BLE Parameter Optimization for IoT Applications

Andreina Liendo, Dominique Morche, Roberto Guizzetti, Franck Rousseau

To cite this version:

Andreina Liendo, Dominique Morche, Roberto Guizzetti, Franck Rousseau. BLE Parameter Opti-
mization for IoT Applications. IEEE International Conference on Communications (ICC’2018), May
2018, Kansas City, MO, United States. <hal-01775056>

HAL Id: hal-01775056

https://hal.archives-ouvertes.fr/hal-01775056
Submitted on 24 Apr 2018

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 BLE Parameter Optimization for IoT Applications
Andreina Liendo∗‡¶ , Dominique Morche‡ , Roberto Guizzetti¶ and Franck Rousseau∗
∗ Grenoble Alps University, Grenoble Institute of Technology, France
Email: [email protected]
‡ CEA-LETI, MINATEC, 38000 Grenoble, France

Email: [email protected]
¶ STMicroelectronics, 38920 Crolles, France

Email: [email protected]

Abstract—Bluetooth Low Energy (BLE) has been designed as and energy consumption. DL is the time it takes for a device
a power efficient protocol for small portable and autonomous to discover or to be discovered, in other words the time to
devices, showing its efficiency for connecting these devices with establish a communication.
smartphones to periodically and frequently exchange data, like
heart rate or notifications. Additionally, BLE is present in almost Table I: BLE Neighbor Discovery parameters
every smartphone, turning it into perfect ubiquitous remote
control for smart homes, buildings or cities. Nevertheless, there Name Notation Value according to the standard
is still room to improve BLE performance for typical IoT use
Scan Interval TSI ≤ 10.24 s
cases where battery lifetime should reach several years. In
Scan Window TSW ≤ TSI
this paper we propose an extension to a model for evaluating Advertising Interval TAI 20 (or 100) ms ≤ TAI ≤ 10.24 s
BLE performance, latency and energy consumption, in order to integer multiple of 0.625 ms
provide realistic results based on various scenario conditions. In Advertising Delay ρ pseudo-random value in [0, 10] ms
addition, we propose a parameter optimization of the BLE Neigh- Advertising Event TAE = TAI + ρ
bor Discovery process, in order to obtain the best performance Adv. Indication interval TAW ≤ 10 ms (low duty cycle)
possible depending on the constraints of specific use cases. Our
results on two typical IoT test-cases show that advertiser battery
lifetime can be increased up to ≈ 89× (9.55 days to 2.32 years) BLE is suitable for a very wide range of application
for a first case, and ≈ 281× (9.55 months to 7.36 years) for a scenarios, but to maximize the lifetime of the devices while
second case. guaranteeing the user experience expected for a given appli-
Index Terms—Internet of Things; Bluetooth Low Energy; cation, it is necessary to find a suitable set of parameters that
Neighbor Discovery; Discovery Latency; Energy Consumption; are specific to the type of communication patterns being used
Battery Lifetime; Testbed; Measurements.
and the constraints specific to the application. In this work,
we propose a method to find optimal sets of parameters to
I. I NTRODUCTION
minimize the energy consumption for a wide range of BLE
BLE has two communication modes according to its spec- applications, thus maximizing battery lifetime.
ification: Neighbor Discovery or ND and Connected Mode In previous work [2]–[6] authors have proposed models
or CM. ND is the first phase, where two devices discover for energy consumption during ND from the advertiser’s per-
each other to establish a link. CM is the second phase, where spective and considering that the scanner is already listening
devices can exchange data over their connection, which is when the advertiser starts advertising. These models are based
teared down once the session is over. While in ND, the on quantitative models of the physical layer via BLE SoC
central device is the scanner and the peripheral device is the measurements of current consumption [2], and probabilistic
advertiser. When reaching CM, the central device becomes the models of the Link Layer of the protocol [3]–[6]. For certain
master and peripheral the slave. During ND, communication use cases the role of the devices might be reversed, for example
is asynchronous and devices transit through 4 of the 5 BLE in the case where there are fixed advertisers and scanners are
states [1]: standby, scanning and initiating for passing by. In this situation, the advertiser is waiting for a
scanners, or standby and advertising for advertisers, scanner, which is not taken into account in existing models.
using only the 3 advertising channels. When in CM, a device In this paper, we present an extension and optimization of
transits through connection and standby using the 37 the model proposed by Kindt et al. [4] in Section III. Then we
data channels. The time spent in CM depends on the size propose a method for parameter optimization in Section IV.
of the data and/or the duration of the session, whereas the The objective is to obtain the best BLE performance possible
time a BLE transceiver will spend transmitting, receiving or during ND, with respect to two important metrics: Discovery
in standby mode during ND is determined by the parameters Latency (DL) and energy consumption. Based on our parame-
given in Table I. The values assigned to these parameters ter optimization, we give DL, energy consumption and battery
determine two important and interdependent characteristics of lifetime results for two IoT use cases in Section VI. These
the ND phase for BLE devices that have a huge impact on results are compared with performance obtained when using
user experience and device lifetime: Discovery Latency (DL) configurations recommended by the Bluetooth SIG Profiles
[7]. DL model and parameter optimization, are validated via suitable when the scanner behaves as the peripheral device
testbed experiments in Section V. at the application level.
