Energy Efficiency in Hpc Systems

1.1 INTRODUCTION Power consumption of High Performance Computing (HPC) platforms is becoming a major concern for a number of reasons including cost, reliability, energy conservation, and environmental impact. High-end HPC systems today consume several megawatts of power, enough to power small towns, and are in fact, soon approaching the limits of the power available to them. For example, the Cray XT 5 Jaugar supercomputer at Oak Ridge National Laboratory (ORNL) with 182,000 processing cores consumes about 7 MW. The cost of power for this and similar HPC systems runs into millions per year. To further add to the concerns, due to power and cooling requirements and associated costs, empirical data show that every 10 degree Celcius increase in temperature results in a doubling of the system failure rate, which reduces the reliability of these expensive system. As supercomputers, large-scale data centers are meant to be clusters composed by hundreds of thousands or even millions processing cores [11] with similar power consumption concerns. Existing and ongoing research in power efficiency and power management has addressed the problem at different levels, including, for example, data center design, resource allocation, workload layer strategies, cooling techniques, etc. At the platform level (individual node or server), current power management research broadly falls into the following categories-processor and other subsystems (e.g. memory, disk, etc.) level, Operating System (OS) level and application level. Although the processor is the most power consuming component, other subsystems have incorporated energy management functionalities such as memory, storage and network interfaces (NIC). Within the OS, there are fewer power management techniques available, and include OS control of processor C-states, P-states and device power states or sleep states. At the application level several approaches have been also proposed such as those based on exploiting communication bottlenecks in MPI programs. In this chapter, we study the potential of proactive application-centric aggressive power management of data center's resources for HPC workloads. Specifically, we consider power management mechanisms and controls (currently or soon to be) available at different levels and for different subsystems, and leverage several innovative approaches that have been taken to tackle this problem in the last few years, that can be effectively used in a cross-layer application-aware manner for HPC workloads. To do this, we firstly profile standard HPC benchmarks with respect to behaviors, resource usage and power dissipation. Specifically, we profile the HPC benchmarks in terms of processor, memory, storage subsystem and …