[petsc-users] parallel ILU

[Craig Tanis]

Hi,
I'm trying to get ILU(k) preconditioning to work with an MPIBAIJ matrix. It appears that I need to use an MPIAIJ matrix and that I also need to build petsc with the blocksolve95 interface (though some google searches suggest that this was removed in Petsc 2?).

1) Should I be able to replace a call to MatCreateMPIBAIJ with a sequence of MatCreateMPIAIJ and MatSetBlockSize?

2) Is blocksolve95 still necessary, and if so where can one download it? It appears to have moved from it's original location on the mcs.anl.gov site.

It is possible to do ILU-k preconditioning on an MPI matrix, right?

Thanks!
Craig Tanis

[Matt Knepley]

ILU(k) was never in Blocksolve, just ilu(0). Also, Blocksolve is no
longer supported. You can try Euclid from the Hypre package at LLNL.
Configure can download Hypre for you.

Matt

[Jed Brown]

Post by Craig Tanis
It is possible to do ILU-k preconditioning on an MPI matrix, right?

Note that ILU (regardless of fill) scales very poorly, so it really only
makes sense on a small number of cores, and even then, only if you can't
find anything else that works. As Matt says, build with Hypre, then

-pc_type hypre -pc_hypre_type euclid -pc_hypre_euclid_levels k

Jed

[Barry Smith]

On Mon, 15 Feb 2010 12:45:21 -0500, Craig Tanis <craig-

Post by Craig Tanis
It is possible to do ILU-k preconditioning on an MPI matrix, right?

Note that ILU (regardless of fill) scales very poorly, so it really only
makes sense on a small number of cores, and even then, only if you can't
find anything else that works. As Matt says, build with Hypre, then
-pc_type hypre -pc_hypre_type euclid -pc_hypre_euclid_levels k
Jed

Also note that there is no support for BAIJ matrices. It is only a
"scalar" ILU, not one that works on little blocks.

Barry

[Craig Tanis]

Post by Jed Brown

Post by Craig Tanis
It is possible to do ILU-k preconditioning on an MPI matrix, right?

Note that ILU (regardless of fill) scales very poorly, so it really only
makes sense on a small number of cores, and even then, only if you can't
find anything else that works. As Matt says, build with Hypre, then
-pc_type hypre -pc_hypre_type euclid -pc_hypre_euclid_levels k
Jed

Also note that there is no support for BAIJ matrices. It is only a "scalar" ILU, not one that works on little blocks.
Barry

could you clarify this? i'm running the euclid preconditioner on a mpibaij matrix, and I get no errors or warnings. a ksp_view verifies that things are configured as expected.

thanks
craig

[Barry Smith]

Post by Craig Tanis

Post by Barry Smith

On Mon, 15 Feb 2010 12:45:21 -0500, Craig Tanis <craig-

Post by Craig Tanis
It is possible to do ILU-k preconditioning on an MPI matrix, right?

Note that ILU (regardless of fill) scales very poorly, so it
really only
makes sense on a small number of cores, and even then, only if you can't
find anything else that works. As Matt says, build with Hypre, then
-pc_type hypre -pc_hypre_type euclid -pc_hypre_euclid_levels k
Jed

Also note that there is no support for BAIJ matrices. It is only a
"scalar" ILU, not one that works on little blocks.
Barry

could you clarify this? i'm running the euclid preconditioner on a
mpibaij matrix, and I get no errors or warnings. a ksp_view
verifies that things are configured as expected.
thanks
craig

As the matrix is passed to hypre it is passed as a AIJ matrix
without a block size. This is because hypre has no concept of a BAIJ
matrix. So yes, your PETSc part of the code is using a BAIJ matrix but
hypre is doing a scalar ILU without a block size.


Barry

[Craig Tanis]

On Feb 16, 2010, at 2:51

As the matrix is passed to hypre it is passed as a AIJ matrix without a block size. This is because hypre has no concept of a BAIJ matrix. So yes, your PETSc part of the code is using a BAIJ matrix but hypre is doing a scalar ILU without a block size.

Thanks for the clarification.Is there much overhead in this conversion? Would I be better off working with a AIJ matrix in the "driver code"?

-craig

[Barry Smith]

I would stick to the BAIJ, since that will give you better
performance in the non-hypre part of the code, while the hypre part
has the same performance because it does not "know" about BAIJ. The
conversion cost is in my experience very small relative to the
computation times.

Barry

Post by Craig Tanis
On Feb 16, 2010, at 2:51

Post by Barry Smith
As the matrix is passed to hypre it is passed as a AIJ matrix
without a block size. This is because hypre has no concept of a
BAIJ matrix. So yes, your PETSc part of the code is using a BAIJ
matrix but hypre is doing a scalar ILU without a block size.

Thanks for the clarification.Is there much overhead in this
conversion? Would I be better off working with a AIJ matrix in the
"driver code"?
-craig