Segfault in ELPA

[oschuett]

Since #3184 the QS/regtest-almo-2/ion-pair.inp has been segfaulting:

==130454== Invalid write of size 8
==130454==    at 0xA8E793B: memmove (in /usr/libexec/valgrind/vgpreload_memcheck-amd64-linux.so)
==130454==    by 0x4CBCE6F: __elpa2_compute_MOD_tridiag_band_real_double (elpa2_tridiag_band_template.F90:1165)
==130454==    by 0x4C83332: __elpa2_impl_MOD_elpa_solve_evp_real_2stage_a_h_a_double_impl (elpa2_template.F90:1155)
==130454==    by 0x4C00A8A: __elpa_impl_MOD_elpa_eigenvectors_a_h_a_d (elpa_impl_math_solvers_template.F90:126)
==130454==    by 0x34534DC: __cp_fm_elpa_MOD_cp_fm_diag_elpa (cp_fm_elpa.F:536)
==130454==    by 0x344E910: __cp_fm_diag_MOD_choose_eigv_solver (cp_fm_diag.F:228)
==130454==    by 0x20AF98E: __preconditioner_makes_MOD_make_full_all.constprop.0 (preconditioner_makes.F:499)
==130454==    by 0x20B27B4: __preconditioner_makes_MOD_make_preconditioner_matrix (preconditioner_makes.F:126)
==130454==    by 0x1EBB40F: __preconditioner_MOD_make_preconditioner (preconditioner.F:202)
==130454==    by 0x1EBC37B: __preconditioner_MOD_prepare_preconditioner (preconditioner.F:430)
==130454==    by 0x1B3584E: __qs_scf_MOD_scf_env_do_scf (qs_scf.F:839)
==130454==    by 0x1B41E53: __qs_scf_MOD_scf (qs_scf.F:246)

I found that changing the OT preconditioner from FULL_ALL to FULL_KINETIC mitigates the problem.

Does anybody have an idea what might be going on?

[fstein93]

What happens if you run the test with a single thread? The pdbg-toolchain (single thread) passes with ELPA but the psmp-testers (two threads) do not.
Regarding the preconditioner: FULL_ALL, FULL_SINGLE and FULL_SINGLE_INVERSE require a diagonalization, whereas FULL_KINETIC and FULL_S_INVERSE do not.

[oschuett]

The psmp binary also crashes with a single thread and rank. Those binaries that don't segfault, produce wrong results instead. Note that I had to lower the threshold in my original PR.

Either this bug is quite old or ALMO somehow assumes a block size of 32.

[oschuett]

It seems that this is not so much an issue with ALMO, but rather with ELPA (or our integration of it).

As a workaround I've now switched regtest-almo-2/ion-pair.inp to scalapack.