Re: diskstriping or muliple cachedirs

Does the stripe size really matter on a web cache disk? I don't think
so, based on the following assumptions:
1. The stripe size is not in any way related to data cluster/block size
for I/O operations.
2. Most objects are small.
3. Disks are faster than the typical network speed used by the clients,
making the total I/O bandwith more important than how fast you can
read/write a single file.
4. Access pattern from squid is quite random... (small objects, hashed)
causing a lot of seeks.

Based on these assumptions I would predict approximately the same
performance on disks striped with any stripe size when the system has
been in use for a time, assuming you have the same number of disks &
controllers in both configurations.

If the disks are only concatenaded, then the disk load will probably be
unbalanced until the disks are filled, then things will even out.

If the disks are separate (multiple cache_dir) then the load should be
even at all times (almost anyway..)

I think that assuption 1 might be false on some platforms, requiring a
small stripe width to not waste I/O when using striping.

So my recommendation is to select a method that allows you to add more
disks to the filesystem as needed. This is generally concatenation, but
might be possible with striping as well (probably requires adding a
whole stripe set, and not only a single disk). It should NOT be
neccecary to wipe the whole cache only to add a new disk.

If your system don't allow you to add more disks without wiping the
filesystem, then there is not much of a point to stripe/concatenate your
disks since the load will be quite even anyway, and it is not that hard
to set up multiple cache dirs compared to building a stripe...

---
Henrik Nordström