Age of Information and Energy Consumption in

IoT: an Experimental Evaluation

Federico Cristofani Valerio Luconi Alessio Vecchio
University of Pisa, Italy IIT-CNR, Italy University of Pisa, Italy
f.cristofani@studenti.unipi.it valerio.luconi@iit.cnr.it alessio.vecchio@unipi.it

Abstract—The Age of Information (AoI) is an end-to-end of AoI on energy but were mainly aimed at studying AoI-
metric frequently used to understand how “fresh” the infor- aware systems in some operating scenarios [12], [13].
mation about a remote system is. In this paper, we present
arXiv:2405.05849v1 [cs.NI] 9 May 2024

In this work, we aim to evaluate the relationship between

an experimental study of the relationship between AoI and the
energy spent by the device that produces information, e.g. an AoI and energy consumption in an experimental publish-
IoT device or a monitoring sensor. Such a relationship has subscribe scenario. We built a testbed made of an IoT device
been almost neglected so far, but it is particularly important that publishes updates at a given rate, and we measure AoI at
whenever the sensing side is battery-operated. The study is a subscriber. Simultaneously, we measure energy consumption
carried out in a scenario where access is achieved via the cellular via a hardware power monitor on the IoT device. The devices
network and information is transferred using MQTT, a popular
messaging protocol in the IoT domain. Numerous parameters of are connected via a cellular network and communicate through
operation are considered, and the most efficient solutions in all the MQTT protocol on top of QUIC and TLS. We run an
configurations are provided. extensive set of experiments to explore how a wide range
Index Terms—Age of Information, energy, Internet of Things, of parameters of operations can influence the two considered
QUIC, TCP, MQTT metrics. In detail, we considered the rate of generation of
messages on the IoT device, the latency between the IoT
I. I NTRODUCTION device and the MQTT broker, the messages’ payload size, the
Age of Information (AoI) has emerged as a critical metric computational capacity of the IoT device, and the impact of
reflecting the timeliness and relevance of data delivery. In the underlying transport protocol. Thanks to the visualization
a monitoring system, AoI is defined as the time elapsed instrument of Pareto fronts, we point out the most efficient
since the generation of the latest received packet at the solutions for every configuration of parameters. We show that
monitor, and it can capture the freshness of the received there is no clear indication of a universally optimal solution,
data better than classic metrics such as delay [1]. In today’s but the context and the system’s requirements should indicate
interconnected world, where real-time data processing and which of the most efficient solutions to adopt. However, we
transmission are paramount, understanding and optimizing believe that our experimental results are extremely precious
the AoI is pivotal for a spectrum of applications, including in helping researchers and practitioners in future protocol
multimedia streaming, Internet of Things (IoT) deployments, designs, IoT system architectures, and environmental sustain-
and financial transactions. Since its introduction [2], AoI has ability initiatives.
been used as a metric to analyze and characterize a wide The rest of the paper is structured as follows. Section II
variety of systems, especially IoT-related such as intelligent provides an overview of the existing literature on the subject.
transportation systems, vehicular networks, smart agriculture, Section III describes our experimental setup. In Section IV,
and augmented reality [3]–[5]. we describe the metrics and the aggregation techniques that
Alongside the quest for timely information delivery, the we use to analyze experimental results. Section V shows our
energy footprint of network protocols and infrastructure has findings. Finally, Section VI concludes the paper.
gained significant attention. With the exponential growth of
digital traffic, data centers, and communication networks, II. R ELATED W ORK
the environmental impact of energy-intensive operations has The performance of transport and application protocols in
become a focal point for researchers, policymakers, and in- an IoT context has been extensively studied from various
dustry stakeholders alike. Especially in the IoT context, energy points of view. Several works have evaluated the adoption
consumption considerations gain paramount importance, since of the recently introduced QUIC transport protocol in IoT.
most of the devices are battery-operated [6]–[9]. In [14] the authors demonstrated the feasibility of a QUIC
Recently, various studies have investigated the trade-off standard implementation designed for tiny devices, by quanti-
between maintaining a desired freshness level of information fying storage, computing, memory, and energy requirements,
while minimizing energy consumption in IoT systems, mainly however, without involving any application protocol. Other
from a theoretical perspective [10], [11]. Conversely, recent studies have used HTTP as an application protocol over QUIC
experimental works still have not focused on the implication in an IoT environment. In [15], the authors studied the latency
and scalability of HTTP/3 over QUIC in an IoT system in Server

the cloud/edge continuum, showing results in the order of

hundreds or thousands of milliseconds depending on the load Internet Broker
of the system. The energetic performance of HTTP protocols
have been studied for both IoT devices and smartphones in [8],
[9], where the authors compared different versions of the
University
HTTP, showing that the ones using QUIC as an underlying network
Power
protocol consumed more energy. The authors also showed Monitor
uart
that their results could be due to the current lack of maturity
and standardized adoption of QUIC implementations, among
other factors. Fatima et al. compared QUIC and TCP/TLS as Publisher Subscriber
underlying protocols of MQTT, for what concerns latency, ethernet

