[petsc-users] Setting up MUMPS in PETSc

[Jinquan Zhong]

Thanks, Alexander.  

That explains the phenomena I needed to use 1152 cores for 576 partitions and only 432 cores for 144 partitions.  I will come up with more questions for you.

Jinquan
________________________________________
From: petsc-users-bounces at mcs.anl.gov [petsc-users-bounces at mcs.anl.gov] on behalf of Alexander Grayver [agrayver at gfz-potsdam.de]
Sent: Wednesday, October 24, 2012 2:10 AM
To: PETSc users list
Subject: Re: [petsc-users] Setting up MUMPS in PETSc

Jinquan,

As a rule of thumb, you should always try to run MUMPS on as few nodes
as possible which have as much memory as possible. Given the same
matrix, MUMPS requires more memory when run with, for instance, 4
process than with 2. You gain nothing trying thousands of processes
without even looking at statistics mumps gives you (through its output).

If your matrix is symmetric you should use cholesky (LDL^T)
factorization in MUMPS. This will save a lot of memory and time.

To estimate conditioning of the problem you may also consider the
algorithm described here:
http://scicomp.stackexchange.com/questions/3486/what-are-the-symptoms-of-ill-conditioning-when-using-direct-methods

It is easy to implement it using MUMPS.

--
Regards,
Alexander