Re: diskstriping or muliple cachedirs

On Wed 17 Sep, 1997, Markus Storm <storm@uni-paderborn.de> wrote:
>Can anyone comment on issuing 'sync's every few seconds ? How many seconds
>would be a good value ? I have seen running a syncloop greatly increase
>CPU usage :-(

from solaris 2.5 onwards, it doesn't ever perform a sync on all
buffers every thirty seconds (as per a classic 'update' program) but
syncs a proportion every n seconds, and completes 100% within y
seconds (I can't remember the values off the top of my head, but
they're kernel tuneables that you can put into /etc/system)

>BTW, try setting cache_swap_low equal to cache_swap_high, this results in
>equal disk usage over time as opposed to a big hit when cache_swap_high is
>reached.

but you shouldn't actually make them the same: the code, as it stands,
subtracts swap low from swap high, and uses that as a divisor = divide
by zero (which on Solaris you get away with, but it isn't a sensible
value you then use in the calculation.) That should probably be
considered a bug in the Squid code that it does this, of course. (I
haven't noticed that this has been fixed, yet, and it wasn't obvious
to me what the right fix was.)

As to stripe widths, it really does vary with how many spindles and a
whole host of other things (the book I was referring to is a book by
Brian L. Wong, also of Sun Microsystems and the group that Adrian
Cockroft works in.)

Under Solaris you have the disadvantages of the UFS filesystem when
confronted with large directories, so you may do better to treat the
disks as indvidual filesystems anyway.

-- jrg.
Received on Wed Sep 17 1997 - 09:09:37 MDT