Do you mean using "-pc_factor_mat_ordering_type  wbm"? It does not seem to be 
in the manual. What exactly is it? I will try it anyway.
 
Many thanks,
Qin



On Tuesday, February 18, 2014 10:11 AM, Jed Brown <[email protected]> wrote:
  
Qin Lu <[email protected]> writes:


> RCM did not help much for my case. This case represents some strong
> connectivities in channels. I am wondering if PETSc has some
> reordering algorithm or solver that is not based on connectivity
> graph, but is based on connectivity strength (such as percolation type
> of reordering). Any information is appreciated.    Thanks, Qin

There isn't such an ordering, but it would be a welcome contribution.
One challenge is that features like anisotropy are not necessarily
apparent in the matrix entries.  In that case, and for vector-valued
problems, you would ideally use a better strength-of-connection measure.
If you computed a strength-of-connection, then thresholded (or used edge
weights as a "priority"), followed by something like an RCM ordering,
you could find something good for low-fill incomplete factorization in
problems with hidden anisotropy.

Reply via email to