[Trilinos-Users] Different results from AztecOO/ML when running serial vs parallel

[Heroux, Michael A]

Eric,

It is fairly common to see an increase in iteration count to reach the same residual value.  The quality of the preconditioner tends to decay as the number of processors increases.  

Mike


________________________________________
From: trilinos-users-bounces at software.sandia.gov <trilinos-users-bounces at software.sandia.gov> on behalf of Eric Marttila <eric.marttila at thermoanalytics.com>
Sent: Friday, January 30, 2015 1:21 PM
To: Trilinos-Users at software.sandia.gov
Subject: [EXTERNAL] [Trilinos-Users] Different results from AztecOO/ML when running     serial vs parallel

I'm using AztecOO with ML to perform linear solves within a non-
linear outer iteration loop.  Experimentation has shown that I
get the best overall convergence (i.e. fastest solution) of my
non-linear solve if use a small number of iterations in AztecOO
when doing each linear solve.

The problem I have is that when running the code with different
numbers of processes I see significantly different results from the
linear solve.

I've attached a small code sample that exercises this issue by
performing a single linear solve.  When running this code with
np=1 and np=4, I see differences in results that are about 0.5%.

If I let AztecOO do more iterations, most of the differences go
away... but this isn't what I want to do since it causes my total
non-linear solution time to increase.

I was not expecting to see the 0.5% difference in results when
running serial vs parallel.  I was expecting differences that are
closer to machine precision (like I see when I have AztecOO
iterate more).  Is what I'm seeing the expected behavior? Or am I
perhaps doing something wrong in setting up AztecOO?

Thanks,
--Eric

p.s. Attached is my source code, as well as a separate .tar.gz
that contains the input files and sample output files.

--
Eric A. Marttila