--GIP5y49pbaVPin6k
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable

On Wed, Apr 06, 2005 at 10:39:22AM -0700, Konrad Schroder wrote:
> On Wed, 6 Apr 2005, Daniel Carosone wrote:
>=20
> >I *think* due to this change, LFS seems to be taking inordinately more
> >CPU (sys) time than when I was running tests last night.
>=20
> Hm, that's possible...that patch appears to be suboptimal, though without=
=20
> it my test box wedges within a few minutes of multiple simultaneous "tar =
|=20
> tar".=20

FWIW, my testing method was essentially running a bunch of parallel
local rsyncs of pkgsrc into /mnt/[a-z]/, including distfiles.  That,
plus a simultaneous "cd /mnt/a; pax -rwl . /mnt/x" or two to create
hard-link trees within the LFS (and touch all the existing inodes to
up link count), and also a copy of postmark running just for good
measure.  After I had, say 5 or 6 full copies of pkgsrc in there, I'd
rm earlier trees and start a new rsync, to give the cleaner something
to do.

This pretty closely matches my intended workload, too.

This ran quite happily for a couple of hours or so until I
deliberately let the fs fill, and even for perhaps 10-15 mins after
that with postmark complaining about errors writing some files
(ENOSPC, I assume) before it locked up.

> chs dislikes the patch for another reason, and he and I are talking=20
> about how the situation might be improved.  Stay tuned :^)

Shall do; I also intend to test on an SMP LOCKDEBUG machine, though
the above was not.

--
Dan.
--GIP5y49pbaVPin6k
Content-Type: application/pgp-signature
Content-Disposition: inline

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.0 (NetBSD)

iD8DBQFCVFaJEAVxvV4N66cRApNbAKCAs/MhOqM5CcznH6b0TK+7NqcZTACcC20M
LJHK3yWZfoD1jD27Bx/A2L0=
=/xUl
-----END PGP SIGNATURE-----

--GIP5y49pbaVPin6k--