- Transonic Aerodynamics
- External Aircraft Component Design
Models active in this Validation Case
- Compressible flow
- Energy models
- Ideal gases
- Mixed Precision
To profile the accuracy and performance of Envenio’s EXN/AERO manycore CFD solver on a challenging transonic aerospace test case. We undertook this demonstration to show that DES and LES are realistic options today for solvers that are properly optimized for low-cost heterogeneous computing architectures. Engineers working in aerospace have been seeking this capability for some time now in order to explore corners of the operating envelope, to understand complex interactions of wing components (e.g. flaps) and to identify ways to reduce noise.
Reference Case Description
The ARA M100 wing body geometry is based on a scale wind tunnel model and is referenced by NASA as a validation case for CFD codes with compressible / transonic flow capability. The configuration of the model is similar to what is seen in civil and military transport aviation and features a finite wing, a realistic wing root and a fuselage. Dimensions are shown in Figure 1.
Figure 1: ARA M100 general dimensions in mm. Image source refefenced in bibliography.
Surface Mesh: The surface mesh used to create the CFD grid is available on the nasa.gov site (link in bibliography). General dimensions are shown in Figure 1.
Coordinates & Domain Size
X direction: Roll axis, positive streamwise
Y direction: Pitch axis, positive to starboard
Z direction: Yaw axis, positive to top of aircraft
X dimension ~41 m
Y dimension ~21.75 m
Z-dimension ~41 m
Position of model:
Upwind face relative to inlet: ~20.5 m
Y position: Centred
Z position: Attached to symmetry plane
Mesh Resolution in the vicinity of the fuselage;
This 3D wing/body test case is at Ma.=0.803 and chord Reynolds number Rec = 13.1 × 106 (max chord = 0.378 m). The mesh has a y+ distribution as follows:
The mesh generator was instructed to use a wall function over the fuselage and direct solve-to-wall was employed on the wing. Freestream turbulence levels used as mesh generator inputs are Tu = 0.5% and µt/µ = 25.
Total Mesh Size: ~100.3 Million
Simulation Completion Criteria
The simulation is run until the coefficients of lift and drag achieve a time-averaged variation less than 1e-4 for CL and 1e-5 for CD. In this analysis the averaging window size, (4750 timesteps, 0.0095 seconds), is equal to the fuselage length (1.7m) divided by 0.7*Vfreestream,(0.7*257.4m/s = 180m/s) which is taken to be the convection time for the near-body fluid past the body. Successive averaging windows are computed 100 timesteps apart. The results of this analysis are presented in Figure 2. From this figure we see that CL converged in approximately 7200 timesteps and CD converged in approximately 8400 timesteps.
Time Step: 2e-6 seconds
EXN GPU Allocation: 6 Nvidia K80 GPUs (3x K80 cards, 2x GPU per card)
EXN CPU Allocation: 10 Intel Xeon 2.6GHz
X-axis orientation: Positive downstream
Y-axis orientation: Positive to the left of the body, looking downstream
Z-axis orientation: Positive upward, normal to ground plane
Turbulence Kinetic Energy: 0.0001 m2/s2
Turbulence K Dissipation: 0.0003 m2/s3
Wall model: Smooth wall
Outlet: Constant Specified Pressure
Initial Velocity: [257.1, 0, 12.9] m/s
Angle of Attack: 2.870 deg
Mach Number: 0.803
Temperature: 255 Kelvin
Pressure: 315.98 kPa
Initial Velocity: [257.1, 0, 12.9] m/s
Initial Turbulence Kinetic Energy: 0.0001 m2/s2
Initial Turbulence K Dissipation: 0.0003 m2/s3
Turbulence Model: DES SST k-omega
Flow type Compressible, Ideal Gas (R = 286.9 J/kg K)
Constant Viscosity: 1.715 x 10-5 kg/m s
Precision Mixed precission 93.3M Single precission, 7.0M Double precission
Body convection time: 0.0095 seconds (4750 time steps)
Total simulated time: 0.01842 seconds (9210 time steps, 1.94 washthroughs)
Mesh Topology Structured multiblock, one-to-one connections at block interfaces, data written as structured arrays in CGNS format
Simulation Outcomes, Timing, and External Factors
The pressure coefficient profiles at different spanwise locations along the wing, namely at span-normalized cross section y/B = 0.123, y/B = 0.231, y/b = 0.325, y/B = 0.455, y/B = 0.633 and y/B = 0.817 are presented in Figures 2 thru 7. The pressure distribution across upper wing surface and fuselage is shown graphically in Figure 8. Table 1 shows simulation performance outcomes and Table 2 shows approximate simulation cost information.