We extend and optimize this model to estimate BLE perfor-
II. O PTIMIZING BLE FOR A WIDE RANGE OF SCENARIOS mance, DL and energy consumption, regardless of the role of
the devices. In addition, we propose a parameter optimization
The BLE standard supports a wide range of parameter method based on this model, in order to obtain the lowest
values for ND. The choice of these parameters directly af- BLE energy consumption possible, while taking into account
fect the performances of BLE devices. Therefore, it is of use case requirements such as maximum acceptable Discovery
capital importance to use an optimal parameter configuration Latency. Finally, we evaluate the Bluetooth SIG Profiles for
in order to ensure a fair tradeoff between DL and energy some typical IoT use cases and compare with the results we
consumption. According to the BLE specification, before find with our new model. We present the results in terms of
establishing a connection, the scanner must first listen to the battery lifetime based on the battery model presented in our
advertising channel during the ND in order to synchronize other work [8].
with the advertiser. The scanner/master is the initiator and is
responsible for establishing a connection with one or several Table II: Recommended Advertising Interval, Scan Interval
advertisers/slaves. Once in Connected Mode, or CM, master and Scan Window Values (Table 5.1 and 5.2 in [9])
and slave will periodically exchange application data. From the
Advertising/Scan Duration Parameter Value
point of view of BLE communication, the scanner/master is the
central device whereas the advertiser is the peripheral which is First 30 seconds (fast connection) Adv. Interval 20 ms to 30 ms
After 30 seconds (reduced power) Adv. Interval 1 s to 2.5 s
typically at the sensor side. However, for certain use cases, the
First 30 seconds (fast connection) Scan Interval 30 ms to 60 ms
scanner may play the peripheral role at the application level. Scan Window 30 ms
Applications based on proximity sensing are an example After 30 seconds (reduced power) Scan Interval 1.28 s
where the advertiser is configured to broadcast data to adver- – Option 1 Scan Window 11.25 ms
After 30 seconds (reduced power) Scan Interval 2.56 s
tise periodically. Multiple scanners in range can receive data – Option 2 Scan Window 11.25 ms
without the need for a connection establishment. From the
application level point of view, the advertiser and the scanner
would play the central and the peripheral role respectively, and III. M ODEL OPTIMIZATION
furthermore, the scanner is at the user side, meaning that for We based our work on the more complete and precise model
such use cases, DL and energy consumption criticality are at available in the literature [4]. The model considers that the
the scanner side. This shows that efforts to improve perfor- advertiser starts at a given phase offset called φ after t = 0,
mance should be focused on one side or another, depending moment at which the scanner starts scanning on channel 37 as
on constraints that are imposed by the use case. shown in Figure 1a. In order to determine scanner performance
The Bluetooth SIG has given a list of profiles [7] intended to when constrained DL and energy consumption is at the scanner
provide efficient configurations for BLE devices for different side, we propose to consider the case where the advertiser
applications, such as collecting sensor information, health, starts advertising at a given phase offset φ before t = 0.
sports and fitness, environmental sensing and proximity. These In other words, in this case, the scanner starts scanning at
profiles define the behavior for both central and peripheral a random time within one advertising cycle as illustrated in
devices, and they include recommended scanner/advertiser Figure 1b.
configurations to ensure optimal DL and energy consumption Algorithm 1 from [4], is used to estimate the advertiser DL.
during ND. There are 24 profiles in total, where the same It is the elapsed time between the start of the first advertising
configuration is recommended for 10 different profiles and event, until the start of the advertising event received by the
it is depicted in Table II, 4 profiles recommend different scanner. The duration of an advertising event successfully
configurations, 6 profiles state that devices should be config- received by the scanner tadvEvnt (ch) is added to obtain the
ured with consideration for user expectations of connection total advertiser DL. It depends on the channel over which
establishment, and the rest state that connection establishment reception takes place, which is estimated at line 9.
requirements are defined by a higher layer specification. The condition to consider a successful reception is that the
To the best of our knowledge, there is no performance evalu- starting point of an advertiser event n falls into a region within
ation using these profiles and there is no evaluation considering a scanner interval called effective scanning window which
the scanner as a peripheral device with constrained DL and goes from tsemin(k) to tsemax(k). The effective scanning
energy consumption. The analytical model proposed by Kindt window is different for each of the advertising channels and
et al. [4] can be used to determine the right configuration depends on four parameters: scanning interval TSI , scanning
depending on application constraints. But results from this window TSW , duration of an advertising packet transmission
kind of model are very general and might require very long Tapk and the duration of hopping to the next channel tch (see
computation time. Getting results from Algorithm 1 can take Figure 1, where TAW = Tapk + tch ). The effective scanning
from minutes to several hours when evaluating configurations window is estimated as seen in Table III, where tearly is the
that uses large parameters. Furthermore, this model is not time with respect to each advertising channel from which, if
 Scanning Cycle

