Re: developer release 3.1.2
Adrian Cox wrote:
> Adam C Powell IV wrote:
>
> > Why not just include multiple optimized binaries for different arches in
> separate
> > dirs, e.g. /usr/lib/atlas-athlon etc., then use a postinst script to scan
> > /proc/cpuinfo, move the appropriate libs into /usr/lib, and remove the dirs?
>
> This would cause problems in mixed-architecture clusters. Currently you
> can use the same /usr for a PIII and an Athlon, or for a G3 and G4
> PowerPC. I wouldn't like to break that property, because it's very
> useful for machines mounting /usr from a network or cluster filesystem.
Oh yeah, forgot about that. :-)
> I'd much rather have a solution which was decided at boot time or
> dynamically. For instance, an init script which dynamically changed
> links in /etc/alternatives. But I do use the current ability to mount
> /etc read-only, so perhaps we could have a new /var/alternatives
> directory for alternatives selected at boot time?
Hmm, why does it need to "dynamically" change links in /etc/alternatives?
I guess the question is better asked this way: do people NFS-mount /etc?! (It's
just 3 MB on my system!) If yes, then you're right, we'd need
/var/alternatives, as that would have to be a local fs, right? If no, then
installing any package involving /etc/alternatives on such a network involves
hand-creating the links at *install time* (not boot time or run time) anyway, so
what's the difference?
Oh yeah, I guess that wouldn't survive a CPU up/downgrade, so you do need to
select at boot time. Oh well.
> > Does anyone know how Microsoft/Adobe/etc. ships software optimized for the
> latest
> > chips, which doesn't break on old ones? E.g. the Altivec instructions in
> > Photoshop which only are exercized on G4s...
>
> If you're shipping an application, rather than a library, it's trivial
> to decide at runtime which version to load. You can put all the CPU
> dependent parts into shared libraries, and load the correct one when you
> need to.
Aha, and this is a shared library... Well, the static lib would have to be
subarch-independent. So I like your /var/alternatives approach, but yes, it
would require a policy modification.
> This all breaks on a MOSIX cluster with mixed architecture. I guess
> they're just too weird to live.
That's too funny!
Jens Müller wrote:
> ----- Original Message -----
> From: "Adrian Cox" <apc@agelectronics.co.uk>
> To: <debian-beowulf@lists.debian.org>
> Sent: Tuesday, August 01, 2000 5:55 PM
> Subject: Re: developer release 3.1.2
>
> > I'd much rather have a solution which was decided at boot time or
> > dynamically. For instance, an init script which dynamically changed
> > links in /etc/alternatives. But I do use the current ability to mount
> > /etc read-only, so perhaps we could have a new /var/alternatives
> > directory for alternatives selected at boot time?
> >
> Why not decide at start time? Replace the binary by a script which chooses
> the correct binary...
For a binary executable, sure, but the context of the discussion is ATLAS, a
shared library, for which that would be impossible.
-Adam P.
Reply to: