[petsc-dev] petsc-dev: ASM with sbaij matrix and block size=1 broken?

[Matthew Knepley]

On Sat, Nov 5, 2011 at 5:44 PM, Brad Aagaard <baagaard at usgs.gov> wrote:

> On 11/5/11 10:25 AM, Jed Brown wrote:
>
>> On Sat, Nov 5, 2011 at 11:13, Brad Aagaard <baagaard at usgs.gov
>> <mailto:baagaard at usgs.gov>> wrote:
>>
>>    The error is in our inner solve (associated with computing the
>>    adjustment to the solution based on the friction criterion) for
>>    cases where not all processors have points on the fault. On
>>    processors without points on the fault the block size is -1 and the
>>    local size is 0. For processors with points on the fault, the block
>>    size is 3.
>>
>>
>> How is this happening? MatSetBlockSize_MPISBAIJ raises an error if you
>> try to change the block size. Are you messing with the PetscLayout
>> directly or something?
>>
>>
>>    I tried altering DMMeshCreateMatrix so that it uses a block size of
>>    1 if bs==-1, but I still get the error. Does the block size need to
>>    be consistent across all processors, even those a local size of 0?
>>    Alternatively, is the local size of 0 causing problems?
>>
>>
>> Yes, bs must be the same.
>>
>> Since bs must be set for PetscLayoutSetUp() to succeed, the fact that
>> you have bs=-1 means that function wasn't called. It could be memory
>> corruption or it could be some manual fiddling with the PetscLayout.
>> Somehow, you have bypassed several guards in the code that are supposed
>> to cause errors with invalid input like this.
>>
>> The local size of 0 should not cause a problem.
>>
>
> It looks like Matt is calling Sieve layout stuff in DMMeshCreateMatrix().
> I don't see any calls to MatSetBlockSize.


This is now fixed.

Problem: MatSetBlocksize() would check that blocksize was identical over
all processes, but the blocksize passed
               into Mat***SetPreallocation() was not checked.

Soluion: Put the check into the preallocation functions.

    Matt


>
> Brad

-- 
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20111106/80008985/attachment.html>