Table 1: Simulation performance outcomes
|Simulation starting conditions||
Simulation initial conditions were interpolated from a previous
coarse mesh, 25M control volume, simulation.
|Time to 1st wash-through||
Simulated time of 0.0095 sec requires 4750 time steps.
This is equivalent to real-time ~70 hours (2.9 days).
Time to completion
(CL & CD criteria)
Simulated time of 0.01842 sec requires 9210 time steps
This is equivalent to real time ~136 hours (5.6 days).
|Real time per time step||53sec|
|CPU type||Intel Xeon @ 2.6GHz|
|GPU type||NVIDIA Tesla K80|
|GPU cores||6 x 2496 CUDA core per GPU|
|Available Memory||128GB system, 24GB each K80 card (12GB each GPU)|
Table 2: Simulation cost information, assuming an owned system and a single license of EXN/Aero, operating year-round; figures in US dollars.
|Req’d CPU Processors 192 192 10||192||10|
|Req’d GPU Processors||-||6|
|Req’d System Hours||720||135|
|Cost per system-hour||$17.28||$7.50|
|Total Compute Cost||$12,441.60||$1,012.50|
|Pro-rated license (req'd hrs / 8750 hrs) *annual license||$23,918||$462.86|
|Cost of Simulation||$36,359||$1,475|
Figure 2: Convergence of lift and drag coefficient during the the EXN/Aero simulation.
Figure 3: Comparison of EXN/Aero simulation and experimental results at y/B=0.123
Figure 4: Comparison of EXN/Aero simulation and experimental results at y/B=0.231
Figure 5: Cmparison of EXN/Aero simulation and experimental results at y/B=0.325
Figure 6: Comparison of EXN/Aero simulation and experimental results at y/B=0.455
Figure 7: Comparison of EXN/Aero simulation and experimental results at y/B=0.633
Figure 8: Comparison of EXN/Aero simulation and experimental results at y/B=0.817
Figure 9: Pressure coefficient contours at the cross section
The comparison of pressure coefficient profiles with the experimental profiles reveals that our model matches closely with the experiments in subsonic regions of the wing. For the two outboard wing sections (y/B 0.633) the Cp distribution is accurately represented by simulation data. For inboard sections (y/B < 0.633) the drop in Cp is delayed and smoother than the experimental data. Inset images from a Cobalt simulation using the Spalart-Almaras RANS model show an early drop in Cp on inboard sections. It is expected that mesh refinement in the vicinity of the shocks will bring EXN/Aero results more in line with experimental data sets; this work is planned as part of Enveino’s QA process and will be updated regularly on Envenio’s wiki site (accessible on envenio.ca).
We completed the 100 million node transonic DES simulations for the ARA-M100 wing body in 5.9 days. It had a Reynolds number of 13.2M. It ran on one of our desktop compute nodes and the computing burden was shared by 3 Nvidia K80 cards and 6 CPUs.
Recall that this is run on 3 Nvidia K80 GPU cards and 6 Intel CPU on a desktop-scale computer, which was the minimum hardware arrangement. Even in the minimum arrangement, this result is fast by DES standards. A properly resourced simulation (e.g. a desktop equipped with 6 or more K80 GPU) has the potential to be much faster.
- External Flow
- DES SST k-omega
- Energy Models
- Double Precision
- Integrated Boundary Values
- External Flow
- Lift & Drag
- Vehicle Maneuvering
- Transonic Flow
- Wing geometry available from: https://cfl3d.larc.nasa.gov/Cfl3dv6/cfl3dv6_testcases.html
- Figures 3-8 insets from: https://cfl3d.larc.nasa.gov/Cfl3dv6/3DTestcases/ARA_M100/compare_cp_m100.gif
- ARAA-M100 Diagrams (Figure 1) http://www.memoireonline.com/05/12/5815/m_Calcul-des-performances-aerodynamiques-de-la-configuration-aile-fuselage-Ara-M100-par-maillage-hybr27.html