[petsc-users] Tuning the parallel performance of a 3D FEM CFD code

[Henning Sauerland]

On 15.05.2011, at 15:59, Jed Brown wrote:

> This is a problem, but you can still order unknowns in the vector so that velocity and pressure all come together, then AIJ/Inode will treat the components together which will tend to make ILU stronger.

The results using the interlaced ordering are more pronounced than expected. The number of iterations did increase significantly using ILU(0) (~430 vs. ~170 per Picard iteration on 4 PEs) which makes it 4-5 times slower. However, the iteration count almost stays constant when moving to more processors! ILU(2) again reduces the iteration count (~100) while being slower compared to ILU(0).


I really appreciate your suggestions so far!

Henning