Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
232 Views

Fortran Usage recommendation

I have a couple of options for running a fortran (grid-partitioned; openmpi; OpenSuSE LEAP linux) CFD code: 

1. A single Opteron 6380 (64 cores) box (Dell R815), or...
2. three Xeon x5680 (dual hexacore) boxes connected by infiniband (ConnectX-2 10GbE) directly (no IB switch)

I can choose between gfortran and ifort (I purchased parallel_studio_2019xe...), and also those two machine setups.  Would it be safe to assume ifort would produce the fastest code?  If so, would there be a big difference between the Opteron vs xeon boxes?  I ask that because the Opteron box doesn't need to compile infiniband drivers and I was unable to get infiniband to work with ifort (although IB worked OK with gfortran - with gfortran I just had to compile openmpi using the "--with-verbs" option).

Thank You Very Much!  

Michelle

PS:  I see "impi" in the directory tree of my parallel studio install - but I'm not sure intel MPI comes with parallel studio.  If so, is it significantly better than openmpi on both setups?  I'm sorry to have so many questions - I'm much more a scientist than a sysadmin! 

0 Kudos
5 Replies
Highlighted
Moderator
192 Views

These are good questions and it's really hard to come up with recommendations for you since so often performance is application dependent.

Generally, applications compiled with ifort will run faster.

Considering just the communications between MPI processes single node vs a cluster, that is faster on a single node, but also depends on CPU and memory use.

Yes, you do see pieces of MPI in the Composer edition (just compilers) because a Fortran feature called "coarrays" is implemented using MPI. It is not a complete MPI package. You can get Intel's MPI in the Cluster edition.


0 Kudos
Highlighted
Moderator
187 Views

Did you write this CFD application or someone else? If it's someone else, you might check with that person for recommendations.



0 Kudos
Highlighted
Valued Contributor III
167 Views

@sheaffer__patti ,

You may also want to review MPI site (https://www.mpi.org/) and resources there such as Benchmarks, Fortran options, etc. and also ask on their Community forum(s).

With Fortran, you can also consider

  1. https://fortran-lang.org/ and the community there some of whom are very familiar with MPI and other aspects of the so-called high-performance computing (HPC): https://fortran-lang.discourse.group/ and also the Fortran standard Coarray facilities which in some cases are built upon MPI,
  2. Also Fortran Wiki at http://fortranwiki.org/fortran/show/HomePage
  3. and comp.lang.fortran newsgroup

Focusing on benchmarks and testing with your own specific needs with your CFD code can guide you best i.e., you may "trust" the information online, but do "verify" it yourself.  And which will be just being a  good scientist!

Highlighted
New Contributor I
128 Views

I have not done work on a cluster for more than 10 years (I had an IBM blade cluster with 32 units running Red Hat). But I have done a project (continuum mechanics)  in the last few years with Intel Fortran and Microsoft compilers (C++, C#). Other than installation issues, I think Intel compilers, including FORTRAN, are the best around I tried both on x86 and 64 (I7 and Xeon), the resulting code generated by FORTRAN is extremely stable and real, real fast. You may find also limited multi-language support very useful like I did.

The small group of people who participate in this forum are exceptionally good and very helpful (and patient).

I. Konuk

Highlighted
New Contributor II
125 Views

The small group of people who participate in this forum are exceptionally good and very helpful (and patient).

Perhaps with one exception, I merely tag along and have some fun -- it is like reading Asimov or Heinlein

I agree with writing your own code the Commercial versions ones leave me cold and you can never solve the problem you want.  

0 Kudos