Energy characterization of IoT systems through design aspect monitoring

Abstract

The technological revolution brought by the Internet of Things (IoT) is characterized by a high level of automation based, to a large extent, on battery autonomy. Important risks hindering its wide adoption, though, are associated with device battery lifetime, which is affected by system design aspects such as connectivity, data processing and storage, as well as security protection against cyber-threats. Even though simulation can help for the energy cost estimation of IoT applications before their actual deployment, it is still challenging, and extensive effort is required to converge to a feasible architectural deployment scenario. This article introduces a method to address this challenge by estimating the energy cost of the IoT design aspects and identifying the feasible deployment scenarios, for an IoT system architecture. The method is illustrated on a smart city application that consists of subsystems for building management and intelligent transportation. These two subsystems employ a variety of IoT devices connected to an Orion border router. We estimate the feasibility of various architectural deployments with respect to the system requirements and conclude to those that are possible, as feedback to the system designers. The case study results include quantitative metrics for the evaluation of system requirements using a new aspect monitoring technique and the well-established Statistical Model Checking (SMC) approach.

Appendices

Appendices

A Energy cost for IoT aspects

This appendix refers to the computation of the energy cost for energy terms presented in Sect. 2.2. Initially, the cost for each aspect $E_{\mathrm{ASP}_{i}}$ is computed as follows.

For the connectivity aspect:

$$\begin{aligned} \begin{aligned} E_{\mathrm{ASP}_\mathrm{CONN}} = \mathop {\sum _{j = 1}^{N_{Tx}}I_{Tx}*V_{Tx}*\Delta t_{{Tx}_{j}}} + \\ \mathop {\sum _{k = 1}^{N_{Rx}}I_{Rx}*V_{Rx}*\Delta t_{{Rx}_{k}}} \end{aligned} \end{aligned}$$

(5)

For the processing aspect:

$$\begin{aligned} \begin{aligned} E_{\mathrm{ASP}_\mathrm{PROC}} = \mathop {\sum _{z = 1}^{N_\mathrm{CPU}}I_\mathrm{CPU}*V_\mathrm{CPU}*\Delta t_{\mathrm{CPU}_{z}}} \end{aligned} \end{aligned}$$

(6)

Energy consumption for the security aspect is linked to both connectivity and data processing aspects; however, the contribution percentage for each one varies and depends on the energy parameters of the IoT application. Hence:

$$\begin{aligned} \begin{aligned} E_{\mathrm{ASP}_\mathrm{SEC}} = \Delta E_{\mathrm{ASP}_\mathrm{CONN}} + \Delta E_{\mathrm{ASP}_\mathrm{PROC}} \end{aligned} \end{aligned}$$

(7)

An additional energy term that should be considered on top of these aspects is the energy consumed for data exchange or control actions that are handled by the IoT device peripherals, where

$$\begin{aligned} \begin{aligned} E_\mathrm{PER} = \mathop {\sum _{w = 1}^{N_\mathrm{PER}}I_\mathrm{PER}*V_\mathrm{PER}*\Delta t_{\mathrm{PER}_{w}}} \end{aligned} \end{aligned}$$

(8)

where $N_\mathrm{PER}$ indicates the relative number of occurrences that the IoT device has interacted with its peripherals either for data exchange or control actions.

The energy consumed in the energy saving (i.e., LPM) mode is computed as:

$$\begin{aligned} \begin{aligned} E_\mathrm{LPM} = \mathop {\sum _{h = 1}^{N_\mathrm{LPM}} I_\mathrm{LPM}*V_\mathrm{LPM}*\Delta t_{\mathrm{LPM}_{h}}} \end{aligned} \end{aligned}$$

(9)

where $N_\mathrm{LPM}$ indicates the relative number of occurrences that the IoT device switches off its radio to save energy. $N_\mathrm{LPM}$ depends on the RDC protocol that the IoT device is using.

