ABINIT Discussion Forums

Hi.

I'm not a user myself, but am in charge of updating the version of abinit on a Cray XE6 system. It all seems to install fine, but on running some of the test cases one of our users has reported problems. The case from tutorial 3 with the input file "tbase3_5.in" is the one I've been focusing my efforts on trying to get working.

I've been testing it in parallel when using 1-32 processors. All work except for 16, 17, 19 31 or 32. They all ran fine in our previous version of 7.0.3.

The problem seems to be to do with the checking of the FFT grids in hdr_check(). For instance at 32 processors two processors are reporting differences between 20 and 24 (lines 284-294). The other processors are all happy. There were warnings about

--------------------------------------------------------------------------------
Subroutine Unknown:0:WARNING
The second and third dimension of the FFT grid, 0 0 were imposed to be multiple of the number of
processors for the FFT, 3
For input ecut= 8.000000E+00 best grid ngfft= 20 24 24
max ecut= 9.454813E+00
However, must be changed due to symmetry => 24 24 24
with max ecut= 0.136149E+02
--------------------------------------------------------------------------------

but they seem to be there in the cases that work too.

I have no idea if this is a bug with ur installation and these processor numbers for this case work elsewhere, or if this is a case which should no longer work for some reason, or if it is just that a bug has crept in somewhere.

Any hints would be much appreciated.

Thanks.

More details...

I've gone back and played at installing old versions. Basically 7.0.5 and before work, 7.2.1 and later don't work on the aforementioned processor counts.

I'm guessing that this comes from something I have seen mentioned in the release notes for 7.2 from http://www.abinit.org/downloads/source- ... notes.html :

B.4 Automatic determination of the best processor repartition is now available for ground-state
and response function calculation:
- by means of a simple heuristics
- by means of a series of small benchmarks using the scheme applied for the LOBPCG algorithm.
The size of communicators for parallel matrix algebra routines can be optimized, as well
as the use (or not) of matrix algebra on GPU.
See the description in the input variable autoparal, values 1, 2 or 3..
Automatic tests should still be written.
Work by F. Dahm, M. Torrent and F. Jollet.

My guess is that some parts are changing the processor counts successfully, but maybe the header reading is being done additionally on processors that shouldn't be taking part. Maybe this part is totally unrelated. Hopefully someone else is better placed than me to work out what has gone wrong rather than my uninformed guesswork!

Thanks.