[petsc-users] MatView with relatively large matrix in Parallel

[Jed Brown]

On Tue, Sep 21, 2010 at 16:30, Gaetan Kenway <kenway at utias.utoronto.ca> wrote:
> I am trying to output a matrix to file so I can load it instead of computing
> it for faster debugging.

It is almost certainly faster to assemble it using the physics module
than to read it from file.

> The matrix I'm trying to output is drdwt. It is a
> parallel block aij matrix with block size of 5. The matrix is assembled and
> the following code works when I run it in serial:

Are you using a DA?  If so, then try setting the format to
PETSC_VIEWER_NATIVE.  This will use the "PETSc" ordering instead of
transforming the system to the "natural" ordering which is an
expensive operation.  Note that the matrix will have a different
ordering on a different number of processors if you do this.

> I've let it go for 20 minutes with no sign of stopping.

Is the system swapping?  Have you tried attaching a debugger to see
where it was hung ('pgrep yourapp' to see process IDs, then gdb -pid
<PID>, backtrace)?

Jed