Last week's "Windows Server 2003 better than Linux" test

[David Collier-Brown -- Customer Engineering]

jra at dp.samba.org wrote:
> Hmmm. To be honest I think this is due to the Solaris UFS 
> behavour - it seems to be very sensitive to write sizes
> (that's why I added the write cache code). Linux doesn't
> seem to need this as the kernel does the buffering.

   Yes, I can confirm that in part: up to 8 update 3 UFS (actually
an updated ffs, not the old old ufs) wasn't particularly good at
coalescing writes.
   That changed a lot about then, and I found in traces that
ufs was putting together some very nice large chunks. However,
that mostly avoids small fragmented writes and excess seeks.
Handing ufs multiples of the (block size * stripe width) is
still a **very** good thing on a filesystem residing on
a RAID array.
   On a brand new filesystem, unrelated to ufs, the same is
still true: RAID and big buffers beats ufs and small ones
by ten to one.
   And it applies to read as well: reading large amounts
gets a lot of parallelism on the initial reads, and uses
the readahead well thereafter.

[Just FYI: I've changed my strategy so that samba reading from
a readahead cache so as never to give a "short read", and
the code got simpler...]


   A question for the Linux aficionados: can the OS get
knowledge of the "best size" for RAID devices from the device?
That sound like something a filesystem might be able to know
and use...  albeit not ufs (;-))

--dave
-- 
David Collier-Brown,           | Always do right. This will gratify
Sun Microsystems DCMO          | some people and astonish the rest.
Toronto, Ontario               |
(905) 415-2849 or x52849       | davecb at canada.sun.com