Advertiser Scanner
Scanning Event

Scanning Scanning Scanning

ch.37 ch.38 ch.39
TSW TSW TSW time
TSI TSI TSI
TAW
Advertising Event(TAE)

ch.37

ch.38

ch.39

ch.37

ch.38

ch.39

ch.37

ch.38

ch.39
Φ TAI TAI time
𝜌 𝜌
t=0

(a) The Scanner is available waiting for an advertising packet

AdvertisingCycle
Advertiser

TAW
Advertising Event(TAE)
ch.37

ch.38

ch.39

ch.37

ch.38

ch.39

ch.37

ch.38

ch.39
TAI 𝜌 TAI time
Scanner

Scanning Event

Scanning Scanning Scanning

ch.37 ch.38 ch.39
TSW TSW time
Φ TSI TSI

t=0

(b) The Advertiser is available waiting to be discovered by the scanner

Figure 1

an advertising event n begins, the packet can be succesfully function (PDF) of the start of an advertising event over time
received even if tearly is not within the scanning window. t (which depends on both n and φ) is required. The shape of
tlate is the time with respect to each advertising channel the distribution depends on n. For more detail on the shape
from which, if an advertisign event n begins, the packet can’t of the distribution we refer the reader to the paper [4].
be successfully received, even if tlate is within the scanning Then the consumed charge in Coulombs during DL can be
window. For more details we refer the reader to the paper [4]. calculated knowing the current and voltage of the circuit for
The probability of an advertising event having started each event. The current and voltage vary from one manufac-
between tsemin(k) and tsemax(k) is called Phit , it is turer to the other and are extracted from quantitative models of
calculated at line 10 of Algorithm 1, where n and k are the physical layer which are based on SoC measurements [2].
the current evaluated advertiser and scanner events which are Finally, knowing the charge and DL, energy consumed during
considered according to lines 13 and 6 respectively. P cM is DL can be estimated. Quantitative models of physical layer in
the probability that the nth advertising events don’t lead to our work are based on TI CC2540 and ST BlueNRG current
a successful reception (cumulative miss probability). With in- consumption measurements extracted from technical literature
creasing values of n, the probability that one of the advertising [10], [11].
events is received successfully, grows. Thus, P cM shrinks
with growing n. The algorithm terminates if (1 − P cM ) is Table III: Effective Scanning Window
smaller than a lower bound  = 0.9999. The expected DL for Ch tearly tlate
a given φ offset is calculated at line 11. It is considered for
37 0 Tapk
the expected value of ρ to be 5ms and the error is neglected. 38 Tapk + tch 2Tapk + tch
Then the average DL for a given TSI , TSW , TAI configuration 39 2Tapk + 2tch 3Tapk + 2tch
is obtained [4] by integrating the expected DL results over all Ch tsemin(k) tsemax(k)
possible values of φ. A numerical integration is then performed 37 0 (k − 1)TSI + TSW − tlate
by multiplying the results with ∆ and computing the sum of 38 (k − 1)TSI − tearly (k − 1)TSI + TSW − tlate
these values. 39 (k − 1)TSI − tearly (k − 1)TSI + TSW − tlate

