[petsc-users] Avoid MUMPS ordering for distributed matrices?

Barry Smith bsmith at mcs.anl.gov
Wed Nov 23 09:25:33 CST 2011


On Nov 23, 2011, at 7:18 AM, Jed Brown wrote:

> On Wed, Nov 23, 2011 at 05:43, Dominik Szczerba <dominik at itis.ethz.ch> wrote:
> In my procedure considerable time is spent to partition the domain.
> When using MUMPS as a solver for my matrix I see the message:
> 
> "Ordering based on METIS"
> 
> This is an ordering to reduce fill in factorization, not to to partition the domain. Last I heard, symbolic factorization was done in serial, which explains why you find it taking a lot of time. The right hand side and solution vectors are also passed on rank 0, which presents another inefficiency/imbalance and memory bottleneck. Talk to the MUMPS developers or use a different package if you don't like these properties.

   Mumps does now have an option of parallel ordering. 

   Run with -help and look at the options like 

-mat_mumps_icntl_28","ICNTL(28): use 1 for sequential analysis and ictnl(7) ordering, or 2 for parallel analysis and ictnl(29) ordering

-mat_mumps_icntl_29","ICNTL(29): parallel ordering 1 = ptscotch 2 = parmetis


I apologize that the options are organized in such a silly way but that is how MUMPS is organized.

   Barry




More information about the petsc-users mailing list