[Sluglug] new to the list

[cerise at armory.com]

On Mon, Oct 03, 2005 at 02:37:41PM -0700, Ignacio Solis wrote:
> * cerise at armory.com (cerise at armory.com) said:
> > Either way, the frequency of use would be interesting, but you wouldn't
> > want to put all those files together.  You'd try to spread those out over the
> > platter so that at any one time, you're near a hot file.
> 
> That doesn't make sense. If you want fast access (seek time) you want all the
> files close to you (the head), hence you move them close together.  You'll
> achieve nothing if you spread them out... other than increasing the seek time
> from one file to another.

Where the head is over the disk tends to be a moving target.  If you have 10
frequently used files, it's a reasonable bet that they aren't used by the same 
application.  Further, it's a reasonable bet that they won't be read one after
another.  

Owing to the uncertainty of where the head is over the disk at that point, one
would try to cover as much of a rotation of the disk as possible with data
which are frequently of use.  While it is possible that you will incur a full
rotation around the disk to read data that you just passed, it is equally
as likely that you'll be at the beginning of the data for the next file that
you want.  As a result, it's reasonable to assume that there will be an average
half rotation delay.

If you were to pack them all within the same pie slice of the disk, that gives
a probability of arc-length-of-slice/circumference-of-disk that you'll be in
the right region and 1-(arclength/circumference) that you aren't.  The cost
for that probability is much greater than 1/2 a rotation.

All this assumes that there's no scheduler optimizing access.  A smart I/O
scheduler could handle access to a continuous placement across the platter
much better than putting all important files in one area by sequencing
the reads and writes for each application according to the current position
on the disk.

-Phil/CERisE