[petsc-users] SNESComputeJacobianDefaultColor troubles

Barry Smith bsmith at mcs.anl.gov
Mon Nov 2 16:26:46 CST 2015


> On Nov 2, 2015, at 4:06 PM, Francesco Magaletti <francesco.magaletti at uniroma1.it> wrote:
> 
> Thanks Barry!! 
> I tried and it worked nice, but it will be improved if I could manage the off-diagonal elements block by block. 
> I have a stencil with 5 points (in 1D) and the coupling with the i+1 is different with respect to the i+2 elements. I found on PetsC site a reference to DMDASetGetMatrix function, but there are no examples. Can you give me an example with a simple matrix? Or there is a simpler way to treat the different off-diagonal blocks separately?

  You will need to dig directly into the PETSc code that generates the matrix preallocation and nonzero structure for the DMDA and make a copy of the routine and modify it as needed. See for example DMCreateMatrix_DA_2d_MPIAIJ_Fill in src/dm/impls/da/fdda.c

> 
> Thank you again
> Francesco
> 
>> Il giorno 02/nov/2015, alle ore 17:54, Barry Smith <bsmith at mcs.anl.gov> ha scritto:
>> 
>> 
>>  Don't do any of that stuff you tried below.  Just use  DMDASetFillBlocks() before DMCreateMatrix()
>> 
>> Barry
>> 
>> 
>>> On Nov 2, 2015, at 9:58 AM, Francesco Magaletti <francesco.magaletti at uniroma1.it> wrote:
>>> 
>>> Hi everyone, 
>>> 
>>> I’m trying to solve a system of PDE’s with a full implicit time integration and a DMDA context to manage the cartesian structured grid. 
>>> I don’t know the actual jacobian matrix (actually it is too cumbersome to be easily evaluated analytically) so I used  SNESComputeJacobianDefaultColor to approximate it:
>>> 
>>> DMSetMatType(da,MATAIJ);
>>> DMCreateMatrix(da,&J);
>>> TSGetSNES(ts, &snes);
>>> SNESSetJacobian(snes, J, J, SNESComputeJacobianDefaultColor,0);
>>> 
>>> All it works perfectly but it happens that, when I increase the number of grid points, I go out of memory since the matrix becomes too big to be stored despite my estimate of memory consumption which is much lower than the actual one. Indeed a more careful analysis (with -mat_view) shows that there are a lot of zeros in the nonzero structure of the jacobian. I read in the manual that DMCreateMatrix preallocates the matrix by somewhat using the stencil width of the DMDA context, but in my case I don’t really need all those elements for every equation. 
>>> I then tried with a poor preallocation, hoping petsC could manage it with some malloc on the fly:
>>> 
>>> MatCreateAIJ(PETSC_COMM_WORLD,PETSC_DECIDE,PETSC_DECIDE,size,size,3,0,0,0,&J);
>>> MatSetOption(J,MAT_NEW_NONZERO_LOCATIONS,PETSC_TRUE);
>>> MatSetOption(J,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
>>> MatSetOption(J,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
>>> TSGetSNES(ts, &snes);
>>> SNESSetJacobian(snes, J, J, SNESComputeJacobianDefaultColor,0);
>>> 
>>> but now it gives me runtime errors:
>>> 
>>> [0]PETSC ERROR: --------------------- Error Message --------------------------------------------------------------
>>> [0]PETSC ERROR: Object is in wrong state
>>> [0]PETSC ERROR: Matrix must be assembled by calls to MatAssemblyBegin/End();
>>> [0]PETSC ERROR: See http://www.mcs.anl.gov/petsc/documentation/faq.html for trouble shooting.
>>> [0]PETSC ERROR: Petsc Release Version 3.6.0, Jun, 09, 2015 
>>> [0]PETSC ERROR: ./bubble_petsc on a arch-linux2-c-debug named ... Mon Nov  2 16:24:12 2015
>>> [0]PETSC ERROR: Configure options --with-cc=gcc --with-cxx=g++ --with-fc=gfortran --download-fblaslapack --download-mpich
>>> [0]PETSC ERROR: #1 MatFDColoringCreate() line 448 in /home/magaletto/Scaricati/petsc-3.6.0/src/mat/matfd/fdmatrix.c
>>> [0]PETSC ERROR: #2 SNESComputeJacobianDefaultColor() line 67 in /home/magaletto/Scaricati/petsc-3.6.0/src/snes/interface/snesj2.c
>>> [0]PETSC ERROR: #3 SNESComputeJacobian() line 2223 in /home/magaletto/Scaricati/petsc-3.6.0/src/snes/interface/snes.c
>>> [0]PETSC ERROR: #4 SNESSolve_NEWTONLS() line 231 in /home/magaletto/Scaricati/petsc-3.6.0/src/snes/impls/ls/ls.c
>>> [0]PETSC ERROR: #5 SNESSolve() line 3894 in /home/magaletto/Scaricati/petsc-3.6.0/src/snes/interface/snes.c
>>> [0]PETSC ERROR: #6 TSStep_Theta() line 197 in /home/magaletto/Scaricati/petsc-3.6.0/src/ts/impls/implicit/theta/theta.c
