[Beowulf] GPUs Nvidia C2050 w/OpenMP 4.5 in cluster

[Peter Kjellström]

On Mon, 12 Aug 2019 18:06:01 +0300
"Mikhail Kuzminsky" <kus at free.net> wrote:

> Heterogeneous nodes in my small CentOS 7 cluster have  x86-64 CPUs 
> along with the old Nvidia GPU C2050 (Fermi). New Fortran program uses 
> MPI + OpenMP software.

There are several challenges here. First the rather old C2050. It is if
I remember compute capability 2.0 (current is turing at 7.5...). It is
also rather slow even if it can run the code. To top it off I'd expect
some challenges getting a working driver+cuda that isn't horribly old
(maybe cuda8 would work, ymmv..).

> Does the modern gfortran or Intel ifort compilers give support of
> work through OpenMP 4.5 with these GPUs?

Second challenge, state of compiler support.

Intel ifort will (not surprisingly) not generate OMP TARGET code for
Nvidia GPUs...

GCC/gfortran has some support for OpenMP and OpenACC via nvptx (see for
example: https://gcc.gnu.org/wiki/Offloading) but you may want to run
9.2 or so.

Bottom line: old cuda + old hardware + ??driver + nvptx + bleeding edge
gcc + fortran -> not very likely usable imo.

/Peter K