dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Dec 03, 2007 at 04:36:20PM +0100, Johan Jansson wrote:
> Garth N. Wells wrote:
> >
> >
> > Johan Jansson wrote:
> >> Anders Logg wrote:
> >>> On Mon, Dec 03, 2007 at 01:05:51AM +0100, DOLFIN wrote:
> >>>
> >>>  
> >>>> changeset:   3430:4fa0790a4fc34d3abac3079644368813a252ac61
> >>>> parent:      3428:7967082310c3ab03df9d812a6f471dcf988e7b4e
> >>>> user:        "Johan Jansson <jjan@xxxxxxxxxx>"
> >>>> date:        Sun Dec 02 20:17:50 2007 +0100
> >>>> files:       src/kernel/ode/MonoAdaptiveJacobian.cpp 
> >>>> src/kernel/ode/MonoAdaptiveNewtonSolver.cpp src/kernel/ode/ODE.cpp 
> >>>> src/kernel/ode/TimeSlabJacobian.cpp 
> >>>> src/kernel/ode/dolfin/MonoAdaptiveJacobian.h 
> >>>> src/kernel/ode/dolfin/MonoAdaptiveNewtonSolver.h 
> >>>> src/kernel/ode/dolfin/ODE.h 
> >>>> src/kernel/ode/dolfin/TimeSlabJacobian.h 
> >>>> src/kernel/parameter/dolfin/DefaultParameters.h
> >>>> description:
> >>>> ODE:
> >>>>
> >>>> Added support for computing the Jacobian as a sparse matrix (only for
> >>>> mono-adaptive dG(0) and cG(1) so far, and only for PETSc).
> >>>>     
> >>> Why PETSc? The ODE solvers are tied to uBlas so the vectors won't be
> >>> parallel.
> >>>
> >>> How and when should this be used? Does it change anything in the
> >>> computation that the ODE solver does, or is it just a utility function
> >>> for computing the Jacobian that a user can call?
> >>>   
> >> Ideally the ODE solver should support both uBlas and PETSc, I think 
> >> the plan should be to transition to a general implementation if we 
> >> start to use it more for PDE. PETSc has a higher priority than uBlas 
> >> though, since that's what we use for large computations (due to 
> >> efficiency and robustness), and since it's parallel while uBlas is not.
> >>
> >
> > Why don't you try to implement it through the common interface?  We'll 
> > soon have more linear algebra back-ends (e.g. Trilinos).
> >
> > Garth
> Yes, rewriting the ODE solver with the general linear algebra interface 
> is on my todo-list. If the ODE interface is used more, then the 
> incentive to do this will increase more (since otherwise we cannot use 
> it in parallel for instance).

The main problem here will be how to design the interface so it can be
used both when component-access is possible (uBlas) and when it is not
or very slow (PETSc).

The move to uBlas was to allow component-access.

-- 
Anders