TY - JOUR

T1 - On Stochastic Optimization for Smartphone CPU Energy Consumption Decrease

AU - Pelogeiko, M.

AU - Sartasov, S.

AU - Granichin, O.

PY - 2023/9/25

Y1 - 2023/9/25

N2 - Extending smartphone working time is an ongoing endeavour becoming more and more important with each passing year. It could be achieved by more advanced hardware or by introducing energy-aware practices to software, and the latter is a more accessible approach. As the CPU is one of the most power-hungry smartphone devices, Dynamic Voltage Frequency Scaling (DVFS) is a technique to adjust CPU frequency to the current computational needs, and different algorithms were already developed, both energy-aware and energy-agnostic kinds. Following our previous work on the subject, we propose a novel DVFS approach to use simultaneous perturbation stochastic approximation (SPSA) with two noisy observations for tracking the optimal frequency and implementing several algorithms based on it. Moreover, we also address an issue of hardware lag between a signal for the CPU to change frequency and its actual update. As Android OS could use a default task scheduler or an energy-aware one, which is capable of taking advantage of heterogeneous mobile CPU architectures such as ARM big.LITTLE, we also explore an integration scheme between the proposed algorithms and OS schedulers. A model-based testing methodology to compare the developed algorithms against existing ones is presented, and a test suite reflecting real-world use case scenarios is outlined. Our experiments show that the SPSA-based algorithmworks well with EAS with a simplified integration scheme, showing CPU performance comparable to other energy-aware DVFS algorithms and a decreased energy consumption.

AB - Extending smartphone working time is an ongoing endeavour becoming more and more important with each passing year. It could be achieved by more advanced hardware or by introducing energy-aware practices to software, and the latter is a more accessible approach. As the CPU is one of the most power-hungry smartphone devices, Dynamic Voltage Frequency Scaling (DVFS) is a technique to adjust CPU frequency to the current computational needs, and different algorithms were already developed, both energy-aware and energy-agnostic kinds. Following our previous work on the subject, we propose a novel DVFS approach to use simultaneous perturbation stochastic approximation (SPSA) with two noisy observations for tracking the optimal frequency and implementing several algorithms based on it. Moreover, we also address an issue of hardware lag between a signal for the CPU to change frequency and its actual update. As Android OS could use a default task scheduler or an energy-aware one, which is capable of taking advantage of heterogeneous mobile CPU architectures such as ARM big.LITTLE, we also explore an integration scheme between the proposed algorithms and OS schedulers. A model-based testing methodology to compare the developed algorithms against existing ones is presented, and a test suite reflecting real-world use case scenarios is outlined. Our experiments show that the SPSA-based algorithmworks well with EAS with a simplified integration scheme, showing CPU performance comparable to other energy-aware DVFS algorithms and a decreased energy consumption.

KW - Android OS

KW - dynamic voltage frequency scaling

KW - stochastic optimization

KW - SPSA

KW - energy consumption

KW - Android OS

KW - SPSA

KW - dynamic voltage frequency scaling

KW - energy consumption

KW - stochastic optimization

UR - https://www.mendeley.com/catalogue/ab63818f-807f-3d91-92ca-9286d4d174c3/

U2 - 10.15622/ia.22.5.3

DO - 10.15622/ia.22.5.3

M3 - Article

VL - 22

SP - 1004

EP - 1033

JO - SPIIRAS Proceedings

JF - SPIIRAS Proceedings

SN - 2078-9181

IS - 5

ER -