A study to provide a set of guidelines for calculation and cost-efficient of numerical algorithms in the cloud with security concern.

0.  

   If only integers are manipulated, portability of numerical software is not a big deal. When floating-point arithmetic is used, the situation is quite different. Although the IEEE 754 Standard for Floating-Point arithmetic [1] has greatly improved a situation that was disastrous before the 90’s (see [2]), there are still some problems that cannot be ignored, especially when writing soft- ware for critical applications. Some horror stories (along with recommendations) can be found in [3] by Monniaux, and in Chapter 6 of [4].

   Now Cloud infrastructure are common, nevertheless the problems have been propagated (even amplified) to this kind of architecture as a Service.

   Many application take benefit of arithmetic libraries. In this study we plan to focus on a specific application. Through this use case we will be faced on a real application and we will provide a real experiments to validate the numerical behavior and the Cloud optimization.

   The selected application from Bosch is a large, real-time capable, computation-intensive and optimization-based predictive "eco-driving" application. The goal of this software-function is to save energy, and this is reached, knowing the information of the road ahead, by anticipating the speed profile that minimizes as much as possible braking and acceleration. At the same time, this way of driving has another benefit : it reduces vehicle wear and tear (fewer breakdowns etc.) which again saves money. Relying on vision, human drivers scan relevant features of the road ahead and with or without the help of maps and other ADAS systems are able to anticipate but due to a fuzzy internal representation of the car dynamic not to optimize very well the speed profile.

   That is the motivation for optimal predictive eco driving-functions. The main idea is to online plan the speed trajectory that minimizes a metric which takes into account energy consumption, travel time and driver comfort over a certain distance or time horizon, while respecting look- ahead information of the road, vehicular traffic etc. . The car and its components (battery, electric machine) dynamics and restrictions are represented with a mathematical model. As the car is moving and the horizon is related to the actual position, the optimal speed trajectory is always re-calculated. A first version of this function has been implemented for an embedded device, but now an extended version shall be implemented using the compute capablilities of the cloud

   1. [1]  IEEE. IEEE Standard for Floating-Point Arithmetic (IEEE Std 754-2019). July 2019.

   2. [2]  W. Kahan. Why do we need a floating-point standard ? Technical report, Computer Science, UC Berkeley, 1981. Available at http://www.cs.berkeley.edu/~wkahan/ieee754status/why-ieee. pdf.

   3. [3]  D.Monniaux.Thepitfallsofverifyingfloating-pointcomputations.ACMTOPLAS,30(3):1–41, 2008. A preliminary version is available at http://hal.archives-ouvertes.fr/hal-00128124.

   4. [4]  Jean-Michel Muller, Nicolas Brunie, Florent de Dinechin, Claude-Pierre Jeannerod, Mioara Joldes, Vincent Lefèvre, Guillaume Melquiond, Nathalie Revol, and Serge Torres. Handbook of Floating-Point Arithmetic, 2nd edition. Birkhäuser Boston, 2018. ACM G.1.0 ; G.1.2 ; G.4 ; B.2.0 ; B.2.4 ; F.2.1., ISBN 978-3-319-76525-9.

0. This work is part of a collaboration between Bosch and the Inria Aric and Avalon teams.

   The work will be under the supervision of Christian Bertsch (Bosch Research), Eddy Caron (Avalon Team. Professor University Lyon1), Nicolas Louvet (Aric Team. Associate professor at University Lyon1). The localization of this position will be at LIP Laboratory, ENS de Lyon.

Before a most generic study we want to explore differents problems with a specific application, the "eco-driving" function described above. This application uses different numerical libraries. Based on a previous internal study have shown a different numerical behavior according to the hardware and to the cloud environment.

The first step will be to evaluate the numerical properties of the application deployed in a Cloud environment. In addition we want to evaluate and forecast the Cloud cost. Based on this, the setup in the Cloud shall be optimized. Tune and scale the deployment on different Cloud providers to deal with the portability and find the best fit accross different Cloud Service Providers and by the way taking into account different underlying hardware. As we have seen these choice must be considered as a point of view of numerical stability and in term of performances.

Beyond the IaaS (Infrastructure as a Service) consideration we want to investigate different configuration possibilities and usage of Bosch-internal and external numerical libraries (e.g. for elementary functions).

We expect that this study will provide a set of generic guidelines for calculation and cost- efficient of numerical algorithms in the cloud.

Moreover a specific focus on security aspect around the platform deployment on Cloud will be a main topic of the work.

The following compepencies will be apprciated

• —  Strong background in computer science

• —  Knowledge in cloud technologies / large computing infrastutures / distributed systems / Cybersecurity

• —  Numerical properties of floating point representations and calculations

• —  Numerical mathematics