Energy Efficiency of MapReduce
Yanpei Chen, Tracy Wang, Joseph M. Hellerstein and Randy H. Katz
The energy efficiency of computer systems is becoming more prominent, with growing concern for power consumption and heat dissipation. The energy efficiency of Internet datacenters is especially interesting, with a large number of computers concentrated in a small physical space. Prior work has looked at energy efficient disks, servers, switches, and protocols. In considering a fresh redesign of the Internet datacenter with respect to energy efficiency, we believe that datacenter applications are worthy of study. In this work, we examine the energy efficiency of MapReduce, an increasingly popular application for Internet datacenters. We measure the energy consumption of MapReduce under workloads that stress different parts of the system, and seek to understand the performance and scalability behavior of MapReduce with respect to energy. From these measurements, we propose energy-saving mechanisms and directions for future work in the area. Our key finding is that well-configured system parameters and well-designed workloads can improve the energy efficiency of MapReduce by more than 100% without significant modifications to the underlying MapReduce infrastructure. Our work represents a first step in evaluating the performance of Internet datacenter applications with respect to energy.