[Pw_forum] Off-topic: parallel compilation

[Pang Rui]

Dear Marcos
I have performed a 77-atom B3LYP calculation using VASP on 36 cores. It
crushed and showed segmentation fault. Then I tried on a machine with
larger memory, it worked, but spent 3 days finishing a loop, so I quit.
In my opinion, the hybrid function calculation is quite time consuming and
memory demanding in plane wave basis. It is better to do such calculation
on a localized basis code such as Gaussian. 
On Sun, 18 May 2014 11:10:59 -0300, Marcos Veríssimo Alves
<marcos.verissimo.alves at gmail.com> wrote:
> Hi all,
> 
> This is a bit off-topic, but I'm sure there will be many knowledgeable
> people around. It's related to the parallel compilation of QE.
> 
> I'm trying to run HSE calculations for a 71-atom silicene sheet, using
> Espresso compiled with ifort (composerxe), mpich2 and intel mkl included
in
> this version of the compiler. Mpich 2 is compiled using gcc/ifort.
> 
> All goes well in the scf part until it gets to the hybrid functional
> calculation. Then it crashes with an mpi error (Bad termination, code
> 9) which the mpi faq says is usually due to the fact that there is a
> segmentation fault in the code, not in mpi.  They also state that this
is
> usually a segmentation fault.
> 
> I wonder if anyone has run HSE calculations for systems as large as
this,
> and if they have run into this problem.
> 
> One possibility that came to my mind is that, if this is a segmentation
> fault, maybe 4-byte integers are not large enough to address all the
> indexes of the arrays allocated in the hybrid functional part of the
> calculation. I tried to compile QE with 8-byte integers but I get an mpi
> error, which I suppose is due to integers being compiled with 4-byte
> integers on the C side of the compilation.
> 
> So, I have two questions:
> 
> 1) Is there any way of compiling Mpich 2 with 8-byte integers using gcc?
> 
> 2) If the parallel part were carried out with OpenMP instead of Mpich (I
> have 32 cores in the same node of the computer I'm running the
> calculation), perhaps this could be a workaround for this problem. Is it
> possible to compile QE in parallel mode using only OpenMP and MKL? If
yes,
> what options would need to be set for either configure or make.sys
> directly? Configure already detects the mpi environment, and sets
make.sys
> to compile QE using it.
> 
> Any help would be greatly appreciated, and I would gladly supply the
input
> file along with the pseudo I am using (from the QE database) if the
> developers would like to run tests on it.
> 
> Best regards,
> 
> Marcos

-- 
Department of Physics, South University of Science and Technology of China