A Methodology for Modeling Dynamic and Static Power Consumption for Multicore Processors

Abstract

System designers and application programmers must consider trade-offs between performance and energy. Making energy-aware decisions when designing an application or runtime system requires quantitative information about power consumed by different processor components. We present a methodology to model static and dynamic power consumption of individual cores and the uncore components, and we validate our power model for both sequential and parallel benchmarks at different voltage-frequency pairs on an Intel R Haswell platform.Our power models yield the following insights about energyefficient scaling. (1) We show that uncore energy accounts for up to 74% of total energy. In particular, uncore static energy can be as high as 61% of total energy, potentially making it a major source of energy inefficiency. (2) We find that the frequency at which an application expends the lowest energy depends on how memory-bound it is.(3) We demonstrate that even though using more cores may improve performance, the energy consumed by stalled cores during serial portions of the program can make using fewer cores more energy-efficient.

How to Measure Energy Consumption in Machine Learning Algorithms

“…Category 2 presents papers [14,19,1,2,13,12,15,27,23,8], which introduce empirical models, based on performance counters and at an architecture level. [14,2] are able to model both dynamic and static power of the processor. The authors of [19] on the other hand do not differentiate between static and dynamic power in their model.…”

“…These techniques give an understanding of which hardware components are being used in a specific computation. Papers surveyed with an architecture-level modeling focus are: [22,14,19,5,36,21,1,2,13,12,15,31,23,27,32,8] Finally, to just measure energy consumption, any of the estimating methodologies can be applied, since there are no constraints on optimizing a specific part of the system.…”

“…On the other hand, direct measurements are suitable for these scenarios. Modeling approaches suitable for large-scale datasets are: [14,19,1,2,13,12,15,27,31,23,32,30,8]. Small datasets can use the same models as suggested for big datasets, plus the ones based on simulations.…”

“…Real-time measurements allows for real-time energy optimization based on current load. While all models are suitable for offline learning, since there are no specific requirements on this part, only a set of those are also suitable for online learning scenarios, thus for real-time measurements [14,19,1,2,13,12,15,27,31,23,32,30,8].…”

Comparing optimal and heuristic taskgraph scheduling on parallel machines with frequency scaling

“…The complexity of a model increases significantly with its accuracy. As the power consumption of a processor depends on several factors, such as the temperature, instruction mix, usage rate, and technology of the processor, there exist numerous approaches in the literature to model the power consumption of a processor with varying complexities and accuracies . In general, without considering a specific processor, typically a quadratic or cubic frequency function is assumed, as the frequency and voltage are loosely linearly correlated.…”