[mythtv-users] Multiple directories

f-myth-users at media.mit.edu f-myth-users at media.mit.edu
Wed Nov 21 08:29:04 UTC 2007


    > Date: Tue, 20 Nov 2007 06:52:38 -0500
    > From: "Michael T. Dean" <mtdean at thirdcontact.com>

    > On 11/20/2007 06:27 AM, Nicolas Will wrote:
    > >      4. mv /var/lib/mythtv/recordings/* /srv/mythtv/recordings

    > Too bad this one takes so long.  Perhaps once hard drives get
    > unreasonably large (i.e. Seagate is aiming for 300TB drives by

300Tb, not 300TB, e.g., around 30TB, and I'll bet the first 2010
drives are more like 5-6TB.  There are many misquotes floating
around of a -single- article which seems to have zero independent
verification (though no doubt there are better-sourced figures in
the trade press).  A factor of 30 over the next 3 years is already
enormously larger than the rate at which capacities have been
increasing over the last decade (even given GMR!); a factor of
300 seems unlikely.

[Do a search for ``seagate HAMR'' for more info.  The idea's been
kicking around for at least 5 years (I recall papers in various APS
journals); the basic principle is to get around the superparamagnetic
limit by spot-heating of the medium w/a laser, presumably (I assume;
it's been a while since I've read about it) to decrease the coercivity
of the medium by getting it closer to its Curie temperature.  This is
very reminiscent to me of certain magneto-optical storage systems from
the early 90's; I even have such a drive sitting on a shelf behind me
(though I haven't spun it in years; its capacity was 1GB, back when
1GB of hard disk was a 5 1/4" device that cost $10K, one of which is
sitting on that very same shelf).]

    > 2010--large, even by MythTV standards), manufacturers will stop pushing
    > GB and start pushing performance.

And would you like a pony, too?

(a) The storage interface has continued to increase in speed, from
various generations of ATA, through various generations of SATA, each
one faster than the last.  (And, of course, since long before ATA.)

(b) A single drive in the same form factor will still have the same
number of platters of the same size, with the same number of heads.
Assuming that (a) is still true, then mirroring the drive will still
require exactly one pass over each track.  Densities have been going
up both along the tracks, and in the number of tracks per inch that
can be squeezed in radially.  The former means faster transfer rates
for any given track, but then there's the issue of the latter...

The problem here is that arial density is a quadratic function,
whereas (given roughly constant intertrack stepping speeds) overall
transfer rate is only a linear function.  The only way to beat that
quadratic factor is to start adding heads linearly with the increase
in radial track density, but that's ridiculous---not only would it be
an insane growth rate in heads/surface, you're talking about either (1)
having independent actuators for each head (yeah right) or (2) ganging
them together (more plausible, but -only- useful in the case where
you're mirroring an entire disk track-by-track with no interaction
with the filesystem or any partitioning schemes---and just how often
do users -do- that compared to normal use?).

You -do- get something of a win as the TPI increases, because your
intertrack step size goes down proportionally, which means you don't
have to accelerate the head mechanism as much 'cause you're covering a
smaller distance.  Unfortunately, you're still going to have -some-
step issue and there's gotta be some servo settling there somewhere.
But you've still got to wait for an entire rotation to read an entire
track, so it's a small effect---even if you can go track-to-track
instantly, more tracks total means more rotations total, and there are
very hard physical limits on just how fast you can spin a physical object.
[Be very, very glad that disk sizes have come down as they have, since
the stored kinetic energy of a flywheel goes up as the square of the
radius.  I've seen explosions of old-style "washing machine" disk
drives; one almost killed a colleague of mine.]

You can make small tweaks to the physical system, but they're -all- in
the noise compared to that quadratic factor in arial density---you can
decrease the intertrack step times, or spin faster, or add heads, or
add platters.  All of these get impractical almost immediately:
faster step or faster spin add lots of heat and lots of vibration,
both of which work against high densities; adding heads hugely
increases expense (both the heads, the actuator(s), and the
sense/drive electronics) and cuts down reliability as well [just as
faster step/spin do].  Adding platters is expensive (more heads) and
either makes the drive thicker or makes the platters thinner, but you
don't have infinite space 'cause even if the platters have a thickness
of zero, the heads don't---and thin platters will experience all kinds
of interesting dynamic-wobble effects (map a sine wave around the
circumference of the platter and imagine that's its position in the
z axis), not to mention flying apart and/or just stretching at those
really high spin rates you'd like.  And of course -any- vibration,
internally or externally imposed, will tend to cause gyroscopic
effects on that very-rapidly-rotating set of disks in there...

But for any workload besides drive mirroring, even a little bit of
head seeking will destroy your performance, so all of these heroic
measures are pointless.  And it'd be a rare filesystem where it just
happens that the order in which commands like "mv /mnt/a/* /mnt/b/*"
try to move the files happens to match exactly the way the bits are
laid out on disk (and no, even writing them in order makes no
guarantees that the -drive- has laid them out in the geometry you
thought it was using, and that's before we get into all the crazy
things that -filesystems- might be doing...).

So it sure seems to me that, as long as we're stuck storing data on
spinning surfaces with discrete heads per surface, this quadratic
arial-density term will dominate any other effect.  And if we ever
manage to get some whiz-bang no-moving-parts 3D volumetric storage
(crystalline holographic systems, anyone?), it's only going to get
-worse- (O(n^3) volume instead of O(n^2) area!) unless such a system
can read entire 2D slices at once---reading 1D lines through such a
device would still have a quadratic mismatch in storage vs access,
and reading 0D points (as a disk drive does) would be cubic.


More information about the mythtv-users mailing list