A survey on program-state retention for transiently-powered systems

Abstract

Low-power small-scale embedded sensing systems employing batteries generally impose high maintenance costs. To enable maintenance-free operation, they are powered from energy harvested from the environment thus making them batteryless. However, due to high variance of ambient energy, these batteryless embedded devices are unable to harvest enough energy from the environment required for continuous device operation thus hampering application progress and causing frequent loss of volatile program-state. Therefore, these batteryless devices have to employ state retention mechanisms to save the volatile program-state to non-volatile storage before interruption. These batteryless embedded sensing devices are known as transiently-powered systems (TPS). In this article, we survey existing literature to identify strategies and techniques used by each existing literature to decide what amount of volatile program-state needs to be saved and when to save it. We list the challenges in retaining program-state across periods of energy unavailability and how existing state-of-the-art solutions tackle them. We also describe different memory models and discuss factors governing the choice of each model for TPS deployment.

Introduction

In recent years, the vision of “smart dust” [1] has driven the development of tiny, embedded sensing devices that run autonomously for decades. Equipped with programmable microcontroller units (MCUs), these sensing devices are envisioned to be used in large-scale applications such as wearables [2], implants [2], small satellites [3], and wireless robotic materials [4], [5]. The monitoring of existing infrastructure [6] and the environment using insects carrying the tiny devices has also been recently proposed [7], [8].

To enable these applications, the tiny sensing devices require a continuous and perpetual energy source. Batteries are commonly used to power such embedded devices [9], [10]. However, recent advances in nanotechnology and microelectronics have made it possible for these devices to shrink in size [11]; small enough to fit on the head of a honey bee [12]. Battery designs have not scaled in the same way, thus making it hard for them to fit in such devices despite recent developments [13].

Furthermore, the use of batteries nullifies the initial vision of “smart dust” of being invisible and maintenance-free. Modern-day batteries have only limited number of power cycles, i.e., batteries have one power cycle and rechargeable batteries have a few thousand. As a result, batteries from a high number of devices have to be frequently replaced, preventing a perpetual operation of the devices and resulting in high maintenance costs.

One solution to enable the maintenance-free operation of these devices is to liberate them from batteries by powering them from harvested energy [14], [15], [16], [17], e.g., from light, vibrations, and thermal sources. Miniaturized mechanical systems have enabled the design of tiny energy harvesters at the scale of nanometers that fit into the form factor of these tiny, embedded sensing devices [17], thus making them batteryless. However, the smaller an energy harvesting unit is, the smaller is the harvested energy. Furthermore, the energy that can be harvested from the environment is generally erratic and exhibits high spatial and temporal variations [17].

To cope with the small amounts of harvested energy and to smooth the fluctuations of the feeble incoming energy, devices powered by harvested energy employ an energy buffer. This is typically a capacitor offering unlimited power cycles as opposed to batteries. The form factor of this buffer must be small enough to not only fit in the device footprint but to also allow fast recharge, thus limiting its storage capacity (typically <100$\mu$F). As a result, one power cycle of this buffer cannot supply enough energy needed to complete most of the programs running on MCU.

Intermittent Program Execution. Intuitively, the notion of energy is transformed from “limited but continuous” in traditional battery-powered devices to “unrestricted but intermittent” in these maintenance-free devices. We call these tiny, energy harvesting-based, batteryless, embedded devices transiently powered systems (TPSs) and refer to the computing paradigm as intermittent computing A TPS is active when the energy buffer is full and it is inactive, i.e., it turns off, when the energy buffer is empty and is recharging. Program execution can only occur during active periods and the device loses its volatile state when the energy falls below the minimum energy that is required by the TPS to operate.

The TPS state comprises three sub-states, i.e., program-, peripheral- and timer-state, which have to be saved onto non-volatile memory (NVM) shortly before the device becomes inactive. Forward progress of the application running on the TPS can only be ensured by retaining all three sub-states at the same time. When the TPS switches back to an active period, it restores the saved system state and continues execution. However, retaining each sub-state has its challenges that are orthogonal to each other.

• The program execution can be interrupted at any time during the operation of a TPS. At the time of interruption, size of volatile program-state is different for different applications and depends on the program-point at which energy failure occurs. Therefore, it requires sophisticated techniques to determine volatile program-state that needs to be retained across periods of energy unavailability.

• Peripheral devices require reconfiguration after each reboot. Simply capturing the current state of these peripheral devices can result in incorrect program execution, thus giving rise to new challenges.

• TPSs collect time-sensitive data and require time stamping of the collected data. However, transient energy supply can cause stale values to be processed by the application, thus producing incorrect results.

We discuss these challenges in detail in Section 2.1. Due to their unique set of requirements, existing literature devise different strategies to address each of the above-mentioned challenges [18], [19], [20], [21], [22], [23], [24], [25], [26].

Scope of the Survey. To keep the discussion focused, we survey existing literature to find solutions that address challenges associated with program-state retention, thus enabling energy-efficient TPS operations. This research space has seen rapid and significant progress over the past few years [18], [19], [20], [21], [22], [27], [28], [28], [29], [30], [31], [32], [33], [34], [35], [36], [37]. However, a classification and a common taxonomy in the realm of program-state retention solutions is still missing.

To this end, this paper encapsulates existing state-of-the-art solutions that answer the following research question: How to ensure the successful execution of a program on a TPS under intermittent energy supply? We build a taxonomy of solutions for program-state retention based on how they ensure forward progress of the programs running on TPSs. More precisely, this paper distinguishes between two key aspects of program-state retention. First, we discuss literature proposing strategies for reducing the energy required to retain the program-state before the TPS switches to an inactive period. Second, we focus on techniques that, based on compile-time analysis, identify program points where saving the program-state would require minimal energy.

Our Contribution. There exist well-organized surveys on energy harvesting and wireless energy transfer techniques within the sensor network domain [17], [38], [39]. There are also detailed surveys available for large-scale distributed systems regarding various checkpointing techniques [40], [41]. In contrast, we present a survey on TPSs, highlighting different checkpointing techniques and the underlying memory technologies employed by these tiny, batteryless, embedded sensing devices that have entirely different constraints than the devices used in the distributed and parallel computing domain.

A recent article surveyed several intermittent computing approaches [42] to briefly define challenges and future research directions in the TPS domain. In contrast, this paper gives a detailed survey of all existing state-of-the-art intermittent computing approaches and draws on an up-to-date taxonomy of program-state retention for TPSs. This paper covers all existing techniques, which answer the above-mentioned research question while discussing the trade-offs of different memory models on efficient program-state retention.

Organization. The rest of the paper is structured as follows. Section 2 gives the bigger picture of TPS operation while defining basic terminology and challenges faced by TPSs in retaining overall system state. Section 3 explains the criteria of classification for existing solutions for retaining program-state for TPSs. Sections 5 Trigger mechanism, 6 Checkpointing strategy classify the strategies employed by different intermittent computing solutions to ensure energy efficiency while performing program-state retention. Section 4 discusses different memory models and their trade-offs before we conclude the paper in Section 7.