Re: diskstriping or muliple cachedirs

On Mon 15 Sep, 1997, Oskar Pearson <oskar@is.co.za> wrote:
>I would suggest striping the disks at an OS level... I tried it the
>other way for a while, and had hassles because the disk partitions were
>never quite the same size... this meant that I would suddenly run out of
>space, even though one of the disks had a few 100 megs free..

Something else very important is to move swap.log out of the cache_dir
structure. I've tended to put it in the logs/ area (which I usually put on a
totally seperate disk spindle), but certainly moving it out of the cache
directories helps keep them balanced (otherwise you can get a situation similar
to what Oskar describes above.)

If you are going to Raid-0 stripe disks, your stripe/interlace size
(different names with each OS, same concepts) needs to be picked with some
knowledge of the OS and filesystem in mind. With Solaris, you also want to
build the filesystem with a suitable maxcontig value (chunk size =
maxcontig/spindles,).

A book I have (Configuration and Capacity Planning for Solaris Servers)
recommends that maxcontig and the stripe width should be some multiple of 32kB
(maxcontig defaults to 56kB), and goes into a lot more detail than I could
summarise here. It also points out that for random access (which is pretty much
what we get, though we personally see 2/3 write to 1/3 read the majority
of the time: that's with the VM Squid) changing maxcontig isn't critical.

James.
Received on Tue Sep 16 1997 - 15:47:44 MDT