During the last three and a half years, CFX has actively participated, as CFD software provider, in the European HPNURSA (High Performance Numerical Unsteady Rotor Stator Analysis) research project. The project, which started in 1998, recently closed with a workshop at the 21st IAHR meeting held in Lausanne, Switzerland in September 2002. Within the project, CFX worked closely with VA TECH HYDRO, as industrial end-user, and the Universities of Lausanne (EPFL) and Lyon (ECL) as academic partners.

The complexity of unsteady rotor-stator flows makes them highly computer intensive, as they require a transient simulation with sufficiently small time steps to resolve the unsteady features of the flow. Furthermore, the full 360-degree machine must generally be modelled, leading to multiple blades per row in the simulation. These demands increase CPU and memory requirements by at least an order of magnitude compared to steady-state simulations. If such computations are to be tractable for industrial users, the use of advanced computer and programming technologies is essential. These issues have been addressed by the HPNURSA project, providing improvements to the unsteady time integration formulation, enhancements of the transient rotor-stator interface and the optimization of the parallel performance of CFX-TASCflow and CFX-5.

A regular bottleneck for transient rotor-stator applications is in computing the geometric interface connectivity between the rotating and stationary portions of the grids. This calculation has to be performed for every time step. Studies produced an optimized algorithm which resulted in a speed-up by a factor of almost ten for large test cases. The overall effect on computing time is a reduction of around 10-15%.

The principal challenge in high performance computing of unsteady rotor-stator applications lies in the scalable extension of the software parallelization, which must cope with the changing connectivity due to the relative motion between the domains. With CFX-TASCflow, we could only achieve a moderate speed-up due to its grid block-based parallelization paradigm and its existing data structures. We therefore decided to concentrate on CFX-5, which has been designed from the outset as a scalable parallel code. Both, the CPU and the memory scalability have been optimized and are now comparable to those for single-domain computations.

As the industrial partner in the project, VA TECH HYDRO applied the enhanced software successfully to a transient analysis of a pump-turbine operating in pump mode. This machine consists of five runner blades and 44 stator blades (22 guide and 22 stay vanes). The transient analysis was performed for a reduced configuration with a simplified inlet and outlet region with one runner blade and eight stator blades (about 1.5 million nodes), as well as for the full machine including the draft tube and the spiral casing. For the full configuration, a total of 5.1 million grid points were employed, with a hexahedral grid around the blades and in the draft tube, and a hybrid unstructured mesh for the spiral casing. Better parallel scaling is obtained for the larger grid. Work to improve the performance for even greater numbers of processors is under way.

The unsteady pressure coefficients, velocity vectors and flow angles, which enclose the experimental data, show significant temporal variations. The time-averaged solutions are in very good agreement with the experiments. The superior parallelization capability of CFX-5 has proven to be scalable for both CPU and memory, making it highly effective for rotor-stator modelling.

Parallel speed-up for HPNURSA pump-turbine configurations (click for large image).

Results and experimental data for HPNURSA pump-turbine (click for large image).