For IoT applications that are frequently exchanging data, such as ITS applications that involve continuous broadcasting and listening for traffic awareness data, the time duration that a device remains in an operating mode cannot be easily distinguished. Hence, the calculation of energy cost for the connectivity aspects is based on an alternative form of Eq. 5, focused on number of bytes that are transmitted or received. This equation is:

$$\begin{aligned} \begin{aligned} E \, '_{\mathrm{ASP}_\mathrm{CONN}} = \dfrac{{(I_{T_x} * V_{T_x})}}{bit} * \sum _{}^{} Tx \; bits \\ + \dfrac{{(I_{R_x} * V_{R_x})}}{bit} * \sum _{}^{} Rx \; bits \end{aligned} \end{aligned}$$

(10)

where $\dfrac{{(I_{T_x} * V_{T_x})}}{bit}$ and $\dfrac{{(I_{R_x} * V_{R_x})}}{bit}$ indicate respectively the energy consumed for the transmission/reception of one bit from the device of the IoT application.

For the same type of IoT applications, $E_{\mathrm{ASP}_\mathrm{CPU}}$ is calculated by replacing the term $\Delta t_{\mathrm{CPU}_{z}}$ of Eq. 6 with the maximum CPU load time measurements that are obtained using the Linux dstat library. Moreover, $E_{\mathrm{ASP}_\mathrm{LPM}}$ is obtained by replacing the term $\Delta t_{\mathrm{LPM}_{h}}$ of Eq. 9 with idle CPU time measurements from the dstat library. Finally, the computation of $E_{\mathrm{ASP}_\mathrm{SEC}}$ remains unchanged. The updated equations for energy cost computation were used to calculate the energy cost for the experiments of the case study ITS subsystem in Sect. 4.2.

B Distribution fitting for energy model calibration

In this appendix, we describe the distribution fitting technique that is used in our method to calibrate the energy model (Fig. 3) with the energy measurements obtained from the application execution on the IoT architecture (Step 4 in Fig. 4). In our method, distribution fitting is used under the consideration that the target model is a probability distribution. Through this technique, the energy model includes the energy-oriented behavior of the real system under study in the form of variables that are characterized by a probability law. This constitutes the model faithful and allows us to use it for energy consumption estimation in place of the actual system. In the following part, we describe the employed distribution fitting technique.

Fig. 13

Fitted energy distribution for the transmission (Tx) mode

The distribution technique itself is based on the randomness of input data and thus cannot be applied to deterministic or statistically correlated data. Instead of this, the data should be independent, such that one outcome of a random sample does not affect the outcome of another. This holds for energy data as IoT devices have asynchronous and not correlated changes, which is a consequence of relying in event-driven operating systems as the Contiki OS [7].

The fitting process is using well-known methods, such as moments matching and maximum likelihood. The moments matching method estimates the model parameters by using as many moments as the number of missing parameters of the candidate distribution. These moments depend on the probability law that the chosen candidate distribution follows. On the other hand, maximum likelihood finds the parameters that maximize the likelihood function. Then, the fitted distributions are validated against the input energy data using goodness-of-fit tests, such as the Kolmogorov–Smirnov (K–S).

An example fitted distribution characterizing the energy consumed while a device is in Tx mode for the ITS subsystem of Sect. 4.1 is illustrated in Fig. 13. Horizontal axis reflects the range in which energy values can vary, whereas the vertical illustrates the Probability Density Function (PDF). In this example, the distribution that is selected as a best fit is Cauchy with $\sigma =8.8014, \mu =409.99$ moments. The large energy consumption values are due to the overall size of the CAM and DENM messages (167 bytes) that include internal containers in comparison with the CoAP temperature and light messages of the BMS subsystem with maximum length of 30 bytes. If we consider the energy samples of the Cauchy distribution as $X=[{x_1, x_2, \ldots , x_n]}$, the distribution parameters $\theta _1$ and $\theta _2$ that maximize the likelihood function are computed as follows:

$$\begin{aligned} L(x_1,\ldots ,x_n)= \prod _{i=1}^{n} \frac{1}{\pi (1+{x_i}^{2})} \end{aligned}$$

(11)

During the validation phase, the goodness-of-fit tests have given 0.23228 error for Kolmogorov–Smirnov (K–S).

The fitted distributions are calibrating the energy model in the form of probabilistic variable $t\lambda$ in Fig. 3 (marked with blue color). This variable take values based on a non-deterministic selection that is following the probability law of four distinct fitted distributions for the time duration that the IoT device remains in each operating mode. Based on the chosen time duration, the energy model afterwards uses the equations of “Appendix A” to compute the energy cost.