[Debian-med-packaging] Bug#756392: enable SIMD and threading support

[Elmar Pruesse]

> However, when thinking about, users might be used to call raxmlHPC
> and thus the wrapper should get this name and we could throw away the
> result of Makefile.gcc.  Is this correct or not?

Sounds about right.

> Is there any possibility that the vanilla raxmlHPC might be used
> under some circumstance in practical cases as last resort?

I don't think so, but maybe Alexis knows better than me.

> What about Makefile.AVX2.PTHREADS.gcc?

I had trouble compiling it on some platforms and Alexis said IIRC that
the performance gain was not very high. If it works in Debian, though,
just add it, whether there are 3 or 4 SIMD variants doesn't increase
complexity much.

> Once we are talking about it:  We have some Debian programs were we 
> deliver a "pure" and a "mpi" binary package.  If I understood Elmar 
> correctly the results of
> 
> Makefile.AVX.HYBRID.gcc Makefile.AVX.MPI.gcc 
> Makefile.AVX2.HYBRID.gcc Makefile.AVX2.MPI.gcc 
> Makefile.SSE3.HYBRID.gcc Makefile.SSE3.MPI.gcc

Yes, those would have to depend on openmpi. The detection, whether to
use non-MPI or MPI binaries could probably be based on environment
variables set by mpirun.

AFAIK, PTHREADS splits work along the lengths of the sequences, with
500bp per thread being about the smallest split that is showing gains.
The MPI parallelization splits work along the number of bootstraps.
HYBRID enables both. So choosing the "optimal" binary would require
interpreting the command line options and possibly the input data. Since
I have rarely used RAxML using MPI parallelization, I cannot say much
about this, though.

Elmar
-- 
Dr. Elmar Pruesse

Max Planck Institute for Marine Microbiology
Microbial Genomics Group
Celsiusstrasse 1 | 28359 Bremen | Germany

Jacobs University Bremen gGmbH
School of Engineering & Science - SES
Campus Ring 1 | 28759 Bremen | Germany

Tel: +49 421 2028 984
Fax: +49 421 2028 580

http://www.arb-silva.de
http://www.microbial-genomics.de