forbidden memory area access

[Boone11]

Hi users,

I try to perform a computation on a mesh that is composed of both structured and unstructured hexa but the simulation fails before the first time step with the following error: SIGSEGV signal (forbidden memory area access) intercepted!

I think the error can be caused by the mesh since the same setup works well on another mesh (that is composed of only structured hexa).

I am using GMSH to perform the mesh (.msh). Below are some files that could help understanding the error.
As the mesh is quite big I uploaded it on filesender: https://filesender.renater.fr/?s=downlo ... 5f76eda79a

Notice that there are warnings in the listing file about the mesh but I suppose that they are not responsible for the SIGSEV error since the error occurred before the first time step.

May you help me to understand and solve this issue ?

Best regards

[Yvan Fournier]

Hello,

Do you have user-defined functions/subroutines ? According to your stack trace, the error could be there ?

Regards,

Yvan

[Boone11]

Hi Yvan and thanks for your answer.

Yes I have user-defined subroutines. I perform other tests with a smaller mesh size to run the calculation in sequential and identify the issue. The computation failed due to a user-defined subroutine and more particularly when trying to calculate and store the two vertices of each edge. Here is the code:

Code: Select all

Nseg = 36                       ! number of edges+diagonals per cell
allocate(Vcell(ncelet,Nseg*2))  ! save the 2 vertices coordinates for each edge (for each cell)
allocate(Cptcell(ncelet))       ! counter of Vcell = 6(cells)*6(edges)*2(vertices per edge)=72
Vcell(:,:) = 0
Cptcell(:) = 0

! For all interior faces
! save the 2 vertices of each edge
do ifac = 1, nfac
   iel1 = ifacel(1,ifac)
   iel2 = ifacel(2,ifac)  
   do ii = ipnfac(ifac),ipnfac(ifac+1)-2
      do jj = ii+1, ipnfac(ifac+1)-1
         inod = nodfac(ii)
         jnod = nodfac(jj)

         Vcell(iel1,Cptcell(iel1)*2+1) = inod
         Vcell(iel1,Cptcell(iel1)*2+2) = jnod
         Cptcell(iel1)                 = Cptcell(iel1)+1

         Vcell(iel2,Cptcell(iel2)*2+1) = inod
         Vcell(iel2,Cptcell(iel2)*2+2) = jnod
         Cptcell(iel2)                 = Cptcell(iel2)+1
      enddo
   enddo
enddo

This code works well on fully structured mesh. I do not see what causes trouble when the mesh is also containing non structural cells. Do you have an idea?

The obtained error is:

forrtl: severe (408): fort: (2): Subscript #2 of the array VCELL has value 73 which is greater than the upper bound of 72

Image PC Routine Line Source
libsaturne-5.3.so 00007FB02661A679 Unknown Unknown Unknown
cs_solver 000000000046E391 cs_user_f_initial 428 cs_user_initialization.f90
libsaturne-5.3.so 00007FB024A07353 inivar_ 197 inivar.f90
libsaturne-5.3.so 00007FB0244F67B2 caltri_ 577 caltri.f90
libcs_solver-5.3. 00007FB027EF7835 cs_run 429 cs_solver.c
libcs_solver-5.3. 00007FB027EF7B85 main 654 cs_solver.c
libc-2.28.so 00007FB012CC8CF3 __libc_start_main Unknown Unknown
cs_solver 000000000040E13E Unknown Unknown Unknown
solver script exited with status 152.

Error running the calculation.

Thanks again

[Yvan Fournier]

Hello,

You crash seems to be due to an incorrect estimation of arrays size/allocation. Are you sure about the number of edges + diagonals per cell ? Are you using only hexahedral elements ?

Recent versions of the code (v7.0, or v8.0-beta)have higher level operators (in C) allowing building various mesh adjacencies in a safer manner, so upgrading could be useful (and in any case, v5.3 has been retired since the release of v6.0 in September 2019).

Regards,

Yvan

[Boone11]

Hi Yvan and thanks.

Yes, this routine has been used for a long time in fully structured hexa mesh with success. Yes, according to GMSH there are only hexahedral elements.

I see that Cptcell(iel1) and Cptcell(iel2) sometimes become bigger than the allocated size of Vcell (Nseg*2=72) whereas they should not. I do not see a reason for this behavior.

For the moment, to unblock the situation, I used a if statement to allocate Vcell only if Cptcell(iel1) and Cptcell(iel2) are lower than 72. It works well on single processor but crashes on parallel computation. I do not see why in parallel computation issues occur. There are no error file created, the listing is suddenly interrupted after the partionning step. The slurm file gives:

Abort(17) on node 8 (rank 8 in comm 0): Fatal error in PMPI_Waitall: See the MPI_ERROR field in MPI_Status for the error code

I provide the end of the listing file and the only partition_output file in attached file (i added the .txt extension to be able to upload it). The mesh output can be uploaded following the link: https://filesender.renater.fr/?s=downlo ... 65d6b35bcb


Note that, if I remove the code provided in the message above there is no issue in parallel computation.

The moving to newer versions of the code is planned.

[Boone11]

Hi,

I think that I have found the reason of the error. In the listing file it is written:

Code: Select all

Before joining
     Number of cells:          71056
     Number of interior faces: 202822
     Number of boundary faces: 20692
     Number of vertices:       83825

  Global number of boundary faces selected for joining:      11040

  Element selection successfully done.
  Determination of possible face intersections:

    bounding-box tree layout: 3D

  Global number of intersections detected:        84780
    Vertex-Vertex intersections:           79557
    Other intersections:                    5223

  Joining operation is non-conforming.

  Global number of new vertices to create:       7141

  Equivalences broken for 0 edges.

  Edge intersections and vertex creation done.

  Pre-merge for 9930 global element couples.

  Max size of a merge set of vertices: 3

  Excessive transitivity for 0 set(s) of vertices.

  Edge removed for 0 faces (global).
  Join mesh cleaning done.

  Edge removed for 0 faces (global).
  Join mesh cleaning done.

  Merge vertices and mesh update done.

  *** WARNING ***
  Globally, 250 problem(s) found during the face splitting
               66  open cycles,
              184  edges traversed twice,
                0  faces split into more than max_subfaces (= 200)

    => Eventually modify joining parameters

Joining operation is non-conforming.

I think that I should not use the joining option. The user routines are not working with non conforming mesh.