Section: Application Domains

Simulation

Discrete-event simulation

Simulation is an example of an application with ever increasing computation needs that would benefit from the Scale research results. In emergency planning and response, for example, users need to access the power of large scale distributed computing facilities to run faster than real-time simulations of the situations they face on the field; Such a computation can mix heterogeneous distributing computing platforms (PDA and laptops on the field, Cloud and HPC in background) and use a number of external services (eg. weather forecast).

Simulations made of multi-party contributed software models also demonstrate the need for a unifying and user-friendly programming model. Indeed, since the early 70's, the simulation field have been the subject of many efforts in order to abstract the computation models from their actual application domain. DEVS (Discrete Event Systems specification), is an example of such a popular formalism in the simulation community that breaks-down the representation of a simulation model into hierarchical components.

Our objective is to focus on the operational support of execution for such simulation models. For example, considering that the model of a single node of a Peer-to-peer network requires several (and possibly many) DEVS components, it is easy to see that running simulations of a realistic large-scale peer-to-peer network rapidly ends-up involving millions of DEVS components. In addition to the problems posed by the execution of a distributed simulation application made of millions of components, such a use-case is also challenging in terms of analytics, because when millions of components are instrumented to collect observations, it becomes a typical instance of a big-data analytics problem.

Stochastic simulation platform

Understanding how complex objects, as found in finance/insurance (option contracts), biology (proteins structure), etc. evolve is often investigated by stochastic simulations (e.g. Monte-Carlo based). These can be very computational intensive and the associated communities are always seeking adequate parallel computing infrastructures and simulation software. Being able to harness all the available computing power, while ensuring the simulation is at first performant but also robust, capable to self-adapt, e.g. to failures, is a real opportunity for research and validation of our approach. Many other simulation applications could also benefit from our models and techniques, and we may in the future set up specific collaborations, e.g. in biocomputing, data-center activity management, or other engineering domains. We have recently solved pricing of high-performance demanding financial products on heterogeneous GPUs and multicore CPUs clusters, mixing use of active objects and OpenCL codes. This kind of application could continue to serve as a benchmark for our multi-level programming model.