John Summerfield wrote:
Alan Cox wrote:
* high-end HDs, e.g. 10,000+ RPM. These can improve your performance
somewhat, in the range of 10-20% - but it stacks on top of
performance gain of alternate filesystems. This is probably the
cheapest upgrade path.
Also more but smaller disks. 8 40 GB disks have much better seek
behaviour than a single 320GB disk. They also take up a lot of room and
power 8(
Have you tested that? It's not clear to me that that should be so.
Physical characteristics determine the number of platters that can be
in a drive, and you generally see a manufacturer release several
drives of different capacity at the same time.
3.5" drives have from one to (I think) five platters.
At any moment, the five-platter version would have five times the
amount of data on they cylinder the heads are on compared with the
one-platter model.
Seek times, when seeks are needed, would be the same. For many
applications, seeks would actually be needed less often.
Comparisons between different models need to take into account the bit
density (which translates to the amount of data per track, and the
number of tracks per cylinder), and _all_ comparisons need to be made
comparing like areas of the recording surface. It used to be the case
that all tracks were the same size, but that's no longer the case:
tracks near the outer rim have more data than the short ones near the
centre.
Unfortunately, the _real_ drive characteristics tend to be hidden; the
cylinders/tracks/sectors information on the labels is what DOS might
see. Even as far back as the 80s, some mainframe drives I knew about
reported physical characteristics to the OS (back then it was MVS/XA)
in terms of what the OS supported, and not the true characteristics
(on those drives, OS could not use the full capacity).
I think what Alan had in mind was probably less about the physical
characteristics of the drive(s) but more about the "random" aspect of
random disk access. With lots of small drives, the load is distributed,
so 2 seeks in a row on one drive (which add up) are run in parallel. In
other words, it's not a single-access performance improvement, but
real-load improvement.
Thinking in terms of Squid, this makes perfect sense..
