Parallel running needs more iterations?
 

Dear OpenFOAM community,

I am testing parallel scalability on different machines and I found that sub-linear scaling may be caused by an increase of iterations.

In the attachments you can find my study based on 'cavity' case:
- iterations.png: Comparison of number of iterations: single/parallel
- iterations-case.zip: OpenFOAM 1.7.1 case (If you want to repeat my test, you will need gnuplot for plotting. Calculation should take less than 2 minutes.)

Questions:
1. Why does the solver need more iterations when solving a case in parallel?
2. Do you experience the same behavior in your CFD problems?
3. How can we change convergence criteria in order to have the same (or approx. same) number of iterations for the parallel run as we had for the single run?
4. Are there other possibilities to reduce this undesirable effect?

Thanks,
Zoltan

Does "pIters" refer to the iterations on the pressure equation?

Yes "pIters" refers to the number of iterations on the pressure equation.

Máté (a colleague of Zoltán, this is the reason I know the answer as well)

with assumption that program has no bugs related to parallel programing.


Questions:
1. Why does the solver need more iterations when solving a case in parallel?

It depends on your matrix solver for pressure correction equation. Some matrix solvers require more iterations in parallel mode.


2. Do you experience the same behavior in your CFD problems?

It is known problem.

3. How can we change convergence criteria in order to have the same (or approx. same) number of iterations for the parallel run as we had for the single run?

Again it all depends on what matrix solver you are using. In some cases , it is not possible to change this behaviour. You have to increase iterations in matrix solver.


4. Are there other possibilities to reduce this undesirable effect?

yes, www.columbia.edu/~ma2325/poly.ps