CERFACS CERFACS Scientific Activity Report Jan. 2010 – Dec. 2011

ADVANCED METHODS AND MULTIPHYSICS
expertise and with predefined mesh shapes choices following canonical CADs. Structured grids are suitable
to maintain mesh lines aligned with the flow anisotropy, especially in the boundary layer where the variables
are mainly varying in the direction normal to the wall. Moreover, structured numerical techniques help to
improve the global performance and robustness of the structured approach. The structured mesh generation
process must be performed by experts and the solution can be computed very accurately and quickly when
the mesh lines are aligned with the flow.
Nowadays, an industrial tendency is to increase the CAD complexity in order to compute flows with a
higher accuracy. For planes, one can for instance consider thermal computation between engine and nacelle
or flows around landing gears. For turbomachinery, passive wall treatments are added to increase the
stability area of the propeller system and cannot be meshed easily. For those examples, the global structured
mesh strategy fails at a reasonable human cost and it is clear that complex CAD must be computed with
unstructured capabilities. The unstructured mesh generation process is very efficient but the prize to pay
lays in the global accuracy of the computation : it is very difficult to impose mesh lines which are aligned
with the flow and the accuracy can finally be lower than for structured grids. This point has been highlighted
as a conclusion of the 4th Drag Prediction Workshop held in June 2009 : “More scatter from unstructured
methods than from structured grid methods. Suspect this is more due to grid than to code.” A way to
overcome this drawback is to authorize unstructured grids composed of different element shapes with
hexahedra and prisms in the anisotropy flow region, tetrahedra elements where the flow is isotropic and
pyramids in the buffer region from four-nodes element faces to three-nodes element faces. This is what is
called a hybrid mesh approach in the literature but we prefer to call them mixed-elements grids. Compared
with a structured grid, an unstructured equivalent induces a memory overhead due to connectivity and
indirect memory addressing and since the global mesh structure is left away, it is difficult to implement
efficient algorithms such as implicit schemes.
An interesting approach seems finally to associate both techniques. Therefore, we consider meshes
composed at the same time of structured blocks and unstructured mixed-elements zones : this is what we
call a hybrid approach. This approach corresponds to the current Airbus needs (and strategy) who would
like to benefit from the two worlds at the same time.
In that context and in collaboration with ONERA, CERFACS actively participates to the development of
unstructured capabilities in elsA. We are now able to simulate RANS three-dimensional unstructured grids
on parallel platforms. Advanced validations on industrial configurations are on-going. The next step consists
in handling hybrid grids. CERFACS is involved in the numerical treatment of interfaces between structured
zones and unstructured ones. This is a consequence of the impossibility to implement identical numerical
schemes for structured and unstructured zones. This activity can be related to coupling algorithms.
4.3.2 High Performance Computing (M. Montagnac)
The emergence of large-scale HPC computing platforms based on multicore and manycore technologies,
processor-based CPU and hybrid CPU / GPU, involves the modification of the legacy software to ensure
optimal use of these capabilities in the future. CERFACS is involved in these two aspects. The first point
is on hardware accelerators and especially GPU. elsA has been ported on GPU platforms after rewriting
some parts of the code in CUDA-C. Results show an acceleration factor of about ten between one GPU card
and one core of an Intel Westmere processor. Many GPU cards can be addressed for parallelism but more
validations are required to make conclusions.
The second point focuses on multicore hardware architecture. A simplified model of elsA has been
developed in order to assess many different techniques that aim to maximize both the individual CPU
performance and the shared-memory multicore performance. This effort will continue in the project
SONATE+.
164 Jan. 2010 – Dec. 2011