Subject: Performance on Transmeta Crusoe
To: None <current-users@NetBSD.org>
From: Lasse Kliemann <lasse-list-current-users-netbsd-2004@plastictree.net>
List: current-users
Date: 01/15/2005 16:38:17
--E/DnYTRukya0zdZ1
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable

Greetings,

I am using 2.99.11 on my notebook, a Desknote A530 with a Transmeta Crusoe=
=20
processor. I know that this notebook isn't the fastest one in the world, bu=
t=20
the performance was always acceptable under Linux. However, with NetBSD it=
=20
seems to be significantly slower in some aspects. The Crusoe is clocked aro=
und=20
600 MHz, and I run NetBSD on an 650 MHz Athlon just fine with excelent=20
performance. Hence, I believe it must be some special issue for the Crusoe.

In particular, Mozilla and Firefox (from Pkgsrc) are very, very slow. For o=
ne,=20
the user inteface responds with large delays. Opening a menu can take about=
 3=20
seconds. I first see a black recangle (where the menu is to appear) and the=
n,=20
after some more moments, the menu pops up, finally.

Second, also the loading of webpages only happens with large delays. Browsi=
ng=20
with w3m-img, however, is absolutely fine.

I know that those browsers (Mozilla and Firefox) are large applications, an=
d=20
they do not run absolutely fluently under Linux either. But there is still =
a=20
significant difference. (I'v got two such notebooks and hence was able to=
=20
compare the performance directly: one of them running NetBSD and the other=
=20
Slackware Linux.)

What I did so far is mostly to experiment with compiler optimizations. I=20
compiled all applications from Pkgsrc with -march=3Di686, but that did not =
bring=20
any noticable improvement. I did the same with NetBSD itself with the same=
=20
outcome.  I then realized that the Cursoe is detected at startup as i586-cl=
ass. =20
I so tried to compile NetBSD with -march=3Di585, but the resulting kernel d=
id not=20
run (on none of my machines). I read a note somewhere about i585 being brok=
en=20
in gcc, and then decided to use i486 for NetBSD. There was no performance g=
ain=20
noticable. Well, I haven't tried to compile all applications from Pkgsrc ag=
ain=20
with i486, because I did not believe that is is gonna help. Or is there a=
=20
reason why i386 (what I tried in the very beginning) is much slower than i4=
86=20
optimization? I can try it in my next bulk-build, however.


In the meantime I'd like to collect suggestions from you on the matter. I a=
m=20
unsure what further information about my system to include in this report, =
so=20
feel free to ask for any details that might help shed some light on this. M=
aybe=20
there is some kind of benchmark program that would help to do comparisons o=
r to=20
determine the bottleneck on my system?

Thanks a lot! I'd really love to continue using NetBSD on that notebook, an=
d=20
not to have to go back to Linux.

--=20
Lasse Kliemann
      private homepage: http://plastictree.net
   NO software patents: http://swpat.ffii.org
do NOT use M$ products: http://plastictree.net/articles/noms

--E/DnYTRukya0zdZ1
Content-Type: application/pgp-signature
Content-Disposition: inline

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.6 (NetBSD)

iD8DBQFB6Tjp1gObwed86AkRAtLoAKCiNAWaAvpUZsVE99Up4gk8R2R3cQCgy7Nh
OrhBDkjckDaX0G1NO08OFjY=
=QpCo
-----END PGP SIGNATURE-----

--E/DnYTRukya0zdZ1--