[Pw_forum] Crash on wf_collect in multi-pool spin-polarized calcs
pscherpelz at uchicago.edu
Wed Feb 4 22:05:03 CET 2015
No, I've observed no similar crashes in the spin-unpolarized case, and I
just re-checked two spin-unpolarized toy models and saw no errors (2
k-points / 2 pools, also 7 k-points / 7 pools).
On 02/04/2015 02:46 PM, Paolo Giannozzi wrote:
> It looks like yet another case of the usual mess with wavefunctions
> a. kept in memory, b. stored to a memory buffer, c. written to file.
> With one k-point per pool we are in case a. but apparently the code
> thinks to be in case b. that falls back to case c.. Does this happen
> also in the spin-unpolarized case with # of k-points = # of points?
> On Wed, 2015-02-04 at 12:59 -0600, Peter Scherpelz wrote:
>> I'm hitting a crash that I've traced to a fairly particular set of
>> circumstances, and want to check if this is a known and/or reproducible
>> bug beyond what I've found.
>> In detail: I've been running parallelized, spin-polarized pw.x
>> calculations (scf and relax). Quantum-espresso v5.1, using the MPI
>> version on either a single node or cluster. I find that quantum espresso
>> crashes with a davcio error, during the wf_collect stage of the
>> computation, only if the number of pools I'm using is equal to the total
>> number of k-points after spin-polarization is considered (e.g., gamma
>> only with 2 pools, or 2 distinct k-point locations with 4 pools).
>> If I run on half that many pools, I do not get a crash. If I run on an
>> equal number of pools but double the number of k-points, I also do not
>> get a crash. If I set wf_collect to false, I also do not get a crash.
>> I've attached a toy model (Si crystal) that exhibits this behavior; and
>> can include the successful runs with the alternate configurations if
>> that helps.
>> Thanks in advance for your help! And thanks overall to the developers
>> for the program - I'm a fairly new user and it's been working great
>> Peter Scherpelz
>> Pw_forum mailing list
>> Pw_forum at pwscf.org
> Pw_forum mailing list
> Pw_forum at pwscf.org
More information about the users