State of the Art of Power Saving in Clusters and Results from the EDF Case Study

Abstract

Energy management for mobile devices has been traditionally a well studied topic during the last two decades, as these devices usually do not have a permanent connection to the power grid and thus solely rely on the limited battery charge. However, this trend has been mostly disregarded in the context of HPC systems as the main focus mainly relied on improving the performance at any cost. Therefore, energy costs for operating and cooling the equipment of current data centers have increased signicantly up to a point where they are able to surpass the hardware acquisition costs. In this work we survey the current energy conservation eorts in distributed systems. Hence, we start our work with the introduction of the basic principles of power and energy management. the two often misleading terms This involves the distinguishing between power and energy management. Furthermore, we present several approaches on how to measure the power consumption at system and component level. Moreover, in order to get a deeper understanding on where the most of the power is spent within a server, we take a closer look at its components (i.e. CPU, RAM, etc.) and outline the currently available energy-saving mechanisms. Given that energy consumption strongly depends on the workload characteristics we also discuss the dierent types of workloads (i.e. mobile, commercial and scientic). Afterwards, we continue with the current approaches for energy savings in distributed systems by focusing on both and cluster -level node eorts. Finally, we nish this survey with an outline of some open challenges and the introduction of the EcoGrappe project. Thereby, we detail its objectives and present some of the work done at our project partner EDF.