The Case for Event-Driven Energy Accounting

Abstract:
Energy management requires a precise knowledge of the patterns of energy use. Not only knowing where the energy has been spent is important but also knowing who was responsible for the use of energy. The resolution of power measurement equipment like current meters or smart battery interface is neither sufficient to identify the hardware unit consuming the energy nor to identify the originator (energy principal) of a specific hardware activation. Our measurements have also demonstrated that timing data, the source of information used in current operating systems, is not adequate to estimate the use energy consumption, because the power depends on the patterns for the use of specific functional units. To investigate energy usage patterns we strongly encourage the use of embedded hardware monitors (e.g., processor performance counters) that have also proven to offer valuable information in the field of performance analysis. We use information about active hardware units (e.g., instruction decoder, memory management unit, cache-/memory-interface) gathered by event counters to establish a precise and energy principal-specific energy accounting. By showing the correlation of events and energy values we can provide the necessary information for energy-aware scheduling policies that go far beyond the capabilities of the state of the art power management standards like ACPI. While the aspect of energy saving by driving the system at its optimal operation point is important in mobile devices, the aspect of client- and service-specific throttling is vital in data-centers where servers have to provide mission critical services even when running under an emergency power supply in times of (rolling) black outs. An implementation of the proposed event-driven energy accounting using a Linux/x86 system and extensive measurements prove the concept.