Methods and tools for challenging experiments on Grid’5000 : a use case on electromagnetic hybrid simulation

In the field of Distributed Systems and High Performance Computing experimental validation is heavily used against an analytic approach which is not feasible any more due to the complexity of those systems in terms of software and hardware.Therefore, researchers have to face many challenges when conducting their experiments, making the process costly and time consuming. Although world scale platforms exist and virtualization technologies enable to multiplex hardware, experiments are most of the time limited in size given the difficulty to perform them at large scale.The level of technical skills required for setting up an appropriate experimental environment is risen with the always increasing complexity of software stacks and hardware nowadays. This in turn provokes that researchers in the pressure to publish and present their results use ad hoc methodologies.Hence, experiments are difficult to track and preserve, preventing future reproduction. A variety of tools have been proposed to address this complexity at experimenting. They were motivated by the need to provide and encourage a sounder experimental process, however, those tools primary addressed much simpler scenarios such as single machine or client/server. In the context of Distributed Systems and High Performance Computing, the objective of this thesis is to make complex experiments, easier to perform, to control, to repeat and to archive. In this thesis we propose two tools for conducting experiments that demand a complex software stack and large scale. The first tool is Expo that enable to efficiently control the dynamic part of an experiment which means all the experiment workflow, monitoring of tasks, and collection of results.Expo features a description language that makes the set up of an experiment withdistributed systems less painful. Comparison against other approaches, scalability tests anduse cases are shown in this thesis which demonstrate the advantage of our approach.The second tool is called Kamelon which addresses the static part of an experiment,meaning the software stack and its configuration.Kameleon is a software appliance builderthat enables to describe and control all the process ofconstruction of a software stack for experimentation.The main contribution of Kameleon is to make easy the setup of complex software stacks andguarantee its post reconstruction.