[petsc-users] a direct solver with support for 64-bit integers

[Barry Smith]

On Mar 24, 2010, at 2:20 PM, Jed Brown wrote:

> On Wed, 24 Mar 2010 13:41:03 -0500, Hari Krishna Kodali <hari at iastate.edu 
> > wrote:
>> I have been using superlu_dist and ran into problems.  My two
>> dimensional, structured grid, quadrilateral elements, finite element
>> code creates a square matrix of degrees of freedom (dof)
>> 1,200,000. superlu_dist works fine with this, but when i try a  
>> problem
>> of dof 1,500,000 there are errors.
>>
>> The dof is well below 2 billion but the number of nonzeroes in the
>> matrix may have exceeded 2 billion which could be causing the  
>> problem.
>
> Ah, I see.  MUMPS and PaStiX both claim to support 64-bit integers,  
> but
> I don't think --download-mumps or --download-pastix currently support
> this.  In the case of MUMPS, I appears rather involved to get BLACS  
> and
> SCALAPACK built (portably) with 64-bit integers since they were  
> clearly
> not written with this capability in mind.

    I investigated this and it is all crap stuff. Compile the Fortran  
with -i8 and other nonsense.

    This is becoming a serious hole in the direct solver community  
which none of them has yet portably and properly overcome.

    Barry

>
> Jed