dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, May 14, 2007 at 10:03:04PM -0500, Andy Terrel wrote:
> Hello,
> 
> So I have been trying to run some simulations in parallel, when I 
> noticed that I was getting the same timings and iteration counts on 4 
> processors as on 1.  It seems that all the Petsc matrix, vector, ksp, 
> ... objects are created using PETSC_COMM_SELF, and thus I was running 
> not 1 simulation on 4 processors but 4 simulations on 4 processors.  
> When I go to change the PETSC_COMM I am confronted with lots of errors 
> (mostly about getting non-local rows probably in the PETScMatrix.nz() 
> function.  I also see some comments such as:
> 
> // FIXME: Temporary fix (assumes uniprocessor case)
> 
> I am just wondering if anyone else is running into these sorts of errors 
> or if I have horribly misconfigured something while building things. I 
> was using the 0.6.4 release because I wanted a stable release that I 
> could use on a large cluster, but it seems that the same code is in the 
> dev repo.
> 
> Thanks,
> Andy

Hey Andy,

Here's the situation... DOLFIN uses PETSc for linear algebra (if you
configure it that way) which means DOLFIN runs in parallel
*in principle*, but no one has bothered to add the final touches to get it
running in parallell *in practice*. The reason is we don't yet have
parallel assembly. Once that's in place, we will fix the missing
pieces for parallel linear algebra.

I can't say exactly when this will happen, but it should be soon, like
within 3-4 months.

/Anders