For calculating the probability Phit of whether an adver-

tising event is successfully received, the probability density In order to evaluate DL from scanner perspective when it is
Algorithm 1 BLE Average Advertiser DL Algorithm 2 Proposed BLE Average Scanner DL
1: DLadv ← 0 1: for φ = −TAI + 10ms to 0 step ∆ do
2: for φ = 0 to 3TSI step ∆ do 2: n = 1 , P cM = 1, ε = 0.9999
3: n ← 0 , DLexp ← 0, Phit ← 0, P cM ← 1, ch ← 37 3: while 1 − P cM ≤ ε do
4: while 1 − P cM ≤ ε do 4: tideal = φ + (n − 1)TAI , treal = tideal + (n − 1) ρ 2
5: tideal ← φ + nTAI , Phit ← 0 5: if tideal < 0 then
t tideal +n5ms
6: Kmin = b ideal
TSI c , Kmax = b TSI c 6: P cM = 1 , n = n + 1
7: for K = Kmin to K = Kmax do 7: continue
8: ch ← mod(j, 3) 8: end if
t tideal +n5ms
9: (tsemin(k), tsemax(k)) ← getInterval(ch)) 9: Kmin = b idealTSI c , Kmax = d TSI e , Phit = 0
10: Phitp ← Phit + 10: for K = Kmin to K = Kmax do
pk(treal , n, ρ n 11: tsemin(k, TSI ,TSW , T apk, tch) ,
12 , tideal , tsemin(k), tsemax(k), ρ)
11: DLexp ← DLexp +pk.P cM (n(TAI +5ms)+tadvEvent (ch)) tsemax(k, TSI ,TSW , Tp apk, tch)
n
12: end for 12: pk (treal , n, ρ 12 , tideal , tsemin(k), tsemax(k), ρ)
13: P cM ← P cM (1 − Phit ) , n ← n + 1 13: Phit = Phit + pk
14: end while 14: end for
15: DLadv ← DLadv + DLexp 15: P cM = P cM (1 − Phit ) , n = n + 1
16: end for
DL
16: end while
17: DLadv ← 3Tadv 17: end for
SI
∆ 18: advtreal = treal (max(pk )) + abs(φ) , scantreal = treal (max(pk ))
19: advDL = mean(advtreal ) , scanDL = mean(scantreal )