in a simulated environment. They showed that, for short- RPI Laptop

lived connections, the performance of the two protocols is
comparable, while for long-lived connections, QUIC obtains Fig. 1: Experimental setup hardware and software architecture.
shorter completion times. The same comparison is conducted
in [16] for a ]IoT-cloud environment based on MQTT. Again,
QUIC obtained a much better performance. In [17], the environments. In [26], the AoI of a TCP/IP connection over
authors analyzed MQTT over QUIC and TCP/TLS from the WiFi, Ethernet, and cellular networks has been investigated
energetic point of view, showing that for small-medium delays via emulation and in a physical network. Observations have
QUIC was able to consume less energy, while for large ones been conducted with different sampling rates and degree of
TCP/TLS obtained a slightly better performance. network load. The AoI of TCP and UDP connections over the
Work on AoI focused on theoretically analyzing the perfor- Internet and of lightweight IoT connections in a local WiFi
mance of different systems characterized by different queuing network has been evaluated in [12]. In this work, the authors
models and scheduling algorithms in different application provide insight into how energy can be a factor impacting
scenarios [18]–[21]. AoI in constrained environments such as IoT. In [13], the
The relationship between energy and AoI has been in- authors proposed WiFresh: an AoI-aware wireless architecture
vestigated mainly from a theoretical point of view. In [10], to achieve optimal AoI even in the case of an overloaded
the authors devise two scheduling algorithms for multiple network. They show that compared to classic WiFi, WiFresh
sensors that minimize the AoI at the monitoring station. can obtain lower values of AoI.
Subsequently, energy consumption is brought into play and the The analysis of the previous works reveals considerable
authors show variants to their algorithms that can achieve close attention to topics related to the IoT, highlighting extensive
to optimal performance with significant energy consumption coverage of these issues. In particular, the emerging protocol as
reduction. A framing and scheduling policy for minimizing QUIC is gaining popularity as a modern alternative to the de-
energy consumption in cognitive wireless sensor networks facto standard TCP that showed limitations for lossy networks
has been proposed in [11]. The optimization is conducted and constrained devices. The focus is also placed on the AoI
under strict constraints of AoI. Via numerical and simulation metric, able to represent the freshness of data, that in modern
analysis, the authors provide the optimal number of samples systems assumes a pivotal role. However, significant gaps
per packet under given operational conditions, such as the emerged. Among all, there is a lack of experimental works,
number of nodes in the network and channel quality indicators. especially regarding combined studies on AoI and energy.
In [22], the authors derive closed forms for AoI and energy This work is placed at the intersection of studies exploring
consumption in an IoT monitoring system based on low power energy consumption and AoI in IoT, adopting an experimental
wide area (LPWA) wireless communication technologies and methodology. To the best of our knowledge, ours is the first
adopting a truncated automatic repeat request (TARQ) scheme. experimental work that analyzes AoI and energy consumption
The authors show that under transmit power constraint, the in a real-world IoT scenario.
TARQ scheme obtains lower AoI than a classic ARQ scheme.
III. E XPERIMENTAL S ETUP
The problem of minimizing energy consumption under AoI
constraints has been studied also in [23], where the authors This study aims to evaluate Age of Information (AoI)
derived a lower bound on the competitive ratio of any given and the energy consumption in a IoT scenario, using the
causal policy for choosing which packets to transmit on a given MQTT protocol running on top of multiple transport protocols:
node. In addition, they propose a greedy policy that achieves QUIC and TCP/TLS. We built the experimental setup depicted
that lower bound. Other works have tackled the problem of in Figure 1, composed of multiple hardware and software
AoI and energy from different perspectives, for example for components. The RPI acts as an IoT device, hosting the client
an energy harvesting source [24], or computation offloading side, i.e. the publisher in the MQTT terminology. The RPI is
in industrial IoT [25]. equipped with a cellular network card [27] able to support both
Few works focused on evaluating AoI in experimental Long Term Evolution (LTE 4G) and New Generation (NG 5G)
 TABLE I: Experimental setup hardware specs.

Device CPU RAM NIC

RPI BCM2837B0 ARMv8 64bit1 1GB LPDDR2 SIM8200EA-M2 5G HAT
Laptop Intel Core i7-8565U 16GB DDR4 Intel(R) Wireless-AC 9260
Server Intel Xeon Gold 51202 4GB Ethernet adapter
Ethernet interfaces for RPI and the laptop are not reported in the table

TABLE II: Network connections in the experimental setup. acts as the time-reference server for the RPI. The ethernet
connection is characterized by an extremely low latency, being
Connection RTT Download Upload Hops
the two devices in the same lab room, and is used only for
RPI - Server 80 ms 28 Mbit/s 15 Mbit/s 12 synchronization traffic. All the other traffic exchanged by the
Laptop - Server 5 ms 200 Mbit/s 160 Mbit/s 3 two devices flows via the cellular connection and the Internet.
The overall configuration corresponds to a rather common
scenario in the IoT domain, in which the RPI acts as a device
communication technologies3 . The RPI represents a battery- installed somewhere that produces data, typically collecting
powered device. To measure its power consumption, it is information by monitoring the environment. The subscriber
powered by the Otii Arc Pro from Qoitech power monitor [28]. on the laptop in turn can be considered as a cloud application
The Otii power monitor is featured with a Asynchronous that deals with the collection of the data produced, applying
Receiver-Transmitter (UART) interface that is used to send logic that may be sensitive to the freshness of the information
commands to the RPI device and to annotate the recorded obtained from the data. The two parties exchange information
trace with timestamps to precisely confine the portion of the through the broker, hosted in the cloud as the subscriber, which
total energy required during the execution of an experiment. represents a centralized entity enabling indirect communica-
The RPI publisher is then connected via cellular connection to tion according to the MQTT protocol. Table I summarizes the
a MQTT broker hosted on a server in the University of Pisa hardware specs of the experimental setup, while Table II pro-
cloud network. vides a characterization of the network connections between
The final component of the experimental setup is a laptop, the hardware components.
which has a twofold function. Firstly, it acts as a MQTT
subscriber client, connected to the broker via the University A. Implementation
of Pisa network (with WiFi access). The connection between The client has been implemented relying on the Quinn
the broker and the subscriber is characterized by a relatively library [29], a rust-based implementation of the IETF QUIC
low latency, being the two components in the same network. protocol. We did not use a standard MQTT client as available
Secondly, the laptop acts as the controller of the experiments. QUIC-based implementations are rather limited in number.
It is connected to both the RPI and the Otii to provide Quinn provides an async API and it is based on the Tokio
the configuration parameters and the automation for each async runtime [30]. Quinn also uses the Rustls [31] library
experiment. for the cryptographic functionality. In particular, our client
The goal of the study is to evaluate the system from the also uses the mqttbytes [32] libraries for producing MQTT
point of view of the AoI and the energy needed on the messages according to the specification. The client just imple-
publisher side. We are not interested in the energy consump- ments the subset of the MQTT protocol needed for our exper-
tion of the machine hosting the subscriber as we suppose iments, like sending messages to the broker and receiving the
that the alerting/controlling system mentioned before is not corresponding acknowledgments. To achieve higher message
executed on a battery-operated device. Similarly, the energy rates, the client uses a task to send messages to the broker and
needed to run the broker is not of interest, as it is generally another task to keep track of the corresponding incoming acks.
executed on reasonably powerful machines without constraints The client also includes the buffering mechanisms previously
in terms of energy. With this setup, we can measure both the mentioned: a task produces new information at a nominal rate
energy consumption on the RPI and the AoI, as the difference and uses a buffer to communicate with the sender task. The
between the instant when the message was generated and the latter is responsible for extracting information from the buffer
instant when it was received. To obtain a precise measure and sending it through the network. The buffer is managed
of the AoI, the two devices (the Raspberry Pi 3B+ and the according to two strategies: (i) a FIFO queue with a capacity
laptop) need to be adequately synchronized. To achieve the of 1, 16, and 1024 messages, and (ii) a drop-head-on-full
desired synchronization, we used an approach based on the queue of capacity 1. In the first case, the two tasks can be
Network Time Protocol (NTP) protocol. The two devices are blocked when the queue is full or empty and no messages
connected via a second ethernet connection, and the laptop are lost. In the second case, a message can be lost if a new
one is produced before the current one is transmitted. In this
3 Experiments have been carried out only using 4G access. case, the producing task never blocks and always operates
according to the nominal rate. The strategies will be hereafter
indicated as FIFO 1, FIFO 16, FIFO 1024, and DROP. The
broker is an instance of the Rumqttd server. The subscriber
Time-below
is rather simple and receives messages when relayed by the b
median
5/10 = 0.50

broker. The subscriber computes the AoI, by using a timestamp

collected when the message is received and another timestamp
included in the message by the publisher and indicating the
d f
time when the message was produced. The subscriber has been
implemented using the Paho library [33]. Median = 3

IV. C OLLECTED M ETRICS AND AGGREGATION c

To make the paper self-contained, we here provide a de-

a e
scription of the metrics and the aggregation tools that we
2 1 2
consider in this study.

