[CP2K:5174] Regtest: obviously wrong results in TMC
mba... at gmx.net
Mon Apr 14 13:23:07 CEST 2014
On Sun, Apr 13, 2014 at 10:00:00PM -0700, Kaneta Yuske wrote:
> I recently compiled popt version of CP2K 2.5.1 with gfortran 4.8.2 and open
> MPI 1.8 on TSUBAME 2.5 (arch file attached)
> After compilation, I run regtest with reference to CP2k 2.5 branch uploaded
> to sourceforge, but I got 98 wrong results out of 2469 tests.
> Most of the wrong results seemed to come from math library difference.
> However, TMC tests, which account for 23 out of 98 wrong results, showed
> totally different value like -999. Besides, Most of the test related to TMC
> (23 out of 27) ended with wrong results.
> I don???t plan to use Monte Carlo, but I want to know whether these wrong
> results are related to my compiling procedure. Does anyone help me?
In my experience, this happens at least partially if you run the
regtests on a single core, likely due to this code in tmc_setup.F:
CALL tmc_redistributing_cores(tmc_env%tmc_comp_set, para_env, &
WRITE(output_unit,*)"TMC|NOTcalculatedTotal energy: -999"
which apparently gets parsed by the regtest code for comparison. So I
suggest to check the output of the failing tests.
Also, I believe setting a bogus energy on failure and exiting without
error is not in line with the rest of CP2K code and should possibly be
changed (if that is indeed the culprit).
More information about the CP2K-user