peripheral at the application level, we propose an optimization

as depicted in Algorithm 2. Unlike in Algorithm 1, the phase IV. BLE PARAMETER OPTIMIZATION
offset φ is within a negative range, that is the case where the BLE parameter optimization is achieved using a look-up
advertiser starts advertising before the scanner starts scanning. table based simulation in Matlab using Algorithm 2. For eval-
We consider the advertiser starts a maximum of TAI + ρ uation of DL, the typical procedure is to fix two parameters,
time before t = 0 which is a full advertising cycle plus the usually TSI and TSW , and to vary the other one, TAI . As
maximum value of ρ as shown in Figure 1b. Before t = 0 any mentioned previously, this procedure can take a lot of time.
advertising event n has 0 probability of being received by the We have executed the algorithm for typical BLE parameter
scanner, so P cM is increased to its maximum value and the values and all their possible combinations and stored them in a
next event n is evaluated as shown in line 5 of Algorithm 2. look-up table in order to save computing time when evaluating
The average DL has been computed such that all data related use case.
a specific event can be accessed once the algorithm has been We sweep into the look-up table to determine the
executed as seen in line 18 of Algorithm 2. The foregoing aims (TSI , TSW , TAI ) configuration whose worst DL is no longer
to evaluate not only average but also worst case for parameter than a critical latency (CL) given by the use case: the
optimization purposes, as it will be discussed in Section IV, maximum communication latency that guarantees acceptable
so we don’t make use of DLadv and DLexp variables shown operation or user experience. We have set up the simulation
in Algorithm 1. Finally, average DL is calculated at line 19. such that, the worst case DL has a maximum difference of
Algorithm 2 can be used for both cases: when analyzing 5 ms below the desired value (CL). Several configurations may
BLE performance from the advertiser or scanner perspective. satisfy a given latency requirement, for example, for a required
The main difference between these two cases lies in the defi- advertiser DL of 200 ms at least a dozen of configurations can
nition of φ. When analyzing performance from the advertiser be found, with a scanner duty cycle ranging from 65% to
perspective we use φ = [0, 3TSI ] and when analyzing perfor- 100%, using TSI ≤ 400 ms and an advertiser duty cycle of
mance from the scanner perspective φ = [−(TAI + 10ms), 0] up to ≈ 5%. In this case, we select the configuration that pro-
accordingly with Figure 1. For the sake of simplicity, hence- vides the smallest energy consumption on the advertiser side.
forth when we refer to the performance from the advertiser Similarly, several configuration may satisfy a given latency
point of view, we will refer to Algorithm 1, and we’ll refer requirement when constrained DL is at the scanner side, and
to Algorithm 2 when analyzing from scanner perspective. As we selected the configuration that provides the smallest energy
it has been stated in [4], the model must be calibrated. Its consumption for the scanner.
precision lies in the choice of ∆ (φ step). We have adjusted In Table IV we give a set of parameters (TSI , TSW , TAI )
∆ for Algorithm 1 according to Eq. 1 and for Algorithm 2 obtained for a required latency CL between 200 ms and 1 s,
according to Eq. 2. The accuracy of the estimated advertiser when evaluating BLE from the advertiser perspective (advCL)
and scanner DL, for both average and worst case values, are and from scanner perspective (scaCL).
experimentally validated as explained in Section V.
 V. E XPERIMENTAL MODEL VALIDATION
1ms if 30ms ≥ TSI ≤ 100ms
∆= 5ms if 100ms < TSI < 640ms (1) A. Testbed Setup
93.6ms if TSI ≥ 640ms
 We have validated the proposed model on a real testbed
10µs if 20ms ≥ TAI ≤ TSW + ρ developed on top fo WalT. WalT is a cheap, reproducible and
∆= 19.9µs if TSW + ρ < TSI < 6s (2)
29.9µs if TSI ≥ 6s highly configurable platform for network experiments devel-
oped at Univ. Grenoble Alpes [12]. It is basically composed
 Table IV: Parameter Optimization Respect to CL results show the same pattern as the results obtained from
simulations of the model. As said earlier, details statistics
Respect to advCL Respect to scanCL
for simulation results were omitted, since they fit perfectly
CL TAI TSI TSW TSI TSW
to experimental results, except for extreme mean and max
200 ms 190 ms 400 ms 300 ms 400 ms 300 ms values at the peaks, they are just more regular as you would
300 ms 290 ms 500 ms 400 ms 600 ms 400 ms
400 ms 390 ms 60 ms 50 ms 800 ms 500 ms
expect since they were obtained from an ideal system, with
500 ms 490 ms 500 ms 500 ms 900 ms 600 ms a very large number of trials. In both cases, the peaks are
600 ms 590 ms 300 ms 300 ms 1.28 s 640 ms exactly at the same location, which means that the model and
700 ms 690 ms 200 ms 200 ms 2.56 s 1.28 s
800 ms 790 ms 70 ms 60 ms 3.2 s 1.28 s
experimentation behave similarly. However, we can see that
900 ms 890 ms 2.56 s 1.92 ms 3.84 s 1.28 s there is a significant difference in the extreme values. This
1s 990 ms 1.92 ms 1.28 ms 4.48 s 1.28 s can be explained easily by the fact that we drive experiments
from user space, where high precision timing is impossible
to achieve. This introduces a high variability in the results
of 6 different elements: clients from which to access the where the curve is very steep since a small imprecision in the
server and the nodes, the server which is the brain of the timing, can result in a large variation in the measured delay.
platform, which interacts with nodes on which we run or drive We notice on the graph that the distribution of values is very
experiments, interconnected with PoE switches to manage asymmetric, the median stays low, but the mean grows very
the network and restart nodes, and finally the docker hub quickly at the peaks, the 1st and 3rd quartiles are close, but
on which we store WalT images containing full reproducible the 0.05 and 0.95 quantiles show much more variability: there
experiments. In our case, nodes are Raspberry Pis with one are few occurrences where the DL is very large but they have
Asus USB-BT400 Bluetooth device attached, compatible with a very big impact. Also, these occurrences are catastrophic for
Bluetooth 4.0 Core, running Debian with the Linux BlueZ the user experience, we reached a worst case scanner discovery
stack. latency of 255 s in our experiments, more than 4 min!
We have configured two nodes to play advertiser and
scanner role. They both have a set of scripts that drive the 5
Testbed q5-95
Testbed q25-75
Testbed mean
Average Discovery Latency (s)

