[CPMD-list] [Fwd: Error Encountered During CPMD-3.9.2 Compilation]

[Axel Kohlmeyer]

On Mon, 4 Jul 2005, Jin Hong Mei wrote:

DL> A collegue of mine has the following questions for the list:
DL> ---------------------------
DL> 
DL> Dear Sirs,
DL> 
DL> I am Dave Lin, Systems Administrator from the Institute of High 
DL> Performance Computing, Singapore. I tried to install CPMD 3.9.2 for our 
DL> researchers and encountered problems. My environment are as follows :

[...]

DL> I tried running the same input file with our previous CPMD 3.9.1 and it 
DL> has been running for 19 hours and still in progress. The partial output 
DL> shows that it has passed the error point of the 3.9.2 run.
DL> 
DL> Please kindly advise me how to rectify the problem.

well, i have several comments to the input and what is 
detectable from the output. most of that concerns the
physics of the problem and not the technical side.

first, and this is the technical part, the MPI parallelization
in CPMD is in general much better than the OpenMP parallelization,
yet the job is run with one MPI node but 32 OpenMP threads.
i don't know whether this is causing the problem (i doubt it),
but is surely will make the job run _very_ inefficiently.
on a machine like yours, it would probably better to run
with 32 MPI-tasks and no OpenMP or maybe, 16 MPI-tasks and 
two OpenMP threads each.

second, the input uses vanderbilt ultra-soft pseudopotentials,
and for them a plane wave cutoff of 80ry is _huge_ and
also _very_ inefficient. something between 20ry and 40ry
is usually sufficient. 35ry should be a safe choice.
for that kind of plane wave cutoff one may have to 
increase the GC-CUTOFF value (e.g. to 1.0d-6).

also, for vanderbilt uspps sometimes the default initial 
atomic guess is not very good (see previous discussions
on cpmd-list and in the manual), so using 
INITIALIZE WAVEFUNCTIONS RANDOM or modifying the atomic
guess via ATOMIC CHARGES might help.

but actually, just with a changed plane wave cutoff
it already runs fine. please see attached output, 
which was, btw, created with 6 dual athlon nodes on a
gigabit ethernet using only MPI parallelization.

finally, and this should have little to do with
the reported problems. i noticed that one oxygen
and one manganese atom are _very_ close together
and also the CONSTRAINTS part looks strange. 
with FIX ATOMS one gives the number and then the
indices, but on the same line and no coordinates, i.e.:

FIX ATOMS
 1  3

however, the three following numbers, although ignored
by cpmd, coincide with the coordinateshat of the
third oxygen (i.e. the one at the strange position).
but in this case the input should be (4xLa + 4xMn + 3xO):

FIX ATOMS
 1  11

on top of that, i'm a bit confused about the purpose
of this, since all other atoms will move around it.

finally, i'd like to note, that with a supercell of
this size you may want to check, whether your k-point
sampling with only the gamma point is sufficient.
unfortunately, this is bit of a problematic topic, as CPMD 
currently does not support k-points with ultra-soft
pseudopotentials...


DL> 
DL> Thank you very much.
DL> 
DL> dave lin

with best regards,
	axel kohlmeyer.

DL> 
DL> 
DL> 

-- 

=======================================================================
Dr. Axel Kohlmeyer   e-mail: axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Lehrstuhl fuer Theoretische Chemie          Phone: ++49 (0)234/32-26673
Ruhr-Universitaet Bochum - NC 03/53         Fax:   ++49 (0)234/32-14045
D-44780 Bochum  http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: test-v1_out.gz
Type: application/x-gzip
Size: 36840 bytes
Desc: 
Url : http://cpmd.org/pipermail/cpmd-list/attachments/20050704/5ba7c573/attachment.gz