Section: New Results

Experimental Methodologies

Overall improvement of SimGrid

2012 was the last year of the USS-SimGrid project granted by the ANR. We thus capitalized the results of the first project by properly releasing them in the public releases. Parallel simulation is now stable enough to be used in practice by our users. In addition, the framework is now able to simulate millions of processes without any particular settings in C. The java bindings were also improved to simulate several hundred thousand processes out of the box [25] .

This year was also the first year of the SONGS project, also funded by the ANR. This project is much larger that the previous one, both in funding and targets. In surface, SONGS aims at increasing the scope of the SimGrid simulation framework by enabling the Cloud and HPC scenarios in addition to the existing Grid and P2P ones. Under the hood, it aims at providing new models specifically designed for these use cases, and also provide the necessary internal hooks so that users can modify the used models by themselves.

This project is well started, with three plenary meetings and a user conference organized over the year, but no new publication resulted of this work yet. The first work toward increasing the simulation versatility, initiated last year, was published this year[14]

Dynamic verification of liveness properties in SimGrid

A full featured model-checker is integrated to SimGrid since a few years, but it was limited to the verification of safety properties. We worked toward the verification of liveness properties in this framework. The key challenge is to quantify the state equality at state level, adding and leveraging introspection abilities to arbitrary C programs.

This constitutes the core of the PhD thesis of M. Guthmuller, started last year. A working prototype was developed during this year, described in an initial publication [20] .

Grid'5000 and related projects

We continued to play a key role in the Grid'5000 testbed in 2012. Lucas Nussbaum, being delegated by the executive committee to follow the work of the technical team, was heavily involved in the recent evolutions of the testbed (network weathermaps, storage management, etc.) and in other activities such as the preparation of the Grid'5000 winter school. We were also involved in a publication [33] which is a follow-up to the workshop on Supporting Experimental Computer Science held during SC'11, and in another publication [32] describing the recent advances on the Grid'5000 testbed in order to support experiments involving virtualization at large scale.

More specifically, our involvement in the OpenCloudWare project led us to design several tools that ease the deployment of Cloud stacks on Grid'5000 for experimental purposes. Those tools were also used during an internship that was co-advised with the Harmonic Pharma start-up, exploring how complex bio-informatics workflows could be ported to the Cloud.

On the institutional side, we will also play a central role in the Groupement d'Intérêt Scientifique that is currently being set up, since Lucas Nussbaum is a member of both the bureau and of the comité d'architectes.

Distem – DISTributed systems EMulator

In the context of ADT Solfége, we continued our work on Distem. Three releases were made over the year, with several improvements and bug fixes, including support for variable CPU and network emulation parameters during an experiment. See http://distem.gforge.inria.fr/ for more information, or our paper accepted at PDP'2013 [26] .

Kadeploy3 – scalable cluster deployment solution

Thanks to the support of ADT Kadeploy3, many efforts were carried out on Kadeploy3. Two releases were made, including many new features (many improvements to the handling of parallel commands and to the inner automaton for more fault-tolerant deployments; use of Kexec for faster deployments) as well as bug fixes.

Kadeploy3 was featured during several events (journée 2RCE, SuperComputing 2012), and in two publications: one unsuccessfully submitted to LISA'2012 [35] , one accepted in USENIX ;login: [13] .

Finally, Kadeploy3 was also the basis of submissions to the SCALE challenge held with CCGrid'2012, of which we were finalists, and of the winner challenge entry at Grid'5000 winter school 2012.

Business workflows for the description and control of experiments

We are exploring the use of Business Process Modelling and Management for the description and the control of complex experiments. In [28] , we outlined the required features for an experiment control framework, and described how business workflows could be used to address this issue. In [27] and [15] , we described our early implementation of XPFlow, a experiment control engine relying on business workflows paradigms.

Towards Open Science for distributed systems research

One of our long term goal on experimental methodologies would be the advance of an Open Science in the research domain of Distributed Systems. Scientific tools would be sufficiently assessed and easily combined when necessary, and scientific experiments would be perfectly reproducible. These objectives are still very ambitious for the researches targeting distributed systems.

In order to precisely evaluate the path remaining toward these goals, and try addressing some of the challenges that they pose, we currently host Maximiliano Geier as an Inria intern. While most researchers try to answer brilliant scientific questions with simple scientific methodologies, he is asked to answer a simple question (on the adaptation of the BitTorrent protocol to high bandwidth networks) using an advanced scientific methodology. We are also surveying the experimental methodology used in top tier conferences to gain further insight on this topic.

In addition, we are organizing Realis, an event aiming at testing the experimental reproducibility of papers submitted to Compas'2013. Associated to the Compas'13 conference, this workshop aims at providing a place to discuss the reproducibility of the experiments underlying the publications submitted to the main conference. We hope that this kind of venue will motivate the researchers to further detail their experimental methodology, ultimately allowing others to reproduce their experiments.