[dominic]

Hi,

I would like to know if there are MPI communications within a BCGSTAB iteration ? Seeing through the code, I don't seem to notice that. I can see that all partitions do individual solves and then communicate.

For e.g. in solver_direct_mean.cpp
void CEulerSolver::ImplicitEuler_Iteration(), the Set_MPI_Solution is called at the end, after the solution is computed. May be your convergence would improve if the communication is placed inside the BCGSTAB routine as well ? but many increase your total communication time. - Is this why it was done this way ?

Dominic

[fpalacios]

Originally Posted by dominic

Hi,

I would like to know if there are MPI communications within a BCGSTAB iteration ? Seeing through the code, I don't seem to notice that. I can see that all partitions do individual solves and then communicate.

For e.g. in solver_direct_mean.cpp
void CEulerSolver::ImplicitEuler_Iteration(), the Set_MPI_Solution is called at the end, after the solution is computed. May be your convergence would improve if the communication is placed inside the BCGSTAB routine as well ? but many increase your total communication time. - Is this why it was done this way ?

Dominic

Yep, there is some communication inside de algorithm. In different parts... the most clear ones are when you call the matrix vector multiplication mat_vec(phat, v); or dot product dotProd(r_0, v).
For example the mat_vec is done in

void CSysMatrix::MatrixVectorProduct(const CSysVector & vec, CSysVector & prod, CGeometry *geometry, CConfig *config) {

Cheers,
Francisco

[dominic]

Thank you Francisco, I missed that part ! I can see that there is a
SendReceive_Solution() inside the function you've mentioned.