Convergence of ILU preconditioner with MPI and using IJ interface

[valonbb]

I am mainly using Krylov based solvers with ILU preconditioner and IJ interface. I have observed that for some input matrices,
there is a convergence issue when more than one mpi processes have been invoked. On the other hand, under the same conditions,
this seems to be more robust under the sequential operation - one processor. I have tried adjusting preconditioner parameters such as dropping threshold, level of fill and ilu type, but still did not have much luck with this. Is this something one would expect given that implementation of ILU is based on Domain decomposition framework and Schur complement? Would any better matrix partition help with this or any idea where things may have gone wrong? In contrast, it all works fine when PILUT and EUCLID preconditioner have been used.

Thanks a lot for all the great work you are doing to keep hypre going.

[oseikuffuor1]

@valonbb which options of ILU are you using? What you are seeing may not be unusual, particularly for the block Jacobi options. In this case, using options > 1 (such as 10, 11, 30, 31 ) for the ILU type could be better.

[valonbb]

Hi @oseikuffuor1,
Thanks a lot for your reply.

Actually, I have experimented with most of the ILU options. It is true that for some cases, options you have mentioned lead to a better convergence, but in my case I do not see that being applicable for any given number of processors. For example, a given 10 by 10 matrix with considerably well conditioned number, HYPRE_ILU works well for (number of processors - np) np = 1, 2 and ilu.type = {1, 10, 11, 31}, however when np = 3, result does not converge at all. Does this have to do with the way matrix is partitioned across the processors or what would be the explanation of this different behaviour with varying number of mpi processes (np)?

Thanks a lot for all your help - very much appreciate your time to look at this.