Patrick,
You are absolutely right (its a wonder it never hit us before). I have
updated the manual pages in petsc-dev to reflect this.
Barry
On Dec 12, 2012, at 2:56 PM, Patrick Lacasse <patrick.m.lacasse at gmail.com>
wrote:
> Hi guys,
>
> I got a dead lock by doing something like this :
>
> Vec a;
> [...]
> PetscInt n;
> VecNorm(a, NORM_INFINITY, &n);
>
> PetscScalar* b;
> Proc 0 -> VecGetArray(a, &b); [...] VecRestoreArray(a, b);
> Proc 1 -> nothing
>
> VecNorm(a, NORM_INFINITY, n);
>
> Into VecNorm,
> proc 0 stall in MPI_Allreduce
> while proc 1 just took the norm into its cache.
>
> What I understand is that PetscObjectStateIncrease has being called only for
> proc 0 so that proc 1 doesn't know that the norm must be recalculated.
>
> Should any function that call PetscObjectStateIncrease be consider logically
> collective?
>
> I have corrected my own code to use Vec(Get/Restore)Array collectively.
>
> Patrick Lacasse
>
>
