[Gcl-devel] Re: ParGCL/MPI ready to be put into GCL

[Camm Maguire]

Gene Cooperman <address@hidden> writes:

> You also wrote:
> 
> >  My personal preference is for autoloadable modules
> > (see gcl_collectfn.o pulled in by (compiler::emit-fn t), but we will
> > need to do a little work to get the shared libs to come in at the
> > right time too in this case.
> 
> If you have an example of an autoloadable module that pulls in a
> library, I'd be happy to look at it.  However, my current code is assuming
> that the libraries are static (.a).  So, probably the libraries are not
> shared in the sense you mean.  In that case, it may be simpler to statically
> link to the MPI libraries only if the user configured for mpi.
> (So, if --enable-mpi has default "no", MPI libraries are not linked in
> by default.)                                                  - Gene
> 

Greetings!  Agreed that it should work this way for now.  We don't yet
have the facility to load a module, have that dlopen a shared lib, and
preserve the link across image saves, though one contributor's work
(gcl elf-loader) may facilitate this.  This has moved somewhat down my
priority list for the moment, o I think we should go with what you
outline above for now.  Though you need'nt insist on static linking --
-lmpi setup in configure will create a dynamic deendency that should
work just fine.

Have you given any thought to MOSIX?  Might make parallelizing lisp
applications, at least at the coarse grained level much easier.  If
memory serves they recommend predistribution of the processes via mpi.

Take care,

> On Sat, Jul 23, 2005 at 03:11:36AM -0400, Camm Maguire wrote:
> > Greetings!  Please excuse my delay, and thank you *so* much for
> > working on this!
> > 
> > I've been working on stabilizing a checkin of my latest tree before
> > leaving on vacation for 1 week (tomorrow).  I hope to be able to
> > discuss this with you more when I get back.  Preliminary thoughts below.
> > 
> > Gene Cooperman <address@hidden> writes:
> > 
> > > ParGCL (GCL/MPI) now works fine again.  I'm ready to integrate it into
> > > GCL.  If you want to preview the latest version, it's at:
> > >   ftp://ftp.ccs.neu.edu/pub/people/gene/pargcl/pargcl-0.9.3-beta1.tar.gz
> > > 
> > > There are two remaining issues:
> > > 
> > > 1.  Should I install it directly into the cvs, or do you want to do that?
> > >   (I used to have an account on savannah (username: gene), but I hadn't 
> > > used
> > >    it in more than a year, and it's not currently working for me.)
> > > 
> > 
> > I'll be happy to if your access is down, but I'm sure it would be
> > faster if we could restore your access.
> > 
> > > 2.  The ParGCL library has a small footprint (less than 50 KB), and
> > >   includes its own subset of MPI.  Using
> > >      ./configure --with-mpicc=
> > >   one can also use a different MPI.  Currently, I create a separate
> > >   saved-pargcl, so as to guarantee the integrity of the regular saved-gcl 
> > > .
> > 
> > Great!  On Debian, one doesn't need mpicc and can add the mpi libs to
> > the linker command line if they please -- can this be an option too?
> > we should also remember to set *cc*, *ld* and *ld-libs* in compiler::
> > in the new image in case the user wants to compiler::link.
> > 
> > > 
> > >   A.  Should I continue to create a separate saved-pargcl?
> > >       An alternative is:
> > >   Build only saved-gcl, and since it's in a separate mpi package,
> > >           users who don't know or care about MPI won't be disturbed.
> > 
> > I think at present the interested audience will likely be a minority
> > of users (but not for long!), so perhaps we should not disturb the
> > default setup.  My personal preference is for autoloadable modules
> > (see gcl_collectfn.o pulled in by (compiler::emit-fn t), but we will
> > need to do a little work to get the shared libs to come in at the
> > right time too in this case.
> > 
> > > 
> > >   B.  I imagine that I should add to the configure script for GCL a flag
> > >   such as:   ./configure --enable-mpi
> > >        Should the default for --enable-mpi be "yes" or "no"?
> > > 
> > 
> > I'd think 'no" for the time being, but I'm open to discussion if you
> > disagree.  
> > 
> > 
> > > I'm happy to discuss this by phone if you'd like the added bandwidth.
> > 
> > This might be nice when I get back if you still have time!
> > 
> > Take care,
> > 
> > >                                                           - Gene
> > > 
> > > 
> > > 
> > 
> > -- 
> > Camm Maguire                                                address@hidden
> > ==========================================================================
> > "The earth is but one country, and mankind its citizens."  --  Baha'u'llah
> 
> 
> 

-- 
Camm Maguire                                            address@hidden
==========================================================================
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah