[petsc-users] Block size

[Barry Smith]

> On May 22, 2015, at 11:14 AM, Eric Chamberland <Eric.Chamberland at giref.ulaval.ca> wrote:
> 
> On 05/22/2015 11:36 AM, Barry Smith wrote:
>> 
>>    Eric,
>> 
> > to squeeze the largest problem you can run you would use the point format and it will be somewhat slower but at least you will not run out of memory.
> 
> Ok, my algorithm is already general and can do both.  But my main worry is:
> 
> Do all preconditionners are able to handle "sparse blocks" when block_size > 1 in {SEQ,MPI}AIJ matrix?  Do they still have some gain with "sparse blocks"?  

   This depends on the particular algorithm and the particular implementation. There is no general statement. Because of the various contending issues in terms of performance you can only try them and see what happens performance wise.

> (for example, algebraic multi-grid will be able to pick lines "per-block" when block_size > 1 but blocks are "sparses")?

   Yes, the GAMG (hypre BoomerAMG) does exactly that. 
> 
> Maybe the answer is trivial, but we were discussing it here and we got different positions on this..
> 
> Thanks,
> 
> Eric
>