Successfully navigating spacecraft safely back into the Earth’s atmosphere is one of the pinnacles of human engineering, where the slightest miscalculation can lead to catastrophe. Marco Cisternino of scientific computing SME OPTIMAD has been working alongside PRACE as part of its SHAPE Programme to try and optimise the codes needed for this endeavor for state-of-the-art parallel computing architectures.
One of the most challenging aspects of space travel is the re-entry of spacecraft into the Earth’s atmosphere. Re-entry happens at such high speeds that the friction from the gas molecules heats the capsule to extremely high temperatures and causes high levels of turbulence. A control system is needed to guide the trajectory of the capsule as it enters the atmosphere, and a heat shield is needed on the side that enters first.
From an engineering perspective, the process of re-entry is complex. At its core is fluid dynamics, but other variables such as the combustion of gas and the surface of the capsule must also be taken into account. Even trickier is the fact that at the first point of re-entry, the air is so diluted that a standard statistical description of it cannot be applied. Instead, expensive and complex computational codes must be used to describe each molecule individually. Simplified models can be used in their place, but have a number of limitations.
Rapidly improving computational resources are opening new possibilities for industries and SMEs in the aerospace industry. The aforementioned codes for rarefied gas dynamics are particularly challenging because of their high number of degrees of freedom. Without particular care in the implementation, the execution time of such codes becomes prohibitive. For this reason, an efficient and highly scalable parallelisation of the code is needed to dramatically reduce the overall computation time.
New multi-integrated core architectures for high performance computing allow for the use of a high number of cores for parallel codes. However, the implementation of such codes requires expertise in high performance computing and even then it remains challenging to efficiently exploit these architectures for numerical simulations. Indeed, such expertise is rarely ever present within an SME. However, PRACE’s SME HPC Adoption Programme in Europe (SHAPE) has been supporting one such SME involved in the aerospace industry to facilitate its deeper understanding of how to use HPC effectively.
OPTIMAD is a company that develops software for scientific computing, mainly in the field of aerodynamic analysis. Founded in 2006, it was started by a group of fluid mechanics and design optimisation researchers as a spin-off of the Department of Mechanical and Aerospace Engineering of the Polytechnic University of Turin. In 2015, they spent three months working on the PRACE SHAPE project RAPHI, a feasibility study for porting KOPPA (a state-of-the-art code used for rarefied gas dynamics) to the Intel Xeon Phi architecture.
Marco Cisternino worked together with Florian Bernard of the INRIA Sud Ouest Memphis team and the Cineca computing centre to create a proof of concept that the cost of running a code such as KOPPA using heterogeneous computing architecture could be radically reduced. “This was a special project, because we were not merely looking at ‘wall time’,” says Cisternino. “The aim was to also reduce the overall cost of computation in terms of energy.”
As a company, what the SHAPE project has taught us about working with these kinds of computing architectures is invaluable.
KOPPA (Kinetic Octree Parallel PolyAtomic) is a parallel numerical code which is used for the simulation of rarefied gas dynamics. It is based on a library named PABLO (PArallel Balanced Linear Octree) used to manage octree grids in parallel. The main issue with such numerical codes is the very high execution time which can become prohibitive for some industrial applications. Thanks to the SHAPE project, important improvements have been achieved with respect to execution time and scalability. In particular, some parts of the code have been re-implemented to better suit multi-integrated core architectures. So far, the computational time requirements have been decreased by a factor of almost eight and a good scalability has been obtained up to 64 processors against 16 initially.
The SHAPE project is now complete, but the impact of the results on OPTIMAD reach further than just the provision of a proof of concept. “Many of the things we have learned in the course of the SHAPE project about this specific application have been applied to all of our other codes,” explains Cisternino. “As a company, what the SHAPE project has taught us about working with these kinds of computing architectures is invaluable.”
OPTIMAD has since been involved in a Horizon2020 project called AeroGust in which they used their expertise about high performance computing to help with the optimisation of aeroplane shapes for dealing with gusts. “Heterogeneous computing architectures are seen by many people working with simulations as a burden,” says Cisternino. “If you do not approach them in the right way then you can end up paying a lot in terms of the maintainability of your codes. We were brought into this project to bring some of the insight and knowledge we have gained about striking a good compromise between performance and maintainability of codes.”
For more information: www.optimad.it
Resources awarded by PRACE: Marco Cisternino was awarded 50 000 core hours on Marconi hosted by CINECA, Italy