Runtime distribution-aware dynamic voltage scaling

Yoo, Sungjoo; Jin, Hyunjong; Choi, Kyu-Myung; Kong, Jeong-Taek; Eo, Soo-Kwan

doi:10.1145/1233501.1233622

“…In [7], the authors formulate the problem mathematically with two assumptions. The first assumption is P ~ f 3 .…”

Section: Figure 1 Intra-task Dvs With Runtime Distributionmentioning

confidence: 99%

“…maximum frequency) in the middle of task execution is dynamically determined, mostly, by X/T, where X is the estimated remaining workload and T is the time to deadline. Thus, the accuracy of workload estimation determines the quality of intra-task DVS method.Several methods of workload estimation have been proposed: worst case execution time [3][4], average case execution path [5], average energy execution path [6], and statistical methods [7]. Among them, the statistical method and average energy execution path-based one are reported to give the best reduction in average switching energy consumption since they provide global minimum solutions based on mathematical formulations.…”

mentioning

confidence: 99%

“…the battery lifetime of mobile device is mostly evaluated in an average sense after running an extensive set of benchmarks and use cases), it is required to tackle DVS problems statistically to reduce average energy consumption while meeting the given deadline constraints. In [7], a statistical method based on the runtime distribution of software execution cycle (not 978-3-9810801-3-1/DATE08 …”

mentioning

confidence: 99%

See 1 more Smart Citation

Dynamic voltage scaling of supply and body bias exploiting software runtime distribution

Yoo

¹

,

Bin

²

,

Choi

³

et al. 2008

Proceedings of the Conference on Design, Automation and Test in Europe

Self Cite

View full text Add to dashboard Cite

IntroductionDynamic voltage scaling (DVS) is one of the most effective methods in reducing both switching and leakage power consumption. There have been two classes of DVS methods: intertask and intra-task DVS. Inter-task DVS methods [1][2] determines the performance level at a task granularity while intra-task DVS methods at finer granularitiesIn intra-task DVS, workload estimation plays a central role since the performance level (normalized w.r.t. maximum frequency) in the middle of task execution is dynamically determined, mostly, by X/T, where X is the estimated remaining workload and T is the time to deadline. Thus, the accuracy of workload estimation determines the quality of intra-task DVS method.Several methods of workload estimation have been proposed: worst case execution time [3][4], average case execution path [5], average energy execution path [6], and statistical methods [7]. Among them, the statistical method and average energy execution path-based one are reported to give the best reduction in average switching energy consumption since they provide global minimum solutions based on mathematical formulations. However, the leakage power consumption is not minimized by the methods since they minimize only the switching energy based on the assumption of P ~ f 3 (P ~ CV 2 f ~ f 3 since V ~ f ). Leakage power consumption has already become a real design issue. Especially, excessive leakage power consumption at high temperatures often causes significant product parametric yield drop in reality 1 . Thus, DVS methods need to optimize leakage energy as well as switching energy.In order to reduce leakage energy consumption, we apply combined V dd /V bs scaling [10][11] since body biasing (scaling V bs ) 1 Although the power consumption specification can be met at room temperature, it cannot often be met due to significant leakage power consumption at high temperatures in the product specifications, e.g. 80 or 125℃.is the most effective way to control leakage power consumption. In our work, we extend the statistical DVS method (which originally targets only dynamic energy) to tackle the reduction of both switching and leakage energy by scaling both V dd and V bs . Note that the statistical method covers the method based on average energy execution path [6] as a simplified case. We give a mathematical formulation of the problem of V dd /V bs scaling based on the statistical information, i.e. the distribution of software runtime. The formulation gives a multi-variable non-linear function of total energy consumption. As a practical solution to obtain the workload estimations for the minimum average energy consumption, we present a numerical solution.This paper is organized as follows. Section 2 reviews existing DVS methods. Section 3 explains the mathematical formulation of statistical DVS based on combined V dd /V bs scaling. Section 4 gives a total power function for combined V dd /V bs scaling. Section 5 presents a numerical solution to the problem. Section 6 reports experimental results and Section...

show abstract

“…In the case of n program regions, n workload estimations { X 0 , X 1 , …, X n } for the minimum average energy consumption need to be obtained simultaneously. Thus, the original bottom-up traversal method in [7] cannot be applied in this case.…”

Section: Figure 1 Intra-task Dvs With Runtime Distributionmentioning

confidence: 99%

“…Several methods of workload estimation have been proposed: worst case execution time [3] [4], average case execution path [5], average energy execution path [6], and statistical methods [7]. Among them, the statistical method and average energy execution path-based one are reported to give the best reduction in average switching energy consumption since they provide global minimum solutions based on mathematical formulations.…”

Section: Introductionmentioning

confidence: 99%

Dynamic Voltage Scaling of Supply and Body Bias Exploiting Software Runtime Distribution

Yoo

¹

,

Bin

²

,

Choi

³

et al. 2008

2008 Design, Automation and Test in Europe

Self Cite

15

0

View full text Add to dashboard Cite

IntroductionDynamic voltage scaling (DVS) is one of the most effective methods in reducing both switching and leakage power consumption. There have been two classes of DVS methods: intertask and intra-task DVS. Inter-task DVS methods [1][2] determines the performance level at a task granularity while intra-task DVS methods at finer granularitiesIn intra-task DVS, workload estimation plays a central role since the performance level (normalized w.r.t. maximum frequency) in the middle of task execution is dynamically determined, mostly, by X/T, where X is the estimated remaining workload and T is the time to deadline. Thus, the accuracy of workload estimation determines the quality of intra-task DVS method.Several methods of workload estimation have been proposed: worst case execution time [3][4], average case execution path [5], average energy execution path [6], and statistical methods [7]. Among them, the statistical method and average energy execution path-based one are reported to give the best reduction in average switching energy consumption since they provide global minimum solutions based on mathematical formulations. However, the leakage power consumption is not minimized by the methods since they minimize only the switching energy based on the assumption of P ~ f 3 (P ~ CV 2 f ~ f 3 since V ~ f ). Leakage power consumption has already become a real design issue. Especially, excessive leakage power consumption at high temperatures often causes significant product parametric yield drop in reality 1 . Thus, DVS methods need to optimize leakage energy as well as switching energy.In order to reduce leakage energy consumption, we apply combined V dd /V bs scaling [10][11] since body biasing (scaling V bs ) 1 Although the power consumption specification can be met at room temperature, it cannot often be met due to significant leakage power consumption at high temperatures in the product specifications, e.g. 80 or 125℃.is the most effective way to control leakage power consumption. In our work, we extend the statistical DVS method (which originally targets only dynamic energy) to tackle the reduction of both switching and leakage energy by scaling both V dd and V bs . Note that the statistical method covers the method based on average energy execution path [6] as a simplified case. We give a mathematical formulation of the problem of V dd /V bs scaling based on the statistical information, i.e. the distribution of software runtime. The formulation gives a multi-variable non-linear function of total energy consumption. As a practical solution to obtain the workload estimations for the minimum average energy consumption, we present a numerical solution.This paper is organized as follows. Section 2 reviews existing DVS methods. Section 3 explains the mathematical formulation of statistical DVS based on combined V dd /V bs scaling. Section 4 gives a total power function for combined V dd /V bs scaling. Section 5 presents a numerical solution to the problem. Section 6 reports experimental results and Section...

show abstract