[petsc-dev] Subcomms

[Mark F. Adams]

On Jan 27, 2012, at 8:22 PM, Barry Smith wrote:

> 
> On Jan 27, 2012, at 6:36 PM, Mark F. Adams wrote:
> 
>> 
>> On Jan 27, 2012, at 6:58 PM, Jed Brown wrote:
>> 
>>> On Fri, Jan 27, 2012 at 17:48, Matthew Knepley <knepley at gmail.com> wrote:
>>>  Is now the right time. Shouldn't we wait until MPI's replacement is working and do things with that model?
>>> 
>>> I'm laughing. Am I supposed to be?
>>> 
>>> I'm laughing too.
>>> 
>>> There isn't going to be a replacement for MPI until the smart people that understand parallel programming, performance, and libraries start working on something other than MPI. But most of those people are on the MPI Forum, trying to improve MPI. Now we need a good model for threads, and that might not be based on MPI, but it sure looks like the large-scale distributed-memory model will be MPI for the foreseeable future.
>>> 
>> 
>> I don't think its a matter of smart people not having worked on this, they have IMO, its just a hard problem.
> 
>   I disagree; it is not necessarily hard, it is just that the non-MPI people are pretty fucking stupid.
> 

It is not hard, intellectually intriguing, fundable, and smart people won't do it.  What am I missing?

>   Barry
> 
>> 
>>> 
>>> As for sources of parallel errors, yes, it's somewhat tricky, but as long as the model is to get a sub-object out of a bigger one (submatrix, coarse level, etc), I think we can manage it. At any particular time, the user should still be looking at essentially single-comm collections of objects, but not all processes will end up being called in every context.
>> 
> 
>