[Gllug] Kernel source on Red Hat

Vincent AE Scott gllug at codex.net
Thu Jun 6 12:58:32 UTC 2002


Richard Cottrill(richard_c at tpg.com.au)@Wed, Jun 05, 2002 at 06:29:30PM +0100:
> > From: gllug-admin at linux.co.uk [mailto:gllug-admin at linux.co.uk]On Behalf
> > Of Dave Jones
> [...]
> > As UML matures, I wouldn't be surprised to see more distros shipping it
> > as an option during installation. Especially in years to come when
> > systems with lots of CPUs get cheaper and cheaper.
> 
> On that note; how does UML perform? I think I read something along the lines
> of slow disc performance and I expect that scheduling must be an interesting
> problem. Is there a way to give UML kernels direct access to some hardware
> (in the same manner that some clusters can share a disc perhaps)?

it does not perform aswell as a native system.  but it's performance is
heavily dependant on having host and UML using /tmpfs.  IO seems to be
its weak point, UML *usually* has a file on the host which contains the
rootfs.  any buffering which is done inside UML is dupliated on the
host.  you can use hostfs where a directory on the host acts as a sort
of chroot / for UML, giving you the benefits of cacheing, but file
ownerships are the same in the host and UML.  so unless you run UML as
root, it can be tricky to use.

scheduling works OK, the UML processes gets treated as user process in
the running host.  so if the host is doing nothing its fast, but if the
host is running numerous other programs, UML will only get a proportion
of the available cycles in the host.

and nope, theres currently no direct hardware access.

 
> Is there a way to force a process to remain on one CPU, or even to dedicate
> a CPU to a process in an SMP environment? I wonder about these things now
> that there are at least two vendors (HP + IBM) I know who have, or soon will
> have, released multi-core chips. Isn't one of AMD's next gen chips supposed
> to be multi-cored? I think it's the one that sounds like a Transformer...

as dj mentioned, processor affinity is the answer.  the patches for
which are bundled with severaljumbo kernel patch sets, -jp and -wolk
spring to mind.

IIRC, UML is being integrated into the 2.5 tree, dave?

-v

-- 
keys:  http://codex.net/pgp/gpg.asc http://codex.net/pgp/pgp.asc

 Cannot find REALITY.SYS. Universe halted.


-- 
Gllug mailing list  -  Gllug at linux.co.uk
http://list.ftech.net/mailman/listinfo/gllug




More information about the GLLUG mailing list