High-performance computers simulate combustion and heat transfer in power stations

Mannheim, 20-6-97 The layout and optimisation of super power stations based on utility boilers moves from empirical and experimental to simulation-based methods. Industrial acceptance depends on a short turn-around time of the simulations. Uwe Schnell from the University of Stuttgart discussed at the Mannheim Supercomputer Seminar the parallelisation experiences of AIOLOS, a combustion chamber simulation program for utility boilers, that is based on finite volumes. Schnell uses loop-parallelisation with microtasking on the NEC SX-4. A problem with 50 transport equations and a discretization with up to 1,000,000 grid cells was demonstrated.

The simulation of such plants of 20 m x 20 m and a height of 110 m needs more than 500,000 grid points. Schnell's programme is optimised for the parallel vector processor NEC SX-4.

Several functions have been integrated into the program AIOLOS for 3-D simulation of pulverised coal combustion, such as flow, homogeneous and heterogeneous reactions, turbulences.

To reduce the compute time, higher-order methods and finer grids are used locally in interesting areas, the burner for example. The Fortran 77 program is based on loop-parallelisation with HPF or microtasking elements.

The program was ported on the NEC SX-4 at HWW (HPC for science and industry GmbH in Stuttgart) which has 32 vector processors, 8 GByte memory, an 64 GFlop/s peak performance. The programming models are direct message passing with MPI/SX as communications library, as well as HPF and NEC-specific multitasking.

Multitasking allows the subdivision of a programme and the distribution over different processors, using the concepts of the automatic microtasking parallelisation which is organised by compiler options. A preprocessor analyses the programme, looking for dependencies and generates automatically parallelisation directives. In contrast, macro- and microtasking realise the parallelisation via explicit subroutine calls (intrinsics). Analogous to microtasking, HPF supports the data resp. loop-parallel programming with an explicit data distribution.

Schnell also did write a version based on the NEC-specific MPI version - MPI/SX. Each processor gets own processes that can only access their own memory. Communication is realised with system subroutines. Here the programmer has to distribute the data and to program the communication explicitly.

AIOLOS has been used with different programming models on a Cray C90 with 4 CPUs. The NEC used both, microtasking and MPI. First results showed microtasking achieves a higher parallel efficiency. Because of the data structure - hard to change and load balancing problems - only the microtasking model was used. Schnell demonstrated the approach in a simple subroutine to compute the norm of the residual vector. The structure of the code remains unchanged. The programmer has to add some directives for vectorisation and parallelisation, because of indirect addressing, the compiler could not recognise dependencies and recursions automatically. Additionally synchronisation points were set with a barrier directive, if necessary.

Schell presented measurements for the boiler Fusina 2 of ENEL, an Italian based company. Fusina 2 uses a tangential firing with 16 burners on 4 levels delivering 170 MW electric power. The floor space is 8 x 12 m, the height 35 m. The simulation was realised for two different coal mixtures, South African and American coal. Comparison between the measurement and the simulated results showed a close correspondence.

The program achieved a good parallel speedup - nearly linear - with the fine grid, but the coarse grid did not run optimally on 8 processors, and on 16 processors no speedup can be seen. For the coarse grid the results are 40 minutes on one processor, 7 minutes on eight, reduces to 6 minutes on 16 processors - factor 5.7 and 6.7; the medium grid six hours 42 minutes, 52 minutes down to 31 minutes - factor 7.7 and 12.6; fine grid 22 hours 7 minutes, 2 hours 53 minutes down to 1 hour 36 minutes - factor 7.7 and 13.8.

Uwe Harms