On Fri, Feb 10, 2006 at 08:36:28PM -0500, David H. Lynch Jr. wrote:
>...
Firstly I consider swapping (actually paging) a "bad thing" and quite
frequently configure little or no space for that purpose. I'm of the
opinion that a properly designed system shouldn't need it except under
very rare circumstances. Figure roughly how much memory you will
need, then add a good-sized safetly margin. Monitor memory usage
over time when necessary. If a system really does page heaviliy on
a regular basis, performance is likely to be completely unacceptable
anyway. I think it's largely outlived its usefulness.
>The technical information sounds wonderful. But I have 400MB rotating
>disks that are over a decade old and work fine. I lost a 250GB SATA
>drive with well over a 50,000 hour MTBF that failed within a few
>months.
Which illustrates my point, a brand new flash drive won't fail in a
few months like your new 250GB mechanical drive did ;^)
But seriously,
In my experience, the largest number of hard drive failures seem to
occur either early on, or near the end of service life (bathtub curve),
so that's not surprising at all. I suspect the 400MB rotating disks of
which you speak may not have been rotating continuously for that decade,
let alone in your worst-case-scenario of a heavily utilized swap device.
The newer drive probably had a subtle manufacturing defect resulting in
"infant mortality" (unless it was abused in some way).
In any case, the M-Systems drives (which are the only ones I'm
sufficiently familiar with, but there are other similar products out
there) were designed largely for certain military applications where
failure is unacceptable. They do not lose data without significant
warning (if ever) and they don't silently lose bits every so often
like several of the components in most typical consumer PCs do.
Obviously there is no getting around the fact that flash has a limited
lifespan, but if you use high quality components and you spread
write/erase cycles across the full capacity of the device like these
do, the life is extended by orders of magnitude.
To get a general idea of the lifespan there's a calculator here:
http://www.m-sys.com/site/en-US/DeveloperZone/Software/LifespanCalc.htm
I tried punching in the 4GB drive mentioned in the original post, and
calculated that you ought to be able to write to the drive continuously
at its maximum sustained write rate of 12MB/sec for _54 years_ before
failure. Since this is coming from the company selling the things it
could be exaggerated, but even if it is grossly inflated by a factor
of 10, that's still over 5 years. Thus I would have no problem with
putting swap on the drive, though I probably wouldn't do so anyway for
other reasons.
No mechanical hard drive can even pretend to approach 54 years under
ideal conditions, let alone under shock, vibration, heat, altitude. You
can put these in military aircraft, among other things (and they do).
I'd certainly trust my house to that ;^)
>The bad sector re-allocation and SMART technologies do not actually
>address the problem. SMART predict failure, it does not prevent it,
>reallocation does not manufacture new sectors as old ones die.
Actually in this case they do. M-systems FFD, for example, always
has at least 4% spare blocks initially that aren't counted as usable
capacity and are reserved for the above purpose. The ECC corrects
the errors when the data is relocated and SMART tells you when things
are getting to the point where data might soon be in danger. Then if
you still haven't replaced the drive and a threshold is reached that
indicates further write/erase cycles might lead to actual data loss,
the drive automatically goes into read-only mode, preserving the data.
It's not 100% infalable, but it's extremely good. The key is to make
the risk small enough that it's statistically insignificant for the
given application. TrueFFS technology, and other similar methods do
that.
>The relocation technology you are talking about addresses many of the
>problems trying to use it as a disk with an OS on it - but all it does
>is extends the time before a failure. And unless they actually move
>data from infrequently written sectors to frequently written sectors -
>which I would consider to be a seriously dangerous practice...
They do that too, safely and effectively. The software/firmware is quite
sophisticated. It wasn't developed by ameteurs and it's not exactly
version 1.0. Several companies have been using similar methods for
years.
http://www.vmebus-systems.com/pdf/MSystems.Aug05.pdf
All that having been said, I find the noise generated by a traditional
laptop hard drive to be minimal even in relatively close proximity and
they are already fairly low power and low heat. I'm perfectly happy
with laptop drives for the most part, and the Seagate Momentus series
seem to be particularly good in those respects. So the advantages of a
flash drive in a fixed installation, while measurable, aren't
overwhelming. They are vastly more reliable under adverse conditions
though.
|