Barry :

1) One problem is all the parallel MatMatSolve() use the MPIDense matrices. What is the likelihood that they could be changed to use the elemental matrices (note that with the MPIDense matrices using the simple array format the MatMatSolve() codes are very straight ward and do not need crazy remappings.

2) The MPIDense matrices use MatGetArray() all over the place; which would not work with any non-trivial cyclic storage crap that those guys love. I may need to write a temporary MatMPIDenseGetArray() until the MPIDense business can be resolved.

Do we need to keep MPIDense around and just strip out all of it that uses plapack? If that is possible?

I prefer this approach, i.e., "just strip out all of it that uses plapack", and keep MPIDense, which is useful as petsc dense multi-vectors, and perhaps other usage. Hong

On Aug 13, 2012, at 12:14 PM, Matthew Knepley <[email protected]> wrote:

...

On Mon, Aug 13, 2012 at 12:07 PM, Barry Smith <[email protected]> wrote:

On Aug 13, 2012, at 12:05 PM, Hong Zhang <[email protected]> wrote:

...

Barry, Elemental interface is not well tested yet. I would wait for few months to remove PLAPACK stuff.

PLAPACK stuff is (and never was) well tested. Does anyone rely on it? I hope not. And I hate to have someone new come along and rely on it.

I am for removing it. I wrote it, and so I can guarantee that it is not maintained. We should just fix anything that comes up rather than fooling ourselves that this current stuff works.

Matt

Barry

...

Hong

On Mon, Aug 13, 2012 at 11:50 AM, Jed Brown <[email protected]>

wrote:

...

On Mon, Aug 13, 2012 at 10:40 AM, Barry Smith <[email protected]> wrote: Should I remove all the MPIDense stuff from PETSc now? It is uses PLAPACK which is buggy and unsupported.

That is still a good distribution for MxN matrices with M>>N. We can do them with Elemental, but that would use a different distribution so it will be more complicated to interact with. (The current Elemental interface uses a squarish distribution, but we can tell Elemental to use the [VC,*] distribution (for which fewer operations are supported).

The main thing I care about for that distribution is QR. The best format is dense and row-aligned. I don't care whether it uses MPIDense or a new "multi-vector" thing, but that concept should be somewhere.

-- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener