Preview

PDF, English Download (3MB) | Terms of use

Abstract

In recent years, energy has become one of the most important factors for de- signing and operating large scale computing systems. This is particularly true in high-performance computing, where systems often consist of thousands of nodes. Especially after the end of Dennard’s scaling, the demand for energy- proportionality in components, where energy is depending linearly on utilization, increases continuously. As the main contributor to the overall power consumption, processors have received the main attention so far. The increasing energy proportionality of processors, however, shifts the focus to other components such as interconnection networks. Their share of the overall power consumption is expected to increase to 20% or more while other components further increase their efficiency in the near future. Hence, it is crucial to improve energy proportionality in interconnection networks likewise to reduce overall power and energy consumption. To facilitate these attempts, this work provides comprehensive studies about energy saving in interconnection networks at different levels. First, interconnection networks differ fundamentally from other components in their underlying technology. To gain a deeper understanding of these differences and to identify targets for energy savings, this work provides a detailed power analysis of current network hardware. Furthermore, various applications at different scales are analyzed regarding their communication patterns and locality properties. The findings show that communication makes up only a small fraction of the execution time and networks are actually idling most of the time. Another observation is that point-to-point communication often only occurs within various small subsets of all participants, which indicates that a coordinated mapping could further decrease network traffic. Based on these studies, three different energy-saving policies are designed, which all differ in their implementation and focus. Then, these policies are evaluated in an event-based, power-aware network simulator. While two policies that operate completely local at link level, enable significant energy savings of more than 90% in most analyses, the hybrid one does not provide further benefits despite significant additional design effort. Additionally, these studies include network design parameters, such as transition time between different link configurations, as well as the three most common topologies in supercomputing systems. The final part of this work addresses the interactions of congestion management and energy-saving policies. Although both network management strategies aim for different goals and use opposite approaches, they complement each other and can increase energy efficiency in all studies as well as improve the performance overhead as opposed to plain energy saving.