dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

Performance is another good reason to wrap PETSc (in addition to the
simplicity of using the wrappers).

Setting the options correctly to PETSc can make a big difference in
performance and hopefully we'll be able to set some good default
options in our wrappers.

Another example of this is the assembly. If care is not taken to set
the options correctly before assembly (like we do now automatically),
the assembly will run a factor 1000 slower...

/Anders


On Wed, Mar 22, 2006 at 04:47:38PM +0100, Garth N. Wells wrote:

> Matt told me that the PETSc LU solver is poor, but still I was really
> surprised at the speed-up.
> 
> Garth
> 
> On Wed, 2006-03-22 at 09:39 -0600, Anders Logg wrote:
> > Very nice.
> > 
> > /Anders
> > 
> > On Wed, Mar 22, 2006 at 04:41:54PM +0100, DOLFIN wrote:
> > > One or more new changesets pushed to the primary DOLFIN repository.
> > > A short summary of the last three changesets is included below.
> > > 
> > > changeset:   1794:236275c67493735a92352a3735c6e525313ae723
> > > tag:         tip
> > > user:        "Garth N. Wells <g.n.wells@xxxxxxxxxx>"
> > > date:        Wed Mar 22 16:41:19 2006 +0100
> > > files:       src/kernel/la/KrylovSolver.cpp src/kernel/la/Matrix.cpp src/kernel/la/dolfin/Matrix.h src/kernel/nls/NewtonSolver.cpp src/kernel/nls/dolfin/NewtonSolver.h src/kernel/parameter/dolfin/DefaultParameters.h src/kernel/pde/LinearPDE.cpp
> > > description:
> > > Allow PETSc matrix type to be set through the Matrix constructor.
> > > 
> > > The type of PETSc matrix determines which LU solver is used when doing a direct solve. Currently implemented  matrix types are:
> > >   Matrix::default_solver (PETSc MATSEQAIJS)
> > >   Matrix::spooles        (PETSC MATSEQAIJSPOOLES)
> > >   Matrix::superlu        (PETSc MATSUPERLU)
> > >   Matrix::umfpack        (PETSC MATUMFPACK) 
> > > 
> > > By default, PDE.solve() will use Matrix::umfpack if Umfpack has been compiled with PETSc, otherwise it will use the default PETSc LU solver.
> > > 
> > > 
> > > Benchmark (for the linear solve only) for the problem in src/demo/pde/elasticity
> > >   
> > >   default_solver: 48.4s
> > >   spooles:        3.45s
> > >   superlu:        5.77s
> > >   umfpack:        1.68s
> > > 
> > > 
> > > changeset:   1793:6c11a3f3c36279d71838fba51caf128fd4a70c1e
> > > user:        "Garth N. Wells <g.n.wells@xxxxxxxxxx>"
> > > date:        Fri Mar 17 13:54:30 2006 +0100
> > > files:       src/kernel/la/KrylovSolver.cpp
> > > description:
> > > Temporary fix to avoid PETSc error when using Hypre preconditioner.
> > > 
> > > The function KrylovSolver::init() is not called when a KrylovSolver is created with the Hypre preconditioner. If a user creates a KrylovSolver with the Hypre preconditioner and then attempts to access the pointer to the PETSc solver before the solver has been used to solve a system, an error will result.
> > > 
> > > 
> > > changeset:   1792:339e1b957d0433c13176174e13b3ed2122db7677
> > > user:        "Garth N. Wells <g.n.wells@xxxxxxxxxx>"
> > > date:        Thu Mar 16 15:30:52 2006 +0100
> > > files:       src/demo/fem/convergence/main.cpp
> > > description:
> > > Replace removed KrylovSolve::setRtol() function with solver.set("Krylov relative tolerance, ..).
> > > 
> > > 
> > > -------------------------------------------------------
> > > For more details, visit http://www.fenics.org/hg/dolfin
> > > 
> > > _______________________________________________
> > > DOLFIN-dev mailing list
> > > DOLFIN-dev@xxxxxxxxxx
> > > http://www.fenics.org/cgi-bin/mailman/listinfo/dolfin-dev
> > > 
> 
> 
> 
> _______________________________________________
> DOLFIN-dev mailing list
> DOLFIN-dev@xxxxxxxxxx
> http://www.fenics.org/cgi-bin/mailman/listinfo/dolfin-dev
> 

-- 
Anders Logg
Research Assistant Professor
Toyota Technological Institute at Chicago
http://www.tti-c.org/logg/