Toward an experiment engine for lightweight grids

Abstract

This paper presents a case study conducted on the Grid'5000 platform, a lightweight grid. The goal was to make a rather simple experiment, and study how difficult it was to carry out correctly. This means it had to be correct, reproducible and efficient. The paper shows that despite the precautions taken, many parameters that could have an effect on the result were at first overlooked. It also shows that benchmarking plays a key role on making an experiment correct and reproducible. The process is in the end extremely tedious, and stresses the need for new tools to help users. The contribution of this work is to present a methodology to get correct results on grid architecture, to identify relevant problems and to propose an infrastructure that answers part of the problems encountered during experiments. Additionally, pieces of this infrastructure have been built and are also presented.