A. Age of Information 10

The recording of generation and reception instants of the

Fig. 2: Median definition for AoI function
message published during the time window gives all the infor-
mation needed to compute the AoI not only in correspondence
with the message arrival but in any instant along the entire
duration of the experiment, obtaining an exact form for the Median Value. Besides the mean value, we compute the
sawtooth shape function that characterizes the metric [34]. median value, which is known to be more robust to noise.
Dealing with an analytic expression of the function, rather The problem is to find an extension of the median applicable
than a discrete group of points, e.g. by sampling the time to a function and not to a discrete set of points, obtaining the
evolution of AoI, allows for precise calculations of the statis- same type of relationship that exists between the arithmetic
tical quantities, needed to aggregate the collected data for the mean and integral mean. The median is defined as the value
analysis. The drawback is that those calculations might be not separating the higher half from the lower half of the points
straightforward to implement. The statistical quantities taken in the set. The concept of half the points of the set can be
into consideration are mean and median values. The formula extended to half the length of the interval over which the AoI
applied to compute the exact value of the AoI in each instant function is defined. Based on this new definition, the median
is the following: value can be assumed as the level of AoI for which the sum of
∆(t) = t − U (t) (1) the lengths of the intervals on the x-axis that have associated
AoI less than (greater than) the median is exactly half of the
where ∆(t) is the AoI at instant t, and U (t) the generation total length of the interval over which the function is defined.
time of the newest data received. This formula represents the The visualization of this definition is reported in Figure 2.
time evolution of the AoI metric, which assumes the shape To compute the median of a sawtooth function, we devised a
of an irregular sawtooth, rising linearly while waiting for new method based on projecting the function segments on an axis.
messages and suddenly falling to a lower value at the reception This method obtains the exact median value without involving
of a new message. The function is always positive, with a complex computations. The exploited property is the unitary
lower bound given by the one-way delay of the network. Each slope, which allows us to work on either the x-axis or the y-
linear section has a 45◦ inclination, i.e. slope 1, so the distance axis. The method is based on the projection of each segment
covered on the x-axis is the same as the one on the y-axis for on a vertical axis, referred to as the projection axis. The
each segment. The function is integrable. projections on the projection axis can generate overlapping
Mean Value. The mean value is the simplest form of data regions. For each region, a weight is assigned, computed as
aggregation. However, dealing with a function requires a the number of different projections that fall in it. Next, we
different approach than the simple arithmetic mean applicable compute a weighted sum as the length of each region times
for a discrete set of points. To compute the mean of the the associated weight. This sum will be referred to as the
AoI function we used the integral mean, defined as the ratio total extension. Finally, we iterate over the regions, bottom-
between the definite integral of the function over the given up, computing their extension, i.e. their weighted length, and
interval and the length of the interval: summing them up until half of the total extension is reached.
f : [a, b] → R bounded and integrable in [a, b] The last sum will probably exceed half of the total extension,
Z b
(2) so only a fraction of the weighted length of the last region is
1
M (f, [a, b]) = f (x) dx to be considered. The point on which the iteration stops, i.e.
b−a a the quote on the axis, corresponds to the median value. The
In our case, the interval length is given by the time elapsed construction of the projection axis is reported in Figure 3 and
between the first and last message received by the subscriber. a possible implementation is shown in Algorithm 1.
 Algorithm 1 Projection median
function COMPUTE PROJECTION MEDIAN(timestamps) ▷
Returns a float
projections ← empty list
extension ← 0
x2
for i ← 1 to len(timestamps) − 1 do
start aoi ← timestamps[i−1][”rx”]−timestamps[i−
x3
1][”gen”]
end aoi ← timestamps[i][”rx”] − timestamps[i −
x2 1][”gen”]
projections.append((start aoi, 1)) ▷ +1 marks the
x1
begin of the projection
projections.append((end aoi, −1)) ▷ -1 marks the
end of the projection
extension ← extension + (end aoi − start aoi)
end for
Fig. 3: Projection algorithm. projections.sort(key ← x : x[0]) ▷ Sort list of tuples
using the second element as key
weight ← 0
B. Energy curr ext ← 0
start ← -1
The Otii power monitor returns the total energy absorbed for projection in projections do
during the execution of the experiment. Given that the duration if start ̸= −1 then
of an experiment is not fixed, as we will see in Section V, we region ← projection[0] − start ▷ Region length
normalized the measured energy over the experiment execution if curr ext + region × weight ≥ extension/2 then
time. The result is the average power consumed by the device return start + (extension/2 − curr ext)
in the experiment. end if
curr ext ← curr ext + region × weight
C. Pareto Efficiency
end if
The two metrics, AoI and energy, are considered together weight ← weight + projection[1] ▷ Weight is
for the energy-constrained nature of IoT devices. This can increased/decreased by 1
introduce additional complexity in the analysis: the objective start ← projection[0] ▷ Start of new region on
is to optimize both, but there is a trade-off to be resolved. projection axis
The best method for presenting the results is to use a tool end for
that shows all possible optimal solutions and leaves it up end function
to the reader to determine which of those presented is the
most suitable configuration for a given system. The tool that
perfectly matches the above requirements is the Pareto front, considered system. The points having both coordinates worse
built up by a set of Pareto efficiencies. Pareto efficiency is than other points are referred to as Pareto-dominated and the
a concept introduced by the Italian economist and engineer associated configuration doesn’t provide an optimal choice for
Vilfredo Pareto that states that, given a set of resources any system, thus they can be discarded. An example of the
an allocation is efficient if it is not possible to improve Pareto front is reported in Figure 4.
the condition of one individual without worsening those of
another, i.e. Pareto improvements are not available. This V. E XPERIMENTAL R ESULTS
concept can also be applied in multi-objective optimization Our experiments aim to evaluate the combination of the
when there is no feasible solution that minimizes all objective two metrics of interest when varying different operation pa-
functions simultaneously [35]. Let us consider a set of points rameters. Specifically, we evaluated the impact on energy and
in the plane whose coordinates represent AoI and energy, AoI when varying: (i) the publisher generation rate, (ii) the
respectively, each associated with a particular configuration. additional delay between the publisher and the broker, i.e.
The Pareto efficient solutions are those for which one metric the RPI and the server, (iii) the payload size of the publisher
cannot be improved without worsening the other. The sub- messages, (iv) the number of cores used on the RPI, (v) the
set of Pareto efficient solutions represents the Pareto front on transport protocol. The number of generated messages in one
which a designer can choose the solution to adopt, considering run is equal to the nominal generation rate multiplied by 60,
possible constraints of the system in terms of AoI and energy. which corresponds to a duration of 60 seconds. However, as
The choice can be considered as an a-posteriori assignment we will highlight in the next sections, the real generation rate
of weights to the metrics, based on the knowledge of the can be lower than the nominal rate. In these cases, the duration
 to around the nominal delay value when the rate is 2000 msg/s.
Pareto Front Then, it starts increasing again for all FIFO configurations. For
4.0
Dominated the DROP configurations, instead, it remains at the minimum
Optimal
3.5 Front value. This is expected, as the property of the DROP policy is
to deliver always the most “fresh” message. Figure 5b, shows
3.0
this trend for rates from 100 to 4000. It must be noted that
the QUIC configurations obtain always lower AoI values than
Power [W]

2.5
their corresponding TLS ones. Figure 5c shows the power
consumption over the nominal rate. After this initial sudden
increase, the power continues rising with the rate but with a
2.0
slower slope. This trend is valid for all configurations up to the
rate of 2000 msg/s. From such a rate onward the DROP policy
1.5
stops increasing and the same applies for FIFO 1, but only in
the QUIC configuration. The FIFO-1 TLS encounters an upper
1.0
40 60 80 100 bound around the rate of 5000 msg/s. Other configurations,
AoI [ms]
even if subjected to some fluctuations, present a quite evident
Fig. 4: Example of Pareto Front. increasing trend up to the maximum rate. This suggests that
the power consumption is dominated by the transmission rate.
Once the maximum rate for a configuration is reached, further
of an experiment will be longer. For each configuration, we increases in the nominal rate won’t produce any significant
executed one run, except for the last one, where we executed variation in the average power. It must be noted that below
30 runs. Besides the metrics used to compute energy and AoI, 2000 msg/s QUIC always shows a higher power consump-
the RPI collects also context metrics useful to explain the tion. Over that threshold, things get more confused, even if
obtained results, such as the cellular network quality and the generally QUIC consumes more than its corresponding TLS
clock synchronization precision. configuration.
Figures 5d and 5e show the Pareto fronts for the FIFO poli-
A. Impact of the Generation Rate cies and the DROP policies, respectively. The results for the
For these experiments, we set the message payload to 128 B, rate of 1 msg/s are omitted for the sake of clarity, as they ob-
no additional delay on the cellular link, and a single core tain a median AoI of approximately 500 ms, even if the power
on the RPI. We varied the nominal generation rate on the consumption is extremely low. For the FIFO configurations,
publisher from 1 msg/s to 8000 msg/s, with the following the rates that appear in the front do not exceed 2000 msg/s.
steps: 1, 10, 100, 1000, 2000, ..., 8000. We show results for As previously observed, excessively high transmission rates
both the transport protocols (QUIC and TLS), and all the do not guarantee advantages for either power consumption or
buffering strategies. Figure 5a shows the relationship between AoI. Besides that, in general, the lower the rate the lower
the real generation rate and the nominal generation rate, for the consumption and the higher the AoI, and vice-versa, and
both transport protocols and the different buffering strategies. the QUIC configurations obtain the lower AoI. It is worth
The figure shows that the buffer size poses an upper limit on noticing that the high density of points in the left extreme also
the real generation rate. The bigger the buffer, the higher the shows that, once approaching the lower AoI bound, any further
real rate that can be obtained. For FIFO policies, this happens improvement requires a significant increment in the power
because, once the buffer is full, the publisher blocks. This consumption. In the DROP configurations, we can notice that
automatically reduces the real publishing rate so that it is also rates higher than 2000 msg/s are present in the front. This
compatible with the maximum message throughput that can happens because higher rates, even if higher than the nominal
be sent out via the network. The presence of a larger buffer one, produce “fresher” information, thus helping keep the AoI
allows the publisher to cope with the transient reduction of low, at the cost of possible information loss.
throughput due to changing network conditions and, more
importantly, it allows the aggregation of multiple messages B. Impact of Additional Delay
within a single transport-layer packet. For the DROP buffering For these experiments, we artificially applied additional
strategy, the real rate corresponds to the number of messages delay on the cellular link in both directions (uplink and
that are transmitted by the RPI. It has to be noted that for the downlink). Specifically, the delay was added on the server
FIFO 1 policy, the rate obtained by TLS is much higher than machine via the Linux Traffic Control (TC) command. We
the one obtained by QUIC. This happens because in its base performed runs with additional delays from 20 to 80 ms, at
configuration, TCP with TLS uses the Nagle algorithm, which steps of 20 ms. For what concerns AoI, as could be expected,
performs a much more aggressive aggregation than QUIC. an upward shift can be appreciated in Figure 6, which shows
For each rate and each protocol/transport configuration, we the raw results for the QUIC, FIFO 16 configuration with a
computed median AoI and average power. The median AoI rate of 1000 msg/s. The shift is less evident for lower rates,
starts from high values when the rate is low, and then decreases as the AoI values are already high. The maximum achievable
 Median AoI over rate Average Power over rate
Real and Nominal rate 110
8000 Quic, DROP Tls, DROP Quic, DROP Tls, DROP
Quic, DROP Tls, DROP Quic, FIFO-1 Tls, FIFO-1 Quic, FIFO-1 Tls, FIFO-1
Quic, FIFO-1 Tls, FIFO-1 Quic, FIFO-16 Tls, FIFO-16 Quic, FIFO-16 Tls, FIFO-16
Quic, FIFO-16 Tls, FIFO-16 3.00
7000 100 Quic, FIFO-1024 Tls, FIFO-1024 Quic, FIFO-1024 Tls, FIFO-1024
Quic, FIFO-1024 Tls, FIFO-1024

2.75
6000 90

2.50

Average Power [W]

5000

Median AoI [ms]

80
Real rate [msg/s]

4000 2.25
70

3000 2.00
60
2000
1.75
50
1000
1.50
40
1
1 1000 2000 3000 4000 5000 6000 7000 8000 100 1000 2000 3000 4000 0 1000 2000 3000 4000 5000 6000 7000 8000
Nominal rate [msg/s] Nominal rate [msg/s] Nominal rate [msg/s]

(a) Real rate vs. nominal rate. (b) Median AoI over rate (c) Average power over rate

Pareto Front (FIFO) Pareto Front (Drop)

2K Quic, FIFO-1 8K Quic, DROP
2.4 Tls, FIFO-1 5K Tls, DROP
2.3
Quic, FIFO-16 7K Pareto front
1K2K Quic, FIFO-1024 2K
Tls, FIFO-1024 2.2
1K1K
2.2 Pareto front
Average Power [W]

Average Power [W]

2.1 1K
100
2.0
2.0 1K
1.9
100
1.8 1.8

1.7
10 10 10
1.6 1.6
50 60 70 80 90 40 50 60 70 80
Median AoI [ms]) Median AoI [ms])
(d) Pareto front for FIFO configurations (e) Pareto front for DROP configurations

Fig. 5: Generation rate results.

AoI time evolution for Quic, Rate 1000

Delay 0 ms Delay 20 ms Delay 40 ms Delay 60 ms Delay 80 ms • Aggregation. The aggregation at the transport layer for
Median: 45.15 ms Median: 65.20 ms Median: 79.80 ms Median: 99.22 ms Median: 132.04 ms
500
TLS/TCP is based on the Nagle algorithm, that, for high
400 transmission rates, allows for sending only fully sized
segments, independently from the network delay, while
AoI [ms]

300

200 QUIC adopts a blander aggregation mechanism.

100 • Kernel optimization. The Linux kernel on the RPI
0 20 40 60 0 20 40 60 0 20
Time [s]
40 60 0 20 40 60 0 20 40 60 doesn’t offer optimization for the UDP protocol, such as
Generic Segmentation Offload (GSO), potentially reduc-
Fig. 6: AoI values for different additional delays. ing the performance of QUIC.

We conclude the analysis of the impact of the additional delay

rate is affected too. In general, a long queue allows higher by showing the Pareto fronts for FIFO and DROP buffering
rates, as already observed. However, for longer queues (namely strategies (Figures 7b and 7c, respectively). In the figures,
FIFO 16 and FIFO 1024), by increasing the delay there is an for each additional delay (including no additional delay) a
overall reduction of the maximum rate, mainly for QUIC. The separate front is depicted. The leftmost front is the one with
phenomenon can be visualized in Figure 7a, where, starting the minimum delay and the rightmost front is the one with the
from a 40 ms additional delay, the rate achieved by QUIC is highest delay. For the FIFO configurations, the considerations
progressively lower than the one obtained by TLS. The causes for the case without additional delay apply also to the other
for this behavior are beyond the scope of this work, however, cases, i.e. lower rates guarantee a low power consumption,
they could lie at the intersection of the following aspects: at the cost of higher AoI, and vice-versa. In addition, QUIC
• Congestion control. The congestion control algorithm is configurations obtain lower AoI, while TLS ones obtain lower
the same, Cubic, for both transport protocols. However, energy consumption, as already observed in the case with
the operational parameters may differ, resulting in differ- no additional delay. Rates higher than 2000 msg/s do not
ent performance when varying the network delay. guarantee any benefits except in two cases. In both cases,
• Flow control. The advanced flow control mechanism of these results are obtained with a FIFO 1 queue and the
QUIC based on multiple streams is not exploited by the QUIC protocol. Considering that in the FIFO configurations,
implemented tool, so in this configuration also QUIC may the generation rate adapts to the transmission rate, the result
suffer from the head-of-line blocking problem. is to be considered equal to one obtained by a lower rate.
 Pareto Front (FIFO) Pareto Front (Drop)
2K Quic, FIFO-1 8K5K Quic, DROP
1K Tls, FIFO-1 Tls, DROP
Maximum rate Quic, FIFO-16 2.6 1K
2.6 1K 4K
8000 quic tls 2K Tls, FIFO-16
6K* Quic, FIFO-1024
2K 5K* Tls, FIFO-1024
7000 2.4 2.4 3K 1K
1K1K 8K 100
2K 100 5K 6K

Average Power [W]

Average Power [W]

1K
2K 2K 5K
6000
Maximum rate [msg/s]

7K
2K 1K 3K
4K 1K
1K
1K 1K1K 2.2 1K
2.2 1K 1K 5K3K
5000 1K
1K
100
1K 1K 100 1K 100
4000 2.0 1K 1K
100 1K 10 2.0 100 1K
100
100 100
3000 100
1.8 1.8
2000
10 10 10 1010 10 10 10
1000 1.6 101010 10 10 1.6 10 10
0 40 60 80 100 120 140 160 180 40 60 80 100 120 140 160
0 ms 20 ms 40 ms 60 ms 80 ms Median AoI [ms]) Median AoI [ms])

(a) Maximum rate for different delays. (b) Pareto fronts for FIFO delayed configurations (c) Pareto fronts for DROP delayed configurations

Fig. 7: Additional delay results.

TABLE III: Median AoI varying the payload size for TLS, power consumption for the configurations with higher payload
FIFO-16, rate 100 configuration sizes, while the AoI decreases as the payload increases, for
the rate of 100 msg/s and in some configurations even 1000
Payload size Median AoI
msg/s. Table III shows this behavior for the TLS, FIFO 16, rate
128B 107.38 100 msg/s configuration. Particularly interesting are the Pareto
256B 85.04 fronts for the payload size of 2048 B (Figure 8), which show
512B 65.52 the configurations with a rate of 100 mgs/s as the best ones
1KB 60.24 for AoI. In fact, at higher rates, the AoI values are extremely
2KB 55.46 high. This happens because the system is not able to handle
the required throughput, and this does not depend on the queue
size, as testified by Figure 9, which shows the anomalous
The improvement in terms of AoI with respect to the second behavior for both the FIFO 1024 and the DROP queues.
highest point in the front is negligible, at the unjustifiable cost Therefore, the cause seems to be due to oversized buffering
of higher energy consumption. The same considerations found that takes place somewhere between the transmitter and the
for the case with no additional delay apply also to the DROP receiver, either in the RPI (on-device bufferbloat [36]) or in the
configurations. In particular, these configurations benefit also cellular network infrastructure. It has to be noticed a difference
from higher rates, as already pointed out. between the AoI obtained with QUIC and TLS, with QUIC
that always obtains a better performance. For what concerns
C. Impact of Different Payload Sizes power consumption, the least power-demanding transmission
In previous experiments, we considered a fixed payload size protocol is TLS, however, when the payload size increases the
of 128 B. However, there could be scenarios characterized by difference between the protocols tends to shrink, especially for
advanced sensors producing more complex, thus bigger data the FIFO configurations.
structures. In addition, multiple sensors could be attached to
the same device which would in turn produce bigger packets D. Impact of Multiple Cores
containing more than one message. In these experiments, In this experiment, we evaluate the impact on power con-
we evaluate the impact of greater payload sizes on energy sumption and AoI of enabling an additional core on the
consumption and AoI. We consider the following payload RPI. The software we developed is single-threaded, and con-
sizes: 128 B, 256 B, 512 B, 1024 B, 2048 B. The maximum currency is implemented via asynchronous mechanisms, that
payload size of 2048 B allows us to consider application exploit the idle times dictated by I/O operations to carry out
scenarios where a large amount of data has to be fitted in other tasks. However, in parallel with our software, multiple
a single message, as well as to account for possible effects kernel threads are executed, responsible for handling network
on the AoI of messages that do not fit in a single segment at operations and other system tasks. Thus, enabling more than
the transport layer. For these experiments, we considered only one core could benefit the overall system’s performance.
rates up to 1000 msg/s as (i) the best results for AoI in the Enabling the second core allowed the system to reach
previous experiments were obtained with rates around 1000- higher rates, and, more importantly, a lower AoI. This is
2000 msg/s, (ii) the combination of large payload size and high particularly evident for QUIC in the FIFO 1 configurations, as
transmission rate would produce a very high throughput witch can be observed in Figure 10a. This effect is not found in the
could not be sustainable by the RPI and IoT devices in general. DROP configurations, as the possibility to overwrite the buffer
The analysis of the Pareto fronts highlights a slightly higher content allows always to deliver the “freshest” information.
 Pareto Front (FIFO) Pareto Front (Drop)
100 2.2 100
Quic, FIFO-1 Quic, DROP
100 Tls, FIFO-1 Tls, DROP
2.1 Quic, FIFO-16 Pareto front
Tls, FIFO-1024 2.1
100 Pareto front
Average Power [W]

Average Power [W]

2.0
2.0
100
1.9 1.9

1.8 1.8

1.7 10
1.7 10 10 10
50 60 70 80 90 50 60 70 80 90
Median AoI [ms]) Median AoI [ms])
(a) FIFO. (b) DROP.

Fig. 8: Pareto fronts for 2048 B payload configurations.

Quic Tls Quic Tls

Rate 1000, Quic, FIFO-1024 Median: 1930.64 ms Rate 1000, Tls, FIFO-1024 Median: 2603.49 ms Rate 1000, Quic, FIFO-0 Median: 568.78 ms Rate 1000, Tls, FIFO-0 Median: 1372.76 ms
4000 4000
2000 2000
3500 3500

3000 3000
1500 1500
2500 2500
AoI [ms]

AoI [ms]

AoI [ms]

AoI [ms]
2000 2000
1000 1000
1500 1500

1000 1000 500 500

500 500

0 0 0 0
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time [s] Time [s] Time [s] Time [s]

(a) FIFO 1024. (b) DROP.

Fig. 9: Time evolution for 2048 B payload, rate 1000 msg/s.

Median AoI over transmission rate Drop-Head FIFO-1

Core-1 Core-1 Core-1
100 Core-2 2.5 Core-2 2.5 Core-2

90
2.0 2.0
Median AoI [ms]

80
Power [W]

Power [W]

1.5 1.5
70

60 1.0 1.0

50
0.5 0.5
40
0 500 1000 1500 2000 2500 0.0 0.0
Transmission rate [msg/s] 1 10 100 1K 2K 3K 4K 5K 6K 7K 8K 1 10 100 1K 2K 3K 4K 5K 6K 7K 8K

(a) AoI comparison enabling second core for QUIC, (b) Energy comparison enabling second core for QUIC, FIFO-1 and DROP.
FIFO-1.

Fig. 10: Results for multiple cores configurations.

The positive impact of the second core has a drawback, in Figure 10b. For DROP configurations, it is not possible
paid in terms of additional energy required to power the to spot a clear difference in power consumption. For FIFO
hardware component, but not in all configurations, as shown configurations instead, the impact on the power consumption
 Median AoI over transmission rate Power transmission rate
Quic 2.4 Quic
Tls Tls
140 Tls (nagle disbled) Tls (nagle disbled)

120 2.2
Median AoI [ms]

Power [W]
100
2.0

80
1.8

60

1.6
40
0 250 500 750 1000 1250 1500 1750 2000 0 250 500 750 1000 1250 1500 1750 2000
Transmission rate [msg/s] Transmission rate [msg/s]

(a) AoI over transmission rate. (b) Power over transmission rate.

Fig. 11: Impact of the transport protocol for FIFO 16 configurations.

is significant as the rate gets higher. This is because the second Pareto Front (FIFO)
core allows higher transmission rates, which, in turn, can Quic
2.4 2K Tls
increase the power consumption of the network card. Tls-nagle-off
Pareto front
E. Impact of the Transport Protocol 2K
2.2 1K
1K

Average Power [W]

We finally executed a set of experiments aimed at quantify-
500
ing the impact of the transport protocol on power consumption 1K
500
2.0 200
and AoI. We considered three transport protocols: QUIC, 200 100
TLS, and TLS with Nagle’s algorithm disabled. We added 500 50
the latter because, as aforementioned, the aggregation plays 100
1.8
a crucial role in the maximum achievable rate and the AoI, 50 200
20
often resulting in a discriminant factor between QUIC and 20
1010
TLS. We considered just two buffering policies: FIFO 16 and 1.6
DROP. We used just a single core and a fixed payload size 40 50 60 70 80 90
Median AoI [ms])
of 128 B. We limited the maximum generation rate to 2000
msg/s but we explored more values in the range of generation Fig. 12: Pareto fronts for FIFO configurations.
rates, in particular: 5, 10, 20, 50, 100, 200, 500, 1000, 2000.
No additional delays were considered. Finally, we kept the
duration of a single experiment to 60 s, but we repeated broker), Nagle’s algorithm delays transmission until acks are
each run 30 times. This is to ensure statistical robustness. We received or the segment gets full, and this negatively affects
aggregated the results (power consumption and AoI) of each AoI despite the increase of the transmission rate.
repetition of the same setup by taking the median value. The The AoI improvement obtained by removing the aggrega-
results for the two buffering strategies are extremely similar, tion has a consequence from the energy point of view, as
thus, to avoid confusion, we show just the ones for FIFO 16. shown in Figure 11b. The TLS curve with Nagle’s algorithm
Figure 11a shows the AoI obtained when varying the disabled is much closer to the QUIC curve than the standard
transmission rate. The better performance in terms of AoI of TLS curve. The QUIC configuration seems to be the most
QUIC with respect to standard TLS is confirmed. However, the energy-demanding, at least up to the highest considered rate,
most interesting result is that, by disabling Nagle’s algorithm, for which the curves swap. The lack of aggregation in the
TLS yields the same if not slightly better results as QUIC. This TLS with Nagle’s algorithm disabled may cause the inversion.
result shows that the main difference in terms of AoI between The QUIC protocol is responsible for lighter aggregation than
the transport protocols is given by the aggregation mechanism Nagle’s algorithm, so the effects on the AoI are limited, but
at the transport layer, much more aggressive when Nagle’s when the transmission rate increases, having some messages
algorithm is enabled. Standard TLS shows a peculiar trend shipped in the same transport layer packet allows for fewer
for rates higher than 10 msg/s. AoI suddenly and unexpect- transmissions and therefore less energy spent. A simple anal-
edly increases for a 20 msg/s rate, to start decreasing again ysis of traces collected for a rate of 2000 msg/s shows a mild
afterward, but assessing of significantly higher values than the aggregation for QUIC, around 2 messages per segment, while
other two protocols. This is due to Nagle’s algorithm, which TLS without Nagle’s algorithm doesn’t provide any form of
is triggered by acknowledgments: when the transmission rate aggregation.
produces messages with a period higher than the network RTT Figure 12 shows the Pareto fronts for the three considered
(which we recall is ∼80 ms between the publisher and the protocols. Besides the median values, the 25th and 75th
percentiles for both AoI and power are also reported, which some scenarios. Therefore, FIFO buffering strategies should be
identify three bands around the Pareto front of each protocol. considered as well, especially when employing long queues,
A larger band means a greater variability of data and thus that allow for aggregation at the transport level resulting in
less confidence in the single point obtained via aggregation. a dramatic improvement in terms of reachable throughput.
When the bands are far enough it is quite safe to state that The aggregation at the transport layer was revealed to be a
one solution is better than the other, however, if the bands paramount factor to take into consideration. The aggressive
intersect each other, the two solutions are almost equivalent, aggregation enforced by Nagle’s algorithm is the main factor
thus the choice should be taken considering other aspects. The causing different results for the two transport protocols, TLS
Pareto front identified by the combination of the three fronts and QUIC. QUIC able to guarantee a better AoI than stan-
is finally depicted as a black dotted line. dard TLS at the cost of higher energy consumption, at least for
In the right part of the graph, the three transport protocols the rates employs in the analysis. This difference disappears
behave similarly, as highlighted by the three overlapped bands. when disabling Nagle’s algorithm, making the two protocols
Moving towards the left, TLS dominates the other protocols, comparable.
leaving a space from one band to the others and having Overall, our study confirmed the possibility of adopting
the Pareto front completely in its band. Keeping moving in QUIC as a reliable transport protocol, replacing TCP. How-
the same direction the standard TLS is no longer able to ever, the immaturity of the new protocol may represent an
guarantee the optimal results, so it is replaced by the other obstacle to its adoption, favoring a more standardized protocol
two protocols, for which there isn’t a clear winner. It is worth for which compatibility and support are much larger.
noticing that QUIC presents a larger variability, indicated by
a larger band, especially for slower rates, i.e. on the right ACKNOWLEDGMENT
part of the graphs, while both versions of TLS generate a R EFERENCES
narrower band. Moreover, the Pareto fronts show that the
aggregation of Nagle’s algorithm can save energy, at the cost [1] Q. Abbas, S. A. Hassan, H. K. Qureshi, K. Dev, and H. Jung, “A
comprehensive survey on age of information in massive IoT networks,”
of a higher AoI. The result is that for a middle region, Comput. Commun., vol. 197, pp. 199–213, 2023.
where the AoI is not minimized, the standard TLS seems to [2] S. Kaul, R. Yates, and M. Gruteser, “Real-time status: How often should
guarantee the best trade-off. When the requirements on AoI one update?” in Proc. IEEE INFOCOM ’12, 2012, pp. 2731–2735.
[3] M. Patra, A. Sengupta, and C. S. R. Murthy, “On minimizing the system
are the most stringent, the aggregation limits the performance, information age in vehicular ad-hoc networks via efficient scheduling
thus disabling Nagle’s algorithm or adopting QUIC could and piggybacking,” Wirel. Netw., vol. 22, pp. 1625–1639, 2016.
