[billy]

Hi,

I would like to know what are the major diferences between "Diagonally preconditioned biconjugate gradient" and the "Incomplete-Cholesky preconditioned biconjugate gradient" solvers.

Billy.

[chris]

There's some brief technical info at the following link [paper page 41, PDF page 47]:
http://www-2.cs.cmu.edu/~quake-paper...e-gradient.pdf

[hjasak]

Hi Billy,

This is to do with parallelism. Basically, you can do various kinds of preconditioning with the Conjugate Gradient solver. If you do nothing, the solver is a disaster.

The "best" kind of preconditioning we've got in FOAM is the Incomplete Cholesky. However, the preconditioner works by visiting rows of the matrix (cells) in a particular order and the efficiency of preconditioning depends on that order.

All is well when you're doing simulations on a single CPU. However, when you do a decomposition in parallel, each CPU needs to do its own preconditioning on the local cells (because the cells on other processors cannot be accessed (in order). Therefore, for a parallel run, the incomplete Cholesky preconditioning (ICCG) will give you a slightly different result than the identical run in serial (number of solver iterations may be slightly larger and you get different round-off).

Diagonal preconditioning, on the other hand, does not depend on the ordering of cells, but is much (well, hmm) worse than ICCG. Therefore, when doing single-CPU runs, you really have no reason to run anything else than ICCG (maybe AMG, but that's another, and painful, topic).

However, sometimes you can get into a situation when you absolutely desperately need the same result in parallel and serial, e.g. when chasing bugs. :-) In that case, using Diagonal preconditioning (DCG) will allow you to do that, but you'll pay with the increased simulation time.

Enjoy,

Hrv