TL;DR
This paper introduces a "sleep" state to transient computing systems, allowing them to temporarily pause execution without saving their state when power levels dip. This significantly reduces unnecessary checkpointing, leading to a substantial increase in active computation time, especially for low-power energy harvesting scenarios.
What Problem Does This Paper Address?
Energy harvesting (EH) is a promising technology for powering embedded systems, but the power it provides is often low and fluctuates unpredictably. To handle these power interruptions, transient computing systems typically save their state to non-volatile memory (NVM) before power loss and restore it upon recovery. This process, known as checkpointing or hibernating, is essential for maintaining system continuity. However, existing methods like Hibernus++ can be overly cautious, initiating a full hibernation (saving state) even when the power supply is only temporarily dipping and is expected to recover quickly. This frequent, unnecessary checkpointing consumes valuable time and energy, especially with low-current EH sources, leading to systems spending more time saving their state than performing useful work. The paper aims to address this inefficiency by minimizing unnecessary hibernations and maximizing the system's active computation time.
Methods
The core contribution of this paper is the introduction of a novel "sleep" state to the transient computing paradigm, enhancing the existing Hibernus++ approach.
- Sleep State: When the system detects that its supply voltage is approaching a critical threshold (specifically, dropping below a new "sleep threshold," Vs), it enters a low-power "sleep" mode instead of immediately hibernating. In this sleep state, the system suspends its main execution but does not save its state to NVM.
- Conditional Hibernation: The system remains in sleep mode as long as the harvested energy can sustain its low power consumption. If the supply voltage continues to drop and falls below the original "hibernate threshold" (Vh), then a full hibernation (state saving) is initiated as a fail-safe measure.
- Restoration: If the supply voltage recovers above a "restore threshold" (Vr) without having dropped below Vh, the system can resume from its sleep state directly into active mode without the need for a state restoration from NVM. This avoids the time and energy overhead associated with restoring from a snapshot.
- Mathematical Analysis: The paper provides a mathematical analysis of the proposed scheme, considering voltage thresholds, power consumption in different states, and the time required for transitions. This analysis helps in determining optimal threshold values and understanding the trade-offs.
- Simulation and Experimental Validation: The proposed approach was simulated using MATLAB Simulink and experimentally validated on an MSP430FR5739 development board.
Datasets
Not discussed in the paper. The evaluation uses simulated power sources (2 mA DC current source) and real-world energy harvesting scenarios (photovoltaic cell under a desk lamp).
Key Insights and Ideas
- Sleep vs. Hibernate: The critical insight is that not every power dip requires a full state save. Introducing a "sleep" state allows the system to "wait out" minor voltage fluctuations without the significant overhead of checkpointing.
- Maximizing Active Time: By avoiding unnecessary hibernations, the system can spend a much larger proportion of its operational time performing actual computations, leading to greater progress in applications.
- Low-Current EH Synergy: The proposed method is particularly beneficial for low-current energy harvesting sources, where voltage fluctuations are common, and the energy cost of frequent hibernations is most pronounced.
- Trade-offs in Thresholds: The paper highlights that setting the sleep threshold (Vs) too close to the hibernate threshold (Vh) is crucial for effectiveness, but this involves trade-offs that require further investigation for variable harvesting conditions.
Key Results
- Simulation: In a 10-minute simulation with a 2 mA DC current source, the enhanced Hibernus++ system achieved 57.2% active run-time, a significant improvement over the original Hibernus++ which achieved 37.4%. This represents a 65% improvement in active run-time.
- Experimental Validation: The experimental results confirmed the simulation findings. When powered by a photovoltaic cell, the system oscillated between active and sleep modes, completely avoiding hibernation and thus maximizing active computation time.
Limitations and Future Work
- Variable Harvesting Conditions: The paper acknowledges that the proposed enhancement's benefits are most pronounced when the harvester output is consistently above the system's sleep current draw. For highly variable or pulsed sources, the system might still need to hibernate frequently. The optimal setting of Vs for variable harvesting conditions is identified as a subject for future work.
- Trade-offs: Further evaluation is needed to fully understand the trade-offs between setting Vs very close to Vh and the potential for frequent hibernation if the harvesting source drops below Vh.
- Application Agnosticism: While the approach is generally applicable, the specific benefits might vary depending on the application's power consumption profile and its tolerance for interruptions.
Summary updated