Pipistrel, a Slovenian light aircraft manufacturer, uses computational fluid dynamics (CFD) simulations in order to design its aircrafts. Pipistrel is currently limited to a small in-house cluster, that is why it is exploring different options in order to increase the fidelity and reduce the calculations time of its CFD simulations. Pipistrel’s main CFD software is open source software called OpenFOAM. Pipistrel entered MIKELANGELO project, to explore the possibility to use Cloud infrastructure during aircraft design. The OpenFOAM aerodynamic use case will leverage MIKELANGELO’s virtualization stack to flexibly compute aerodynamic maps for aircraft designs in the cloud.
On one hand the aerodynamics use case plans to use cloud computing in order to run as many similar OpenFOAM simulations simultaneously as possible. On the other hand it plans to run a single computationally intensive simulation using all available cores in HPC cluster.
The former need arises in an industrial environment when a particular aerodynamic configuration, such as an aircraft, sailboat or a car needs to be analysed under a large set of conditions. Such set of conditions can include different airflow incidence directions (angle of attack and sideslip angle), air properties (density, viscosity) or even propeller thrust. While the variation of a single parameter requires only a couple of dozen simulations to be run, the number of simulations increases exponentially with the number of simultaneous parameters to be varied. Such an application is therefore very well suited for cloud computing, as a single simulation may not be very demanding (can be run on maximum one node), but the need exists to run a large number of them on independent virtual machines.
On the other hand running a single computationally demanding problem is needed when a set of parameters is already chosen but a physically or numerically more precise simulation is required. A large and diverse aerodynamic body where a certain accuracy still needs to be assured can lead to a computationally intensive simulation as well. OpenFOAM is based on Finite Volume Method (FVM) that needs a mesh (an assemble of 3D cells that represents a volume around an aerodynamic body) within which the airflow is being simulated. A large and geometrically diverse object therefore needs a large mesh (large number of cells) in order to satisfactorily describe its shape. Such a simulation can become even more demanding when incorporating more accurate physical models or numerical schemes. Even though this is a typical HPC problem, it is still planned to be used as a baseline in order to increase the virtualised I/O efficiency of MIKELANGELO cloud stack.
Pipistrel plans to study two different aerodynamic geometries under the MIKELANGELO project. Both will be run on as many different sets of parameters as possible in a simultaneous manner, while only the computationally more demanding one will be prepared to be run in parallel on multiple nodes.
The simpler one is a 2D airfoil depicted in Figure 1. The shown contour presents a 2D shape of a wing as seen in a cross-section at some chosen point along the wing. A complete airplane wing therefore consists of an array of airfoils along the wing. During an aerodynamic design of a wing the designer must first know the characteristics of its basic building blocks, the airfoils. The designer must know how the airfoil behaves at different incidence angles (angles of attack) and different flow properties. For each airfoil he typically needs its coefficients of lift, drag and moment. Only then will the designer be able to estimate how the wing will behave in different flight regimes (take-off, cruise, landing). For an incompressible steady state flow, the same as Pipistrel is going to use, all flow properties (velocity, density and viscosity) can be combined into a single parameter called Reynolds number (Re). In this way a single parameter is needed in order to describe all flow properties.
An example of a typical lift coefficient dependence with respect to the angle of attack at a single Re is depicted in Figure 2. Although there is a continuous curve presented on the plot, in reality the designer is limited in time and computer resources and is therefore not able to calculate a desired number of points on the curve. The computational complexity scales exponentially with the number of additional parameters introduced such as Re.
Figure 1: An example of a 2D airfoil.
Figure 2: An example of a lift coefficient to angle of attack dependence.
One of the objectives of this use case is therefore to run a large number of simulations simultaneously to obtain results at all needed sets of parameters, that is angle of attack and Re, at approximately the same time. Pipistrel’s in-house cluster currently enables 16 simulations at a time, if each one is run on a single core. Using the MIKELANGELO stack Pipistrel will be able to employ much larger machines, which will allow a larger number of simultaneous simulations. Even more important, Pipistrel will gain the know-how to use not only HPC, but also the cloud-based hardware and software, which will introduce greater flexibility to the workflow and possibly also reduce fixed operating costs in the future.
The second case to be studied under the MIKELANGELO project consists of a wing and a larger number of propellers in front of the wing (Figure 3). A distributed propulsion system will therefore be studied. Parameters of interest are beside angle of attack and Re also the propeller’s position and its thrust. Besides computationally more intensive simulation with respect to the 2D airfoil case, there is also a larger number of parameters to vary. The plan is to start with a single propeller and corresponding wing section (Figure 4) in order to be able to run the problem on a single node and to finish with a complete wing run on multiple nodes. Intermediate steps will consist of gradually increasing the mesh size and the physics of the problem. The final full wing simulation will be run with a single set of parameters, chosen according to the lessons learned from previous steps. The objective is therefore to study the position and thrust of propellers in order to obtain satisfactory wing flight characteristics.
Figure 3: Complete wing with several propellers in front.
The objectives of a single propeller case are similar to the 2D airfoil case, that is to increase the number of cases run simultaneously to accelerate each simulation with MIKELANGELO stack optimization and to improve the agility of application deployment. The latter consists of replacement of current bash scripting with a more efficient, simpler and user-friendly GUI interface that will allow the user to choose the parameters to vary with the corresponding values. On the other hand, the objective in the case of the final full wing simulation at a single set of parameters run in parallel on several nodes is to show the increase of efficiency in virtualised I/O of MIKELANGELO stack.
Figure 4: A single propeller with a corresponding wing section.