Section: New Results

Application level

In general purpose software engineering (as opposed to simulation software engineering), the motivations for reuse have long been advocated and demonstrated: lower risks of defects, collective support of potentially larger user community, lower development costs, and so on. In simulation software architectures, we can also cite business-specific motivations, such as providing a better reproducibility of simulation experiments, or avoiding a complex validation process. In [20] , we show that although it is rarely discussed, reuse is a problem that may be considered in two opposite directions: reusing and being reused.

DEVS is a formalism for the specification of discrete-event simulation models, proposed by Zeigler in the 70's, that is still the subject of many research in the simulation community. Surprisingly, the problem of representing the time in this formalism has always been somehow neglected, and most DEVS simulators keep using Floating Point numbers for their arithmetics on time values, which leads to a range of systematic errors, including severe ones such as breaking the causal relations in the model. In [16] we propose a new data type for discretized time representation in DEVS, based on rational numbers. Indeed, we show that rational numbers offer good stability properties for the arithmetics used in DEVS, with a limited impact on the simulation execution performance.

As a challenging example, we have pursued our work on pricing American multi-dimensional (so very computation intensive) options in finance. From our previous work that achieved pricing a 40-assets based american option within 8 hours of computation on a single GPU, the work in [9] allows us to reach approximatively one hour of computation time. For this, we run using active objects coupled with OpenCL codes, on 18 GPU nodes acquired from the Grid'5000 platform (the maximum amount of available GPU on Grid'5000 that we could book at once).

Moreover, the balancing of work is taking in consideration the heterogeneous nature of the involved GPUs, and is capable to harness the computing power of multi-core CPUs that also support running OpenCL codes. This parallel and distributed pricing approach is also extended in the forthcoming PhD thesis of Michael Benguigui: it successfully tackles the Value At Risk computation of a portofolio composed of such complex financial products.

Software Defined Networks (SDN) is a new technology that has gained a lot of attention recently. It introduces programmatic ways to reorganize the network logical topology. To achieve this, the network interacts with a set of controllers, that can dynamically update the configuration of the network routing equipments based on the received events. As often with new network technologies, discrete-event simulation proves to be an invaluable tool for understanding and analzing the performance and behavior of the new systems. In [17] , we use such smulations for evaluating the impact of Software-Defined Networks' Reactive Routing on BitTorrent performance. Indeed, BitTorrent uses choking algorithms that continuously open and close connections to different peers. Software Defined Networks implementing Reactive Routing may be negatively affecting the performances of the system under specific conditions because of it lack of knowledge of BitTorrent strategies.