5. Troubleshooting

The following examples are the result of lengthy debugging sessions and can be checked if problems arise during the development process. Especially when using MPI, the debugging process can be quite cumbersome.

5.1. WriteArrayToHDF5() and the collective flag

Error Description

  • One or more array elements in a .h5 file are corrupt, e.g., the element contains the value 1.16828195e-314 instead of the expected value of 7960

Necessary Conditions

  • LIBS_USE_MPI=ON: the error only occurs when using more than 1 process (a multi-node run with a large number of processes might yield a high chance to trigger this problem)

Sufficient Conditions

  • CALL WriteArrayToHDF5() with collective=.TRUE. but not all processes enter this function

Finding the Bug

  • This error can be found with a regression test that runs with LIBS_USE_MPI=OFF or one process and again with multiple processes at best using the multi-node feature PICLAS_SHARED_MEMORY=OMPI_COMM_TYPE_CORE

Resolution

  • CALL WriteArrayToHDF5() with collective=.FALSE. when it is not 100% certain that all processes enter this routine

Explanation

  • Setting collective=.TRUE. triggers the usage of H5FD_MPIO_COLLECTIVE_F (collective=.FALSE. uses H5FD_MPIO_INDEPENDENT_F) in H5PSET_DXPL_MPIO_F(PList_ID, H5FD_MPIO_INDEPENDENT_F, iError), which configures the “transfer mode” in the hdf5 output. Collective MPI output requires that all processes take part in the operation!

git hashes

5.2. Seemingly meaningless change in code triggers segmentation fault or slow down of the code

Error Description

  • A user-defined read-in parameter is commented in/out which triggers the bug due to different values that are read into the parameter, e.g., setting 1e-14 or 10e-15 (which is basically the same value!) The suspected problem is that something is read from the memory, which should not be read, e.g., an uninitialized variable (that therefore points to a possibly random location in the memory) This causes an undefined state (non-deterministic outcome, different compilers/different machines yield different effects). The code might not crash but hang for a certain amount of time (this can be used to find the bug).

Necessary Conditions

  • An integer variable that is used for (indirect) allocation of an array is uninitialized

Sufficient Conditions

  • A function or subroutine is called that declares an array depending on an uninitialized integer variable

Finding the Bug

  • When the code does not crash but instead hangs for a certain amount of time, the following print statements can be used (when not utilizing a debugger)

    IPWRITE(UNIT_StdOut,'(I0,A,I0)') ": v "//TRIM(__FILE__)//" +",__LINE__

    to see the exact position in the code where the code hangs (due to a possible function call and subsequent allocation process) for a short period of time

Resolution

  • Nullify all integer variables that are used for allocation per default

Explanation

  • The following variable was declared in a subroutine

    REAL :: DistriOld(SpecDSMC(PartSpecies(iPart1))%MaxElecQuant)

    but the integer SpecDSMC(PartSpecies(iPart1))%MaxElecQuant was not initialized because the corresponding model is not used in this specific case and therefore DistriOld is never used. It is however allocated with an undefined state, with undefined outcome! In this case the bug was fixed by using the “assignment to an allocatable array”

    REAL,ALLOCATABLE              :: DistriOld(:)
...
DistriOld = ElectronicDistriPart(iPart1)%DistriFunc

    For gfortran, Fortran 2003’s assignment to an allocatable array was introduced in 4.6, 2011-01-28.

git hashes

5.3. Possible memory leak detection when using MPICH

Error Description

  • The error output looks like this

    ====================================================================================================================================
PICLAS FINISHED! [ 2.41 sec ] [ 0:00:00:02 ]
====================================================================================================================================
Abort(810645775): Fatal error in internal_Finalize: Other MPI error, error stack:
internal_Finalize(50)...........: MPI_Finalize failed
MPII_Finalize(400)..............:
MPID_Finalize(652)..............:
MPIDI_OFI_mpi_finalize_hook(856):
destroy_vni_context(1094).......: OFI domain close failed (ofi_init.c:1094:destroy_vni_context:Device or resource busy)
[WARNING] yaksa: 4 leaked handle pool objects

    and shows that piclas finishes successfully, but an MPI error is invoked afterwards. Note that last line containing “[WARNING] yaksa: 4 leaked handle pool objects” might not be there and sometimes reflects the number of processes minus one.

Necessary Conditions

  • MPICH must be used instead of OpenMPI. The problem even occurs when only one single process is used.

Finding the Bug

  • Activate PICLAS_DEBUG_MEMORY=ON and check all the pairs of, e.g.,

    myrank=      0               Allocated ElemBaryNGeo_Shared_Win with WIN_SIZE =                  240

    with

    myrank=      0                          Unlocking ElemBaryNGeo_Shared_Win with MPI_WIN_UNLOCK_ALL()
myrank=      0                     Freeing window ElemBaryNGeo_Shared_Win with       MPI_WIN_FREE()

    to find the missing CALL UNLOCK_AND_FREE(ElemBaryNGeo_Shared_Win) by running piclas and storing the output in, e.g., std.out and then running the following command

    STDOUT='std.out'; dashes='----------------------------------------'; for line in $(grep -o -P '(?<=Allocated).*(?=with)' ${STDOUT} | sort -u | xargs); do printf 'Checking [%s] %s' "$line" "${dashes:${#line}}"; NbrOfA=$(grep -iw "${line}" ${STDOUT} | grep -ic "Allocated"); printf ' allocated %sx' "$NbrOfA"; NbrOfDA=$(grep -iw "${line}" ${STDOUT} | grep -ic "Unlocking"); printf ' deallocated %sx' "$NbrOfDA"; if [[ $NbrOfDA -lt $NbrOfA ]]; then echo " ---> Could not find MPI_WIN_UNLOCK_ALL() and MPI_WIN_FREE() for this variable"; else echo "... okay"; fi; done

    Replace std.out in the command if a different file name is used. If a variable is not correctly freed, the output of the script should look like this

    Checking [ElemSideNodeID_Shared_Win] --------------- allocated 2x deallocated 2x... okay
Checking [ElemMidPoint_Shared_Win] ----------------- allocated 2x deallocated 2x... okay
Checking [ElemNodeID_Shared_Win] ------------------- allocated 2x deallocated 1x ---> Could not find all required MPI_WIN_UNLOCK_ALL() and MPI_WIN_FREE() for this variable
Checking [ElemBaryNGeo_Shared_Win] ----------------- allocated 2x deallocated 2x... okay
Checking [ElemRadius2NGeo_Shared_Win] -------------- allocated 2x deallocated 2x... okay
Checking [ElemCurved_Shared_Win] ------------------- allocated 2x deallocated 2x... okay

Resolution

  • Add the missing CALL UNLOCK_AND_FREE(MYNAME_Win), where MYNAME is the name of the shared memory window.

Explanation

  • MPI_WIN_UNLOCK_ALL() and MPI_WIN_FREE() must be applied to shared memory windows before CALL MPI_FINALIZE(iError) is called.

git hashes