[petsc-dev] Apparent bug with PGI on Titan with each thread calling a serial LU solve

[Barry Smith]

On May 7, 2014, at 9:09 PM, Mark Adams <mfadams at lbl.gov> wrote:

> 
> 
> 
> On Wed, May 7, 2014 at 1:02 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
>    Mark,
> 
>    Are you sure the PGI version is built with the correct branch of barry/make-petscoptionsobject-nonglobal ?
> 
>     In that branch line 516 of aoptions.c has         case OPTION_INT:   not a PetscFree
> 
>   Meanwhile in master line 516 has   ierr = PetscFree(PetscOptionsObject.title);CHKERRQ(ierr);
> 
> 
> Whoops, I was under the impression that your fixes had been merged with master.  Did I get that wrong ... How should we proceed?  I build PETSc on Titan for Ed so we are flexible.

   Unfortunately fixing PETSc master to support both thread safe and multiple thread within each object is complicated Ed needs to use the branch listed above. Hopefully this summer we’ll get the thread safety into Master but sadly it is not a trivial operation.

   Barry

>