Segmentation fault likely due to hse

[Vahid999]

Describe the bug

Having compiles the 3.9.0 version with libRI and libxc, I ran a 16-atom GaAs supercell with hse .
The input file is
`INPUT_PARAMETERS

suffix GaAs

ntype 2

pseudo_dir ./

orbital_dir ./

ecutwfc 60 # Rydberg

scf_thr 1e-4 # Rydberg

basis_type lcao # changes the type of basis set

calculation scf # this is the key parameter telling abacus to do a scf calculation

dft_functional hse

mixing_beta 0.8

exx_real_number 1

gamma_only 1

symmetry 1

smearing_method fixed`

The 64-core node has a memory of 2TB. The job crashed with segmentation fault. The log files are attached.

Is this a bug?

Thanks,
Vahid

slurm.log

running_scf.log

warning.log

Expected behavior

I expect to see self consistent calculation to finish with hse.

To Reproduce

I can tar the source code and send it to you.

Environment

Linux
Intel
MKL, FFTW, cereal, libxc, openmpi

Additional Context

No response

Task list for Issue attackers (only for developers)

• Verify the issue is not a duplicate.
• Describe the bug.
• Steps to reproduce.
• Expected behavior.
• Error message.
• Environment details.
• Additional context.
• Assign a priority level (low, medium, high, urgent).
• Assign the issue to a team member.
• Label the issue with relevant tags.
• Identify possible related issues.
• Create a unit test or automated test to reproduce the bug (if applicable).
• Fix the bug.
• Test the fix.
• Update documentation (if necessary).
• Close the issue and inform the reporter (if applicable).

[Vahid999]

When I remove hse, I still get segmentation fault. However, without hse, if I change lcao to pw, the code does not crash.

[Vahid999]

I would appreciate even a single hint.
Thanks

[mohanchen]

Could you run LCAO calculations without HSE?

[mohanchen]

You don't need to specify "ntype" parameter, ABACUS will automatically detect it. Maybe use should first try PBE functional and see if the code ends normally.

[ErjieWu]

I reproduce the segment fault with INPUT and STRU (from running_scf.log) files you offered. This will happens when the number of processors>=12. Seems like a bug?

[Vahid999]

Thank you for your kind feedback. I did compile the 3.9.0 version with both LibCOMM and LibRI and the compilation was successful. I would be happy to send you my source code. The attached zip file contains 1) a parallel test using lcao and no hse and 2) a serial test with lcao and hse. Both crash with the segfault. Thanks, Vahid On Feb 5, 2025, at 4:01 AM, Erjie Wu ***@***.***> wrote: CAUTION: The Sender of this email is not from within Dalhousie. I cannot reproduce the segment fault with INPUT and STRU (from running_scf.log) files you offered. Did you compile abacus with LibComm together with LibRI? Try to use PBE and HSE with single processor (without using mpirun) and see whether it works normally. — Reply to this email directly, view it on GitHub<#5884 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ARHYUBHD7UED67MYI6IGPAL2OHAPJAVCNFSM6AAAAABWMR2MUOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMMZWGAYDKNJQGI>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[Vahid999]

The zip file did not go through the email so here it is. I will try with 12 cpus.

tests.zip

[Vahid999]

Using 8 cores, the lcao job without hse crashed similar to the other jobs. I did build the code using

cmake --build build -j 64

I wonder if 64 is causing the problem. The nodes on our Narval cluster have 64 cores.

[ErjieWu]

The error may cause by the memory cost. In HSE calculation, each process consumes a large amount of memory space. On my machine, when the number of MPI parallel processes reached 11, the running memory was full and almost impossible to compute. For hse, more OpenMP thread parallelism is recommended for acceleration.

ABACUS should throw a warning and quit the programme for such bad settings and this seems not work in v3.9.0. Thanks for pointing this out.

[Vahid999]

I see! In that case, I will use much more OpenMP threads than MPI processes. Many thanks for your help. Vahid On Feb 7, 2025, at 3:24 AM, Erjie Wu ***@***.***> wrote: CAUTION: The Sender of this email is not from within Dalhousie. The error may cause by the memory cost. In HSE calculation, each process consumes a large amount of memory space. On my machine, when the number of MPI parallel processes reached 11, the running memory was full and almost impossible to compute. For hse, more OpenMP thread parallelism is recommended for acceleration. — Reply to this email directly, view it on GitHub<#5884 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ARHYUBBZHIMPP36OPW3WOMD2ORNRHAVCNFSM6AAAAABWMR2MUOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNBSGEZDGNRVGI>. You are receiving this because you authored the thread.Message ID: ***@***.***>