveling is quite predictable, and a well-managed drive will
not fail spontaneously, but give ample warning.

Once the write endurance is out of the way, the MTBU is determined by
the usual MTBF.  For spinning drives, the drive manufacturers tend to
quote MTBF numbers of about a million hours; the MTBF actually observed
in the field under good conditions (enterprise-quality disk arrays
and rack-mount enclosures in conditioned data centers, not consumer PCs
with inadequate cooling in residential areas) is about one order of
magnitude lower (for details, see the proceedings of the two most recent
FAST conferences, there has been much academic work on disk reliability
recently).  So we can conservatively assume the MTBF of a spinning SATA
disk to be about 100K hours; under the assumption of a Poisson failure
process, this means that about 1 of each 12 disks will fail in a year.

For comparison: The MTBF of flash-based enterprise-grade SSDs is quoted
by the manufacturer as about 2.5M hours, and industry tests seem to
confirm that number (unlike spinning disks, where industry tests tend
to be an order of magnitude worse than vendor claims).  So high-end SSDs
are probably dozens of times more reliable than spinning disks.

Now, there is the old trifecta: fast, reliable, cheap, pick any two.
High-end enterprise grade SSDs are not available in a bin at your
neighborhood computer parts store, nor do they cost $99.  They tend to be
special order at industrial and enterprise computing vendors, or they are
available in disk subsystems from high-end vendors (EMC, Hitachi, IBM,
the usual suspects).  They tend to be about 10x more expensive than
enterprise-grade (fibre channel or SAS) disks, and per unit capacity
about 30x more expensive than consumer-grade SATA disks.

I am aware that low-end inexpensive consumer-grade SSDs do exist.  Those
are neither fast (some of them have ludicrously low write speeds), nor
very reliable (some of them have low write endurance, comparable to
little CF cards).  On the other hand, they are priced very attractively.

--
Ralph Becker-Szendy    [EMAIL PROTECTED]            (408)395-1435
735 Sunset Ridge Road; Los Gatos, CA 95033

Reply via email to