communication between the two BLE devices. One task runs 4 Testbed median
Simulation mean

on the advertiser and starts the scanner through the wired 3

network with a delay, before or after it itself starts advertising, 2

depending on the scenario. The advertiser contains all the 1

data related to TSI − TSW − TAI configurations and iterates 0

0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400

over TAI with a given step, repeating every measurement a Tai (ms)

configurable number of times (100 in this case). While running Figure 2: Advertiser Discovery Latency with Algorithm 1
the experiments we capture all the timestamped Bluetooth
traffic with tshark for further analysis, mainly extracting the
discovery latency. Docker images containing all the software 40
Testbed q5-95
Testbed q25-75
Testbed mean
Average Discovery Latency (s)

to run the exact same experiment will be made available. Testbed median
Simulation mean
30

B. Experimental Results 20

Figure 2 and 3 show the results obtained on the testbed 10

compared to results obtained by simulation, using respectively 0

0 2000 4000 6000 8000 10000
Tai (ms)
Algorithm 1 with the modifications mentioned in Section III
and Algorithm 2, for TSI = 400 ms and TSW = 300 ms. Figure 3: Scanner Discovery Latency with Algorithm 2
For the sake of legibility, we show detailed statistics only for
testbed results and only the mean for simulation. In Figure 2 Once the model has been validated we proceed to vali-
the advertiser is peripheral, and in Figure 3 the scanner is date our parameter optimization method. We have tested the
peripheral. High average DL is obtained near the values of TAI configurations contained in Table IV which, as mentioned
that are multiples of TSI , corresponding to the peaks observed in Section IV, are based on the worst DL of a TAI -TSI -
in both figures, as it was observed in previous work too [4]. TSW configuration, such that it does not exceed a specific
Around these values, if an advertising event starts between CL. Table V and VI show the resulting worst DL obtained
the end of the scanning window and the beginning of the next from both simulation and experiments, corresponding to the
scanning event, then the next advertising event is very likely to configurations that provide best energy consumption on the
be missed, as the random delay ρ is not enough to compensate advertiser side (Table V) and scanner side (Table VI). As it
the difference between TSI and TSW . The peaks should be can be seen, results are very accurate with a maximum error of
avoided in configurations, as very large average and worst case ≈ 3% of our simulations compared to the experimental results.
DL are obtained for these configurations, resulting in higher This validates our DL model and parameter optimization, and
energy consumption and bad user experience. Experimental thus energy consumption for any BLE manufacturer.
 Table V: Parameter Optimization respect to advertiser CL Unknown
Far
Worst case
CL TSI TSW TAI simul. expe. error

200 ms 400 ms 300 ms 190 ms 198.1 ms 204.40 ms 3.08 %

300 ms 500 ms 400 ms 290 ms 298.1 ms 303.51 ms 1.78 % Near

400 ms 60 ms 50 ms 390 ms 398.1 ms 406.02 ms 1.95 % Best case

500 ms 500 ms 500 ms 490 ms 498.1 ms 505.82 ms 1.53 %
Inmediate
600 ms 300 ms 300 ms 590 ms 598.1 ms 596.31 ms 0.30 % 30m 2m 0.5m
60m Beacon 10.9m
700 ms 200 ms 200 ms 690 ms 698.1 ms 706.19 ms 1.15 %
800 ms 70 ms 60 ms 790 ms 798.1 ms 803.81 ms 0.70 %
900 ms 2.56 s 1.92 ms 890 ms 898.1 ms 905.61 ms 0.83 %
1s 1.92 ms 1.28 ms 990 ms 998.1 ms 1.006 s 0.79 %

