| |
THOR - Benchmarking a Multiblock Finite Volume CFD Code
Mike Ashworth, STFC Daresbury Laboratory
THOR [1] is a CFD package, which can be used to compute laminar and turbulent
flows in complex geometries, with or without chemical reactions. The finite
volume approach is used to discretise the governing equations expressed in
body-fitted curvilinear coordinates. The multiblock strategy is adopted to deal
with complex geometries and the code implements first and second order turbulence
models, including k-epsilon, k-omega, LRR and SSG. Different discretisation schemes,
including first order upwind and higher order QUICK, SMART and CUBISTA, are
implemented. Furthermore implicit first- and second-order time-dependent schemes
were added to THOR to model unsteady flows, such as oscillating turbulent flames,
large scale explosions and blood flows in veins.
THOR is parallelised with MPI. Domain decomposition is employed and by using
the multiblock structure in the code, the computational domain is decomposed to
multiblocks according to boundary conditions and processor number. Different blocks
can then be allocated to different processors. One block must be within one
processor but to enhance the efficiency, one processor can have more than one block.
Figure 1: Performance of the THOR code on the IBM p690 and p690+ and the SGI Altix 3700/1300 and 3700/1500 systems
The benchmark run here has 2619156 nodes and calculates 50 iterations towards
convergence of a turbulent flow using the k-epsilon turbulence model. Performance
of THOR on 32 and 64 processors is shown in Figure 1. Like PCHAN,
THOR contains
nearest-neighbour halo-type communications in order to update boundary data between
the partitions of the computational domain. It also has high memory access
requirements although whereas in PCHAN the access is sequential, the finite volume
data structures involve a more randomised sequence of accesses. Scaling from 32
to 64 processors on the IBM systems is poor. Both Altix 3700 systems outperform
the IBMs, e.g. by a ratio of 1.6 over the p690+ on 64 processors. The performance of
THOR on the HPCx IBM p690 system has also
been reported elsewhere [2].
References
[1] THOR-2D: A two-dimensional computational fluid dynamics code
X. J. Gu & D. R. Emerson, June 2000
Technical Report, Computational Science and Engineering Department,
STFC Daresbury Laboratory
[2] Application Performance on the High Performance Switch
Mike Ashworth, Ian J. Bush, Martyn F. Guest, Martin Plummer and Andrew G. Sunderland and Joachim Hein, 2004,
HPCx Technical Report HPCxTR0417
|
|
