want to make some changes to PCASM

[Richard Tran Mills]

Folks,

I believe this is orthogonal to what Lisandro is discussing, but I am 
wondering if it might be worthwhile to add support in PCASM for subdomains 
being shared across a subset of processors?  That way subdomain solves could 
be shared across cores in a socket, processors in an SMP node, etc.  All the 
talk about the proliferation of cores in a socket, etc., suggests that this 
might be worthwhile, although I'm not sure.  In practice, I've found that for 
many of the problems I work with, one subdomain per processor still works OK 
even with tens of thousands of processors.  Still, this might be worth exploring.

I have not looked into the PCASM code to see what might be involved with this 
-- it's just an idle thought.  Can anyone more familiar with the issues comment?

--Richard

Lisandro Dalcin wrote:
> I would like to make some changes and enhancements in PCASM:
> 
> 1) Make it take full ownership (allocated array memory and IS
> references) of the subdomain index sets.
> 
> 2) Currently, if the matrix is symmetric, ASM switches to type
> PC_ASM_BASIC in PCSetFromOptions_ASM().
> I believe this should also be handled on PCSetup_ASM(). If not, we get
> different behavior if PCSetFromOptions() is never called.
> Moreover, I believe PCSetFromOptions_ASM() should call PCASMSetType()
> if the '-pc_asm_type' option is passed.
> 
> 3) If more than one block per processor is requested, then ASM
> currently does a row-based partitioning. For unstructured problems and
> LU-based local solves, this is going to be really bad, right?
> 
> I've already implemented an 'smart' subdomain partitioner based on
> MatPartitioningXXX. I'm thinking about adding a new utility routine
> (in the spirit of PCASMCreateSubdomains2d)
> 
> PCASMCreateSubdomains(Mat A, MatPartitioningType mptype, PetscInt
> local_blocks, IS *is[])
> 
> which could be called in order to get a good subdomain partitioning,
> and next pass it to PCASMSetLocalSubdomains().
> 
> 
> Does all this make sense?
> 
> 
> 


-- 
Richard Tran Mills, Ph.D.            |   E-mail: rmills at climate.ornl.gov
Computational Scientist              |   Phone:  (865) 241-3198
Computational Earth Sciences Group   |   Fax:    (865) 574-0405
Oak Ridge National Laboratory        |   http://climate.ornl.gov/~rmills