--LyciRD1jyfeSSjG0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline

On Fri, Dec 02, 2005 at 01:37:25PM -0500, der Mouse wrote:
> However, I'm seeing a tremendous performance difference between ld0 and
> ld1.

Yup. Sadly, twe(4)'s only do their "R5 fusion" trick for one ld.

This is the trick that lets them avoid r/m/w cycles without using lots
of nvram cache on the card: write requests that would trigger this are
queued, but deferred and not processed until other writes are
collected to make up a full stripe, or a timeout occurs.

It's the same thing an nvram cache would do, but uses the host's
memory by not pretending the writes have completed.

As you can see, r/m/w's are expensive, and it makes quite a difference
- especially if your fs isn't aligned with the raid stripe size for
data as well as small metadata writes.

--
Dan.
--LyciRD1jyfeSSjG0
Content-Type: application/pgp-signature
Content-Disposition: inline

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2 (NetBSD)

iD8DBQFDkK4MEAVxvV4N66cRAi0cAKDpN24rt2Tf/ts/517d9ZmmPSHi3QCgxo7O
f2mt+N80HJbt7pffgYDuKJg=
=5Ddh
-----END PGP SIGNATURE-----

--LyciRD1jyfeSSjG0--