[Llvm-bgq-discuss] pulling in mpi using clang

[Jeff Hammond]

No, I have never explored this.  If you use the script I put on
https://wiki.alcf.anl.gov/parts/index.php/MPICH2_on_Blue_Gene/Q#Building_BGQ-MPI_from_source
you can trivially build MPICH-BGQ with LLVM (at least V1R1M2) and see
how it does.

Jeff

On Fri, Feb 1, 2013 at 11:34 PM, Michael Blocksome <blocksom at us.ibm.com> wrote:
> Do you guys have any benchmark numbers for the bgclang version of mpich -
> and how it compares to the bgsys xl and bgsys gcc versions of mpich?  Just
> curious ...
>
> Michael Blocksome
> Blue Gene Messaging
> blocksom at us.ibm.com
>
>
>
>
> From:        Jeff Hammond <jhammond at alcf.anl.gov>
> To:        Geoffrey Irving <irving at naml.us>,
> Cc:        llvm-bgq-discuss at lists.alcf.anl.gov
> Date:        02/01/2013 12:22 AM
> Subject:        Re: [Llvm-bgq-discuss] pulling in mpi using clang
> Sent by:        llvm-bgq-discuss-bounces at lists.alcf.anl.gov
> ________________________________
>
>
>
>> What's the proper way to get mpi when using
>> /home/projects/llvm/current/bin/bgclang++?
>
> I thought Hal made wrappers at some point.  Maybe on another machine.
>
>> Previously, I stole the flags from the gcc mpi wrapper (I unfortunately do
>> not remember how I
>> did this), and then manually specifiied
>
> You can use mpicxx wrapper with MPICH_CXX=bgclang++ instead to get the
> same thing with less typing.
>
>>     -I/bgsys/drivers/V1R1M2/ppc64/comm/sys/include
>> -I/bgsys/drivers/V1R1M2/ppc64
>> -I/bgsys/drivers/V1R1M2/ppc64/spi/include
>> -I/bgsys/drivers/V1R1M2/ppc64/spi/include/kernel/cnk
>> -I/bgsys/drivers/V1R1M2/ppc64/comm/gcc/include
>>     -Wl,-rpath=/bgsys/drivers/V1R1M2/ppc64/comm/gcc/lib
>> -Wl,-rpath=/bgsys/drivers/V1R1M2/ppc64/comm/sys/lib
>> -Wl,-rpath=/bgsys/drivers/V1R1M2/ppc64/spi/lib
>> -L/bgsys/drivers/V1R1M2/ppc64/comm/gcc/lib
>> -L/bgsys/drivers/V1R1M2/ppc64/comm/sys/lib
>> -L/bgsys/drivers/V1R1M2/ppc64/spi/lib -lmpich -lopa -lmpl -lrt -ldl
>> -lpami -lSPI -lSPI_cnk -lpthread
>>
>> This seems fairly sketchy, so I'd like to do it the right way if such
>> exists, especially if the driver versions are going to change soon.
>
> I don't think it is sketchy.  It's what I've been doing for all my
> LLVM codes.  LAMMPS (C++ molecular dynamics code using MPI C
> interface) had no issues with this recipe.
>
>> Are there wrapper compilers set up?
>
> As noted above, I don't know.
>
> I can build you a bona fide LLVM MPICH though if you think it would
> help.  I did this a long time ago and just need to update it.
>
>> If I do the above manual way, and qsub a job, it errors out during launch
>> with
>
>> VST-20400-31511-32:14558:ibm.runjob.client.Job: Load failed on
>> Q02-I4-J04: Application executable ELF header contains invalid value,
>> errno 8 Exec format error
>
>> It's possible this is unrelated, of course.
>
> Looks like an LLVM-related issue but not an MPI one, at least not directly.
>
> Jeff
>
> --
> Jeff Hammond
> Argonne Leadership Computing Facility
> University of Chicago Computation Institute
> jhammond at alcf.anl.gov / (630) 252-5381
> http://www.linkedin.com/in/jeffhammond
> https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond
>
> _______________________________________________
> llvm-bgq-discuss mailing list
> llvm-bgq-discuss at lists.alcf.anl.gov
> https://lists.alcf.anl.gov/mailman/listinfo/llvm-bgq-discuss
>
>



-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond