[Trilinos-Users] PETSc interface

[Jed Brown]

"Heroux, Mike" <MHeroux at CSBSJU.EDU> writes:

> The situation here is wrapping a PETSc AIJ matrix to implement the
> Epetra_RowMatrix interface.  Doing this permits the PETSc matrix to be
> used with any Trilinos solver or preconditioner.

In our ML interface, we maintain code that wraps PETSc matrices for ML.
Most of our users get ML via --download-ml, which is pretty reliable on
most architectures.  Based on the number of build failures our users run
into with much simpler CMake packages such as METIS and ParMETIS (from
buggy/incompatible CMake versions or other environment problems), and
the time to build a full Trilinos stack (as compared to just ML), I'm
not sure we can afford to support a --download-trilinos.  Some users
install Trilinos on their own and tell PETSc to link ML from there.
That also works, but takes more thinking on the part of the users.  I'm
afraid that an Epetra interface distributed with PETSc will not get used
much.  What are your thoughts on who should "own" that interface?

On a related topic, there has been significant ML development since the
last stand-alone release.  Would it be possible to make another
stand-alone release?  Many users are happy to add --download-ml to their
PETSc configure, knowing that it only adds a few seconds to the
configure/build time and gives them the option of trying another
smoothed aggregation package.  Most of those users would not be willing
to wait for --download-trilinos, even if we could provide equivalent
portability/robustness.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 835 bytes
Desc: not available
URL: <http://software.sandia.gov/pipermail/trilinos-users/attachments/20140430/94ce1641/attachment.sig>