Abstract de la publi numéro 12915

Energy consumption is rising year after year in com- puting centers. And this trend is even accelerating as scientific computing becomes the key in a large number of scientific discovery. Scientific computing community has long been only fo- cused on raw performance, but due to the energy crisis looming, energy consideration are starting to be put on light. In this article, we present how to use phase detection of scientific algorithms, together with system tuning (changing processor frequency, network link rate) and virtual machines migrations. Using this method, we can reduce energy consumption of large scale comput- ing infrastructure with low impact on performance.