"Cholesky failed" problems new in cp2k 3.0 and trunk versions

[Christian Tuma]

Dear cp2k developers,

one of our cp2k users at HLRN (Cray-XC machine) provided me with an MD 
input (230 atoms - Au slab plus CO molecule) that works flawlessly with 
cp2k versions 2.4 and 2.6.2, but fails with the recent 3.0 and current 
trunk versions. Here the error message is

"Cholesky failed: the matrix is not positive definite or ill-conditioned."

and appears just before the first SCF iteration. Switching off the 
preconditioning is a workaround, but then performance is very poor. I have 
attached all files necessary to reproduce the problem.


Some technical details: Tried with GCC 6.1 (with Cray's LibSci 16.03.1) and 
Intel 16.0.3 (with MKL), also in debug mode. Configured and linked against 
external libraries (ELPA, libint, libxc, libsmm, FFTW3). Always crashes 
with the 3.0 and trunk versions, no problems with 2.4 and 2.6.2. So the 
only thing that correlates with the crash is the cp2k release 3.0 and 
younger, it is certainly not due to the compiler or one of the external 
libraries ... 


Now I wonder whether something new was introduced since version 3.0 in the 
context of the preconditioning which causes the problem. Somebody who can 
advice?


Sincerely,
Christian Tuma



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20160628/30b863a6/attachment.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: problem.tar.gz
Type: application/octet-stream
Size: 111780 bytes
Desc: not available
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20160628/30b863a6/attachment.obj>