# SSOR problem

SUN Chun Chun.SUN at 3ds.com
Mon Apr 27 13:45:45 CDT 2009

```Hello,

Sorry to raise the question again. It's a bit detailed..

I was trying to make the column indices a->j sorted for each row for my
SEQAIJ Matrix (reason explained below). I looked up the source code and
found MAT_COLUMNS_SORTED option, and I can do a MatSetOption to set
that. However, even now with a->sorted=PETSC_TRUE, I still see my a->j's
unsorted for each row after a bunch of MatSetValues. Is there a
particular procedure that I should follow to have each row of my matrix
sorted? Or I have to do it outside PETSc?

I should not care how PETSc store matrix, however, the reason that I
need it to be sorted is that in SSOR implementation, there seems to be
an assumption that the column indices being sorted for zero initial
guess ( n=diag[i]-a->i[i]; in MatRelax_SeqAIJ )... I suspect I might
have misunderstandings here. But at least I would like to try a sorted
matrix on SSOR to clarify.

Any suggestions?

Thanks,

Chun

From: petsc-users-bounces at mcs.anl.gov
[mailto:petsc-users-bounces at mcs.anl.gov] On Behalf Of Matthew Knepley
Sent: Monday, April 27, 2009 11:33 AM
To: PETSc users list
Subject: Re: SSOR problem

The function MatMarkDiagonal_SeqAIJ() takes care of this.

Matt

On Mon, Apr 27, 2009 at 9:34 AM, SUN Chun <Chun.SUN at 3ds.com> wrote:

Hello,

I have an update to this problem:

I found that in MatRelax_SeqAIJ function (mat/impl/aij/seq/aij.c), I
have:

diag = a->diag and:

diag[i] is has exactly the same value of a->i[i] for each row i. This
gives me n=0 when doing forward pass of zero initial guess. That
explains why setting -pc_sor_forward will give me identical results as
if I run pure DSCG.

I assume that this a->diag[] stores the sparse column index of diagonal
entries of a matrix. Now it seems to be improperly set. I will pursue
this further in debugger. Do you know which function it should be set
during the assembly process? That would point a short-cut for me....

Thanks again!
Chun

-----Original Message-----
From: petsc-users-bounces at mcs.anl.gov
[mailto:petsc-users-bounces at mcs.anl.gov] On Behalf Of SUN Chun
Sent: Monday, April 27, 2009 9:13 AM
To: PETSc users list
Subject: SSOR problem

Hello,

I have an *particular* Ax=b which I want to solve with CG preconditioned
by SSOR using PETSc. Then some specific strange things happen. Please
allow me to describe all the symptoms that I found here. Thanks for your
help:

0) All solves are in serial.

1) A 20-line academic code and another matlab code converge the solution
with identical residual history and number of iterations (76), they
match well. If I run without SSOR (just diagonal scaled CG): PETSc,
academic code, and matlab all match well with same number (180) of
iterations.

2) PETSc with SSOR seems to give me -8 indefinite pc. If I play with
omega other than using 1.0 (as in Gauss-Seidel), sometimes (with
omega=1.2) I see stagnation and it won't converge then exceeds the
maximum iteration allowed (500). Residuals even don't go down. If I
don't say -ksp_diagonal_scale, I get -8 too. So, PETSc with SSOR either
gives me -8 or -3.

3) The above was run with -pc_sor_symmetric. However, if I ran with
-pc_sor_forward, I got a convergence curve identical to what I have
without any preconditioner, with same iterations (180). If I ran with
-pc_sor_backward, it gives me -8 indefinite pc.

4) If I increase any of the number of -pc_sor_its (or lits) to 2, it
converges (but still don't match the matlab/academic code).

5) The matrix has good condition number (~8000), maximum diagonal is
about 6, minimum diagonal is about 1.1. There's no zero or negative
diagonal entries in this matrix. It's spd otherwise matlab won't be able
to solve it.

6) The behavior is independent of rhs. I've tried random rhs and get the
same scenario.

7) Here is the confusing part: All other matrices that we have except
for this one can be solved by PETSc with same settings very well. And
they match the academic code and matlab code. It's just this matrix that
exhibits the strange behavior. I tend to eliminate the possibility of
interface problem because all other matrices and other preconditioner
settings work well.

We're running out of ideas here, if you have any insight please say
anything or point any directions.

Thanks a lot,
Chun

--
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which