Table VI: Parameter Optimization respect to scanner CL

CL TSI TSW TAI simul. expe. error
Figure 4: iBeacon regions. Worst case and best case walking
200 ms 400 ms 300 ms 190 ms 204.2 ms 200.76 ms 1.68 %
300 ms 600 ms 400 ms 290 ms 304.2 ms 306.76 ms 0.83 % distance around iBeacon
400 ms 800 ms 500 ms 390 ms 404.2 ms 403.79 ms 0.10 %
500 ms 900 ms 600 ms 490 ms 504.2 ms 504.82 ms 0.12 %
600 ms 1.28 s 640 ms 590 ms 604.2 ms 601.83 ms 0.39 %
700 ms 2.56 s 1.28 s 690 ms 704.2 ms 688.89 ms 2.17 %
In accordance with the above mentioned, we have consid-
800 ms 3.2 s 1.28 s 790 ms 804.2 ms 801.84 ms 0.29 % ered a maximum acceptable latency CL of 4.36s. In addition,
900 ms 3.84 s 1.28 s 890 ms 904.2 ms 892.81 ms 1.26 % with the aim of providing reliability in case of packet loss
1s 4.48 s 1.28 s 990 ms 1004.2 s 987.87 ms 1.63 %
where an entire period would have to pass before the packet
can be received, we consider CL = 2.18s, thus ensuring that
the packet can be received while the user is still in range with
VI. A PPLICATION TO REALISTIC SCENARIOS the beacon. This is the CL reference value used for parameter
In order to provide results for typical IoT scenarios based optimization. The chosen configuration is shown in Table VII.
on our parameter optimization, we propose two test cases Since the advertiser is considered to be active permanently,
and compare them with results when using recommended SIG we do not consider dynamic reconfiguration after 30 s as in
Profiles configurations, more specificaly the Proximity Profile the SIG Profiles. We evaluate performance of this use case
[9]. These two test cases are a retail store and a medical based on Algorithm 2, where the advertiser is considered to
telemetry system which are described below. be available waiting to be discovered by the scanner. Table VII
shows the simulation results for average energy consumption
(Eavg ), average DL (DLavg ), worst case DL (DLwc ) and
A. Retail store use case
lifetime for advertiser and scanner of a TI CC2540 and a ST
This test case targets iBeacon technology (a Proximity BlueNRG. These values represent the typical expense during
application). A device which generates iBeacon advertisements one connection between advertiser and scanner during ND.
is called beacon. Beacons establish a region around them by Our proposed configuration meets the requirements of the use
iBeacon signals as shown in Figure 4. A device supporting case, while implementing a duty cycle that provides a longer
an iBeacon application can determine if it has entered or lifetime for both scanner and advertiser.
exited from the region, and can approximate its distance to the Using the CC2540 device from TI, advertiser (beacon)
beacon via signal strengths. Beacons transmit advertisement lifetime can be 105 times higher, compared with recommended
data frames containing information about what’s on sale so the configurations from the Bluetooth SIG. At the same time pro-
user can find the item and receive extra information about it. viding a scanner DL that satisfies application requirements and
Supposing the user is walking around the beacon, the time to providing a good user experience. When using ST BlueNRG
pass near it must be considered to ensure a connection between device, advertiser lifetime can be 89 times higher using our
the beacon (advertiser) and the smartphone (scanner); so the proposed configuration compared to the Bluetooth SIG and
user can successfully receive the notification. scanner lifetime can be twice higher with our proposition
In the worst case, where the user walks by 29.5m far from for both TI and ST devices. Scanner results are obtained
the beacon, for a total walking distance of 10.9m as shown based on a 225 mAh coin cell battery, and although the
in Figure 4, the user has a maximum time to establish a scanner is expected to be running on a smartphone which
connection of 11.22 s if the user walks at speed of 3.5 km/h has higher capacity, this proves that smartphone battery usage
[13] or 4.36 s at a speed of 9 km/h [14]. The user passes can be more efficient. Additionally, we can see that using the
through the region at a constant speed in a straight line Bluetooth SIG profile, it would be necessary to connect the
without stopping. BLE on the user’s smartphone is active at beacon to a permanent source of energy. In contrast, with our
the moment to enter the region. For performance evaluation, proposition, the beacon can operate for more than 2 years on
we suppose ideal channel conditions and consider the worst a single coin cell battery when using ST BlueNRG, which
case. achieves the longest lifetime.
Table VII: Retail store use case DL, energy consumption and Table VIII: Medical telemetry system DL, energy consumption
lifetime results and lifetime results
TI BlueNRG TI BlueNRG
SIG configurations Advertiser Scanner Advertiser Scanner SIG configurations Advertiser Scanner Advertiser Scanner
TSI = 60ms, TSW = 30ms Eavg 231.16µJ 1.1mJ 66.711µJ 452.11µJ TSI = 60ms, TSW = 30ms Eavg 210.27µJ 3.7mJ 63.32µJ 1.6mJ
TAI = 20ms DLavg 34.9ms 22.5ms 34ms 20ms TAI = 20ms DLavg 26.3ms 131.9ms 25.4ms 129.3ms
DLwc 51.6ms 34.2ms 50.7ms 31.7ms DLwc 103.1s 285ms 102.1ms 282.4ms
Lifetime 3.82 days 1.76 days 9.55 days 2.56 days Lifetime 3.82 days 1.76 days 9.55 days 2.56 days
TSI = 60ms, TSW = 30ms Eavg 202.38µJ 1.24mJ 57.78µJ 476.78µJ TSI = 60ms, TSW = 30ms Eavg 208.33µJ 3.9mJ 62.78µJ 1.7mJ
TAI = 30ms DLavg 41ms 23.6ms 40.1ms 21.1ms TAI = 30ms DLavg 35.4ms 140.3ms 34.5ms 137.7ms
DLwc 71.6ms 39.2ms 70.7ms 36.6ms DLwc 143.1ms 337.7ms 142.1ms 335.1
Lifetime 5.54 days 1.76 days 13.82 days 2.56 days Lifetime 5.54 days 1.76 days 13.82 days 2.56 days
Proposed in this work Advertiser Scanner Advertiser Scanner Proposed this work Advertiser Scanner Advertiser Scanner
TSI = 10.24s, TSW = 2.56s Eavg 192.11µJ 58.4mJ 60.07µJ 25.7mJ TSI = 500ms, TSW = 300ms Eavg 145.43µJ 204mJ 53.5µJ 88.2mJ
TAI = 2.2s DLavg 2.17s 1.11s 2.22s 1.11s TAI = 10.24ms DLavg 4.1s 7.03s 4.09s 7.06
DLwc 4.41s 2.21s 4.41s 2.21s DLwc 10.25s 21.83ms 10.25s 21.82s
Lifetime 1.1 years 4.05 days 2.32 years 5.62 days Lifetime 4.27 years 1.46 days 7.36 years 2.11 days