help to meet the requirements. The simple fine-tuning of the [4] N. Suma, S. R. Samson, S. Saranya, G. Shanmugapriya, and R. Sub-
hashri, “IOT based smart agriculture monitoring system,” Int. J. Recent
TLS protocol, derived from the underlying TCP option, can Innov. Trends Comput. Commun., vol. 5, no. 2, pp. 177–181, 2017.
guarantee optimal protocol results even in those scenarios [5] C. Chaccour and W. Saad, “On the Ruin of Age of Information in
where QUIC seemed to provide better results, so in many Augmented Reality over Wireless Terahertz (THz) Networks,” in Proc.
of GLOBECOM ’20, 2020, pp. 1–6.
cases the greater availability of TLS may favor this protocol [6] C. Caiazza, V. Luconi, and A. Vecchio, “Saving energy on smartphones
over the emerging QUIC, for which the standardization process through edge computing: an experimental evaluation,” in Proc. of ACM
and availability are not yet comparable to the classic TLS over SIGCOMM NET4us ’22, 2022, p. 20–25.
[7] C. Caiazza, S. Giordano, V. Luconi, and A. Vecchio, “Edge computing
TCP. vs centralized cloud: Impact of communication latency on the energy
consumption of LTE terminal nodes,” Comput. Commun., vol. 194, pp.
VI. C ONCLUSION 213–225, 2022.
[8] C. Caiazza, V. Luconi, and A. Vecchio, “Measuring the Energy of Smart-
In this work, we conducted a combined study of AoI and phone Communications in the Edge-Cloud Continuum: Approaches,
energy consumption in an IoT environment based on MQTT, Challenges, and a Case Study,” IEEE Internet Comput., vol. 27, no. 6,
where a battery-operated device acts as a data publisher. pp. 29–35, 2023.
[9] ——, “Energy consumption of smartphones and IoT devices when using
Experimental results show that improving the AoI requires different versions of the HTTP protocol,” Pervasive Mob. Comput.,
to increase the energy expenditure, as both of them depend vol. 97, p. 101871, 2024.
somehow on the transmission rate. However, our contributions [10] V. Tripathi and S. Moharir, “Age of information in multi-source sys-
tems,” in Proc. of GLOBECOM ’17, 2017, pp. 1–6.
are not limited to this result. We considered a wide range
[11] A. Valehi and A. Razi, “Maximizing Energy Efficiency of Cognitive
of system operational parameters and showed their impact on Wireless Sensor Networks With Constrained Age of Information,” IEEE
both AoI and energy. Our results provide system designers Trans. Cogn. Commun. Netw., vol. 3, no. 4, pp. 643–654, 2017.
with guidelines and numbers useful for finding the desired [12] H. B. Beytur, S. Baghaee, and E. Uysal, “Towards aoi-aware smart iot
systems,” in Proc. of ICNC ’20, 2020, pp. 353–357.
trade-off between AoI and energy consumption, according to [13] I. Kadota, M. S. Rahman, and E. Modiano, “WiFresh: Age-of-
the system requirements. Information from Theory to Implementation,” in Proc. of ICCCN ’21,
Thanks to Pareto fronts, we highlighted all the optimal 2021, pp. 1–11.
[14] L. Eggert, “Towards Securing the Internet of Things with QUIC,” 2020.
solutions, leaving to the final designer the possibility to choose [15] J. Dizdarević and A. Jukan, “Experimental Benchmarking of
which operational point better satisfies the specific system HTTP/QUIC Protocol in IoT Cloud/Edge Continuum,” in Proc. of IEEE
requirements. The results showed that a DROP buffering ICC ’21, 2021, pp. 1–6.
[16] A. Alqattaa, D. Loebenberger, and L. Moeges, “Analyzing the Latency
policy is better for the application queue management, but of QUIC over an IoT Gateway,” in Proc. of IEEE COINS ’22, 2022, pp.
paying the cost of message loss which may not be admitted in 1–6.
[17] S. Jeddou, F. Fernández, L. Diez, A. Baina, N. Abdallah, and R. Agüero, [26] C. Sönmez, S. Baghaee, A. Ergişi, and E. Uysal-Biyikoglu, “Age-of-
“Delay and Energy Consumption of MQTT over QUIC: An Empirical Information in Practice: Status Age Measured Over TCP/IP Connections
Characterization Using Commercial-Off-The-Shelf Devices,” Sensors, Through WiFi, Ethernet and LTE,” in Proc. of IEEE BlackSeaCom ’18,
vol. 22, no. 10, 2022. 2018, pp. 1–5.
[18] S. Kaul, M. Gruteser, V. Rai, and J. Kenney, “Minimizing age of [27] WaveShare, “SIM8200EA-M2 5G HAT,” https://www.waveshare.com/
information in vehicular networks,” in Proc. of IEEE SAHCN ’11, 2011, wiki/SIM8200EA-M2 5G HAT, accessed: April 4, 2024.
pp. 350–358. [28] Qoitech, “Otii Arc Pro by Qoitech,” https://www.qoitech.com/
[19] R. D. Yates and S. Kaul, “Real-time status updating: Multiple sources,” otii-arc-pro/, Qoitech AB, accessed on February 21, 2024.
in Proc. of IEEE ISIT ’12, 2012, pp. 2666–2670. [29] Dirkjan Ochtman and Benjamin Saunders and Jean-Christophe Begue,
[20] Y. Inoue, “Analysis of the Age of Information with Packet Deadline and “quinn: A Rust implementation of the QUIC transport protocol,” https:
Infinite Buffer Capacity,” in Proc. of IEEE ISIT ’12, 2018, pp. 2639– //quinn-rs.github.io/quinn, accessed on February 21, 2024.
2643. [30] Tokio Developers, “Tokio: A runtime for writing reliable asynchronous
[21] F. Chiariotti, A. A. Deshpande, M. Giordani, K. Antonakoglou, T. Mah- applications with Rust,” https://tokio.rs/.
moodi, and A. Zanella, “QUIC-EST: A QUIC-Enabled Scheduling and
[31] Rustls Developers, “rustls: A modern TLS library in Rust,” https:
Transmission Scheme to Maximize VoI with Correlated Data Flows,”
//github.com/rustls/rustls.
IEEE Commun. Mag., vol. 59, no. 4, pp. 30–36, 2021.
[22] Y. Gu, H. Chen, Y. Zhou, Y. Li, and B. Vucetic, “Timely Status Update [32] Bytebeamio and contributors, “mqttbytes: MQTT packet
in Internet of Things Monitoring Systems: An Age-Energy Tradeoff,” encoder/decoder in Rust,” https://github.com/bytebeamio/mqttbytes.
IEEE Internet Things J., vol. 6, no. 3, pp. 5324–5335, 2019. [33] Eclipse Foundation, “Eclipse Paho MQTT Client,” https://eclipse.dev/
[23] K. Saurav and R. Vaze, “Online Energy Minimization Under a Peak paho/, accessed on February 21, 2024.
Age of Information Constraint,” IEEE J. Sel. Areas Inf. Theory, vol. 4, [34] R. D. Yates, Y. Sun, D. R. Brown, S. K. Kaul, E. Modiano, and
pp. 579–590, 2023. S. Ulukus, “Age of information: An introduction and survey,” IEEE
[24] B. T. Bacinoglu, Y. Sun, E. Uysal–Bivikoglu, and V. Mutlu, “Achiev- J. Sel. Areas Commun., vol. 39, no. 5, pp. 1183–1210, 2021.
ing the Age-Energy Tradeoff with a Finite-Battery Energy Harvesting [35] P. Ngatchou, A. Zarei, and A. El-Sharkawi, “Pareto Multi Objective
Source,” in Proc. of IEEE ISIT ’18, 2018, pp. 876–880. Optimization,” in Proc. of ISAP ’05, 2005, pp. 84–91.
[25] J. Huang, H. Gao, S. Wan, and Y. Chen, “AoI-aware energy control and [36] Y. Guo, F. Qian, Q. A. Chen, Z. M. Mao, and S. Sen, “Understanding
computation offloading for industrial IoT,” Future Gener. Comput. Syst., On-device Bufferbloat for Cellular Upload,” in Proc. of ACM SIGCOMM
vol. 139, pp. 29–37, 2023. IMC ’16, 2016, p. 303–317.