Re: [gentoo-cluster] Switchable MPI

[Donnie Berkholz]

Andrew D. Fant wrote:
> I'll dig around and forward the ebuild to the list later tonight or
> tomorrow. At the risk of going off on a rant, I think that the
> /etc/env.d model is a core part of the problem.   If two users want to
> use two different mpi implementations, will eselect allow them to both
> use them at the same time, or will it have to be one or the other at
> once?   As I see it, something will have to be tweaking $LD_LIBRARY_PATH
> for each user's environment,

Yep, that's the plan. Some things just won't work as well because they
don't really respect variables, like the includes.

Something like this:

MPI_IMPLEMENTATION="$implementation"
PATH="/usr/lib/mpi/$implementation/$version/$compiler/bin:$PATH"
LD_LIBRARY_PATH="/usr/lib/mpi/$implementation/$version/$compiler/lib:$LD_LIBRARY_PATH"
CPPFLAGS="-I/usr/lib/mpi/$implementation/$version/$compiler/include:$CPPFLAGS"
MANPATH="/usr/lib/mpi/$implementation/$version/$compiler/share/man:$MANPATH"
INFOPATH="/usr/lib/mpi/$implementation/$version/$compiler/share/info:$INFOPATH"

> or /etc/ld.so.cache will have to be changed
> for each user's login to reflect their prefered mpi environment.

That totally screws with multiple simultaneous logins. Environment
variables will have to be the way to go.

> $PATH
> and $MANPATH can be handled by judicious checks in /etc/csh.cshrc and
> /etc/profile, but since /etc/env.d insists on manipulating the shared
> object cache directly, this has the potential to get really ugly without
> some cooperation from the core system team.  I don't think that the
> cluster herd can fix it on it's own.

I think we can, but I could certainly be proven wrong on this.

> I would love to be proven wrong and will gladly defer to anyone who does
> so,  I just know that I've been trying to outthink this problem for a
> year or so on our cluster and haven't come up with a sane solution for a
> true multi-user environment.

I would love to find a way that works too. =)

Thanks,
Donnie