[petsc-dev] PaStiX solution

[Jed Brown]

On Wed, Dec 28, 2011 at 08:11, Alexander Grayver <agrayver at gfz-potsdam.de>wrote:

> I tried LU, the result is the same (attached). The matrix is recognized as
> a symmetric, even though I put in the code:
>     ierr = MatSetOption(A,MAT_SYMMETRIC,PETSC_FALSE);
>     ierr = MatSetOption(A,MAT_SPD,PETSC_FALSE);
>
> Seems like internal check?
> I checked positive-definiteness of the matrix in matlab (using [L,p,s] =
> chol(A); test) and it's not SPD.
> So it seems like PaStiX is not very robust. This case is not one inch
> close to the bad cases I used MUMPS with.
>

Just for kicks, can you try making the change below and re-running your
tests with LU? This is code that the PaStiX developers sent, but it doesn't
make sense to me as written.

$ hg diff src/mat/impls/aij/mpi/pastix/pastix.c
--- a/src/mat/impls/aij/mpi/pastix/pastix.c
+++ b/src/mat/impls/aij/mpi/pastix/pastix.c
@@ -446,16 +446,6 @@
     ierr = PetscMalloc((lu->n)*sizeof(PetscInt)
,&(lu->invp));CHKERRQ(ierr);
   }

-  if (isSym) {
-    /* On symmetric matrix, LLT */
-    lu->iparm[IPARM_SYM] = API_SYM_YES;
-    lu->iparm[IPARM_FACTORIZATION] = API_FACT_LDLT;
-  } else {
-    /* On unsymmetric matrix, LU */
-    lu->iparm[IPARM_SYM] = API_SYM_NO;
-    lu->iparm[IPARM_FACTORIZATION] = API_FACT_LU;
-  }
-
   /*----------------*/
   if (lu->matstruc == DIFFERENT_NONZERO_PATTERN){
     if (!(isSeqAIJ || isSeqSBAIJ)) {
@@ -529,7 +519,7 @@

   PetscFunctionBegin;
   lu->iparm[IPARM_FACTORIZATION] = API_FACT_LU;
-  lu->iparm[IPARM_SYM]           = API_SYM_YES;
+  lu->iparm[IPARM_SYM]           = API_SYM_NO;
   lu->matstruc                   = DIFFERENT_NONZERO_PATTERN;
   F->ops->lufactornumeric        = MatFactorNumeric_PaStiX;
   PetscFunctionReturn(0);
@@ -545,7 +535,7 @@

   PetscFunctionBegin;
   lu->iparm[IPARM_FACTORIZATION]  = API_FACT_LLT;
-  lu->iparm[IPARM_SYM]            = API_SYM_NO;
+  lu->iparm[IPARM_SYM]            = API_SYM_YES;
   lu->matstruc                    = DIFFERENT_NONZERO_PATTERN;
   (F)->ops->choleskyfactornumeric = MatFactorNumeric_PaStiX;
   PetscFunctionReturn(0);
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20111228/cb3b16fb/attachment.html>