Coarray article for the upcoming GCC Summit.

[jdelia@intec.unl.edu.ar]

Hello there,

Tobias Burnus <burnus@net-b.de> ha escrito:

> Bill Long wrote:
>> 8) In section 7 on implementation,  I would probably lean towards   
>> the general implementation in terms of GASNet rather than MPI.     
>> GASNet is generally a shorter path to the hardware, and hence more   
>> efficient.  Perhaps more important, the GNU version of UPC is done   
>> with GASNet (at least that's my impression), and reusing work   
>> already done within the GNU umbrella seems like a good idea.  In   
>> fact, pointing out that much of the work done for UPC can be reused  
>>  for Fortran is probably a good selling point.
>
> I was also wondering the other day whether UPC could help (see GCC's
> bugzilla PR18918). However, I have the impression that since 2006
> nothing has happend and UPC is not part of the GCC itself. I actually
> do not know  whether they have signed a copyright assignment with the
> FSF. If not, we cannot take anything from UPC.
>
> At least the GCC UPC I found was based on GCC 4.0.x which is already
> slightly outdated:
> http://www.intrepid.com/upc.html
>
> Regarding GASNet (http://gasnet.cs.berkeley.edu/), this would be indeed
> an alternative (as I wrote in PR18918). The advantage of MPI is that is
> ubiquitous. However, GASNet seems to be faster and seems to support
> also MPI as network backend (as TCP/IP, infiniband, myrinet, ....). [By
> the way, GASNet is under a BSD-like licence.]

What about an ARMCI alternative? For example, the CAFC traductor  
(http://www.hipersoft.rice.edu/caf/download.html) can be built with  
GASNet (http://gasnet.cs.berkeley.edu/) or with ARMCI  
(http://www.emsl.pnl.gov/docs/parsoft/armci/), but not both, and in  
each case, with or without a MPI option. A MPI conduit could be  
interesting in a Bewoulf cluster.

> Has anyone experience with GCC UPC or knows their licence status
> (agreement with FSF yes/no) or whether they indent to get it merged
> with standard GCC?

I have have little experience with the GCC UPC  
(http://www.intrepid.com/upc.html) and the Berkeley UPC  
(http://upc.lbl.gov/) compilers. The GCC UPC works well but only  
multiprocessor machines while the second one is perhaps more  
interesting for a Beowulf cluster since it can built with several  
low-level network APIs including MPI.

>> I think that having coarrays available in gfortran would be a   
>> tremendous plus for the Fortran user community.
>
> So do I, although I would expect that it takes quite a while until
> coarrays are wider used. (At least the Fortran programmers I know
> strive for large compiler compatibility (Fortran 95 without allocatable
> TR, maybe even without some new F95 features). Though with conditional
> compiling some Fortran 2003 features are creeping in.)
>
> Tobias

I think as Bill, that is, that having coarrays available in gfortran  
would be a tremendous plus for the Fortran user community, specially  
for the new generation of Fortran programmers that work in parallel  
computing. BTW, in my work, the PETSc library is commonly employed  
(http://www-unix.mcs.anl.gov/petsc/petsc-as/) .

Regards,
Jorge.

----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.