B. Medical telemetry system use case critical latency. We provide results for DL, energy consump-
tion and battery lifetime for two typical IoT test cases: retail
We now analyze a medical telemetry use case where an in-
store and medical telemetry. We show that we can obtain huge
body implanted pacemaker uses BLE to connect to a mobile
gains compared to SIG Profiles recommended configurations
device or other access point supporting BLE, to transmit
and provide multi-year lifetime instead of a few days or weeks.
information to a medical institution. The implanted sensor
operates on a 225 mAh coin cell battery and plays the beacon R EFERENCES
or peripheral role (advertiser), so the challenge is to extend [1] Bluetooth SIG. Specification of the Bluetooth System. 02 December
battery lifetime as long as possible, as it has to remain active 2014. Core Version 4.2. [Online]. Available: https://www.bluetooth.
permanently. A connection is required to collect data from com/specifications/bluetooth-core-specification/legacy-specifications
[2] S. Kamath and J. Lindh, “Measuring Bluetooth Low Energy Consump-
the pacemaker to the smartphone or other mobile device via tion,” Texas Instruments, Tech. Rep. Application Note AN092, 2010.
BLE. The key parameters to be evaluated are the connection [3] J. Liu, C. Chen, and Y. Ma, “Modeling neighbor discovery in bluetooth
time on the user side and its energy consumption during ND. low energy networks,” IEEE communications letters, vol. 16, no. 9, pp.
1439–1441, 2012.
The scanner is active when sending data to the hospital is [4] P. Kindt, D. Yunge, R. Diemer, and S. Chakraborty, “Precise
required, whose frequency can range from every 3 months Energy Modeling for the Bluetooth Low Energy Protocol,” CoