1. Use-case: CFD software with FEM and unstructured meshes
Q. V. Dinh and M. Ravachol
Engineering / R& D and Future Business
Dassault Aviation

2. Rationale for road-map in Mont-Blanc
AeTHER, in-house CFD software
Large and old (1985) industrial code
Legacy: 500 routines, 130,000 lines of Fortran 77 & 90
FEM formulation + unstructured meshes discretization
Data-bound code: lots of indirection…
Parallel implementation
MPI + local vectorization using coloring techniques
Too premature and time-consuming to port AeTHER…
Port a mini-application, representative of AeTHER
MiniFEM-asm
Standard FEM assembling phase
Using the same unstructured meshes as AeTHER
Developed for: MPI bulk synchronous => MPI + local task-based parallelism

3. Application to be tested
miniFEM-asm
Which problem is the application trying to solve?
A “surrogate” for industrial CFD code AeTHER, which mimics:
FEM assembling phase on representative unstructured meshes
Programming languages
C, C++, Fortran.
< 3000 lines of code
Parallel programming models
MPI + task-based parallelism using Divide & Conquer (D&C)
For tasks: Intel Cilk, OpenMP planned, OmpSS ?
Accelerator programming model: none
Libraries
For D&C : DC_lib (open-source, DA & UVSQ), METIS

4. Resources required Memory requirement per node (GB) Storage required
Shared storage
Private / temporary storage
Minimum number of nodes / CPU
Not relevant yet on Mont-Blanc platform, however…
Crude evaluation of memory reqs. on Xeon Phi
1 x 60-cores KNC : 1 Million nodes mesh
4 x 60-core KNC : 7 Million nodes mesh

5. Scalability Not relevant yet on Mont-Blanc platform
Scalability assessment available for Xeon and Xeon Phi
see companion presentation

6. Energy consumption
Not relevant yet

7. Comparison to other platforms
Compare performance (and energy consumption ?) to other platforms
Xeon and Xeon Phi: to be done
Results on Xeon and Xeon Phi available
see companion presentation

8. Feedback on platform usability
Not relevant yet