[Trilinos-Users] crash using Amesos_superludist
Travis Austin
austin at txcorp.com
Thu Mar 14 12:07:44 MDT 2013
Hi Eric,
>> One specific question I have is whether superLU_dist requires the
>> global row numbers to be contiguous. I ask because I do not see this
>> crash in problems where that is the case.
> I do believe that superLU_dist requires the row numbers to be
> contiguous, though if the matrix is redistributed before the
> factorization, the matrix values will be imported into a matrix with
> contiguous, linear maps. But this isn't really an issue since it seems
> Amesos is exiting before even reaching this step.
We have diagnosed our issue to the fact that the row numbers are not contiguous
and it appears when digging into Superlu_dist that it needs that or else it can have
memory issues (stepping outside of array bounds).
When you say redistribute are you referring to a routine in Epetra_CrsGraph or
Epetra_CrsMatrix? I see that Isorropia performs redistribution of Epetra_Matrices
and vectors but we would prefer to avoid that since we are in the last several weeks
of a code release and we would have to deal with any new issues coming from
Isorropia.
The issue is that we have a column index like [7 8 9 11 12 13] but the global number
of columns is 6 so in superlu_dist it allocates an array of size 6 but tries to access its
12th entry.
How can we easily change our Graph from [7 8 9 11 12 13] to [0 1 2 3 4 5]? Any ideas?
>
> Have you tried the SuperLU_Dist interface in Amesos2?\
Same as above. We do not have the time now to look at Amesos2. Possibly in the future.
>
> `~Eric
>
> _______________________________________________
> Trilinos-Users mailing list
> Trilinos-Users at software.sandia.gov
> http://software.sandia.gov/mailman/listinfo/trilinos-users
More information about the Trilinos-Users
mailing list