>>> [0]PETSC ERROR: #7 TSStep() line 3098 in /home/magaletto/Scaricati/petsc-3.6.0/src/ts/interface/ts.c
>>> [0]PETSC ERROR: #8 TSSolve() line 3282 in /home/magaletto/Scaricati/petsc-3.6.0/src/ts/interface/ts.c
>>> 
>>> If I use instead the same preallocation but with the FD calculation without coloring it runs without problems (it is only extremely slow) and it uses the correct number of nonzero values:
>>> 
>>> MatCreateAIJ(PETSC_COMM_WORLD,PETSC_DECIDE,PETSC_DECIDE,size,size,3,0,0,0,&J);
>>> MatSetOption(J,MAT_NEW_NONZERO_LOCATIONS,PETSC_TRUE);
>>> MatSetOption(J,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
>>> MatSetOption(J,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
>>> TSGetSNES(ts, &snes);
>>> SNESSetJacobian(snes, J, J, SNESComputeJacobianDefault,&appctx);
>>> 
>>> What do you suggest to solve this problem? 
>>> I also tried to set up a first step with fd without coloring in order to allocate the correct amount of memory and nonzero structure and successively to switch to the colored version:
>>> 
>>> TSSetIFunction(ts,NULL,FormIFunction,&appctx);
>>> 
>>> MatCreateAIJ(PETSC_COMM_WORLD,PETSC_DECIDE,PETSC_DECIDE,size,size,3,0,0,0,&J);
>>> MatSetOption(J,MAT_NEW_NONZERO_LOCATIONS,PETSC_TRUE);
>>> MatSetOption(J,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
>>> MatSetOption(J,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
>>> TSGetSNES(ts, &snes);
>>> SNESSetJacobian(snes, J, J, SNESComputeJacobianDefault,&appctx);
>>> TSStep(ts);
>>> 
>>> ISColoring iscoloring;
>>> MatColoring coloring;
>>> MatFDColoring  matfdcoloring;
>>> MatColoringCreate(J,&coloring); 
>>> MatColoringSetFromOptions(coloring); 
>>> MatColoringApply(coloring,&iscoloring);
>>> MatFDColoringCreate(J,iscoloring,&matfdcoloring);
>>> MatFDColoringSetFunction(matfdcoloring,(PetscErrorCode (*)(void))FormIFunction,&appctx);
>>> MatFDColoringSetFromOptions(matfdcoloring);
>>> MatFDColoringSetUp(J,iscoloring,matfdcoloring);
>>> SNESSetJacobian(snes,J,J,SNESComputeJacobianDefaultColor,matfdcoloring);
>>> 
>>> TSStep(ts);
>>> 
>>> but again it gives me runtime errors:
>>> 
>>> [0]PETSC ERROR: --------------------- Error Message --------------------------------------------------------------
>>> [0]PETSC ERROR: Invalid argument
>>> [0]PETSC ERROR: Wrong type of object: Parameter # 1
>>> [0]PETSC ERROR: See http://www.mcs.anl.gov/petsc/documentation/faq.html for trouble shooting.
>>> [0]PETSC ERROR: Petsc Release Version 3.6.0, Jun, 09, 2015 
>>> [0]PETSC ERROR: ./bubble_petsc on a arch-linux2-c-debug named ... Mon Nov  2 16:37:24 2015
>>> [0]PETSC ERROR: Configure options --with-cc=gcc --with-cxx=g++ --with-fc=gfortran --download-fblaslapack --download-mpich
>>> [0]PETSC ERROR: #1 TSGetDM() line 4093 in /home/magaletto/Scaricati/petsc-3.6.0/src/ts/interface/ts.c
>>> [0]PETSC ERROR: --------------------- Error Message --------------------------------------------------------------
>>> [0]PETSC ERROR: Corrupt argument: http://www.mcs.anl.gov/petsc/documentation/faq.html#valgrind
>>> [0]PETSC ERROR: Invalid Pointer to Object: Parameter # 1
>>> [0]PETSC ERROR: See http://www.mcs.anl.gov/petsc/documentation/faq.html for trouble shooting.
>>> [0]PETSC ERROR: Petsc Release Version 3.6.0, Jun, 09, 2015 
>>> [0]PETSC ERROR: ./bubble_petsc on a arch-linux2-c-debug named ... Mon Nov  2 16:37:24 2015
>>> [0]PETSC ERROR: Configure options --with-cc=gcc --with-cxx=g++ --with-fc=gfortran --download-fblaslapack --download-mpich
>>> [0]PETSC ERROR: #2 DMGetLocalVector() line 44 in /home/magaletto/Scaricati/petsc-3.6.0/src/dm/interface/dmget.c
>>> 
>>> =====================================================================================
>>> =   BAD TERMINATION OF ONE OF YOUR APPLICATION PROCESSES
>>> =   EXIT CODE: 11
>>> =   CLEANING UP REMAINING PROCESSES
>>> =   YOU CAN IGNORE THE BELOW CLEANUP MESSAGES
>>> =====================================================================================
>>> APPLICATION TERMINATED WITH THE EXIT STRING: Segmentation fault (signal 11)
>>> 
>>> where is my mistake?
>>> 
>>> Your sincerely,
>>> Francesco
>> 
> 



More information about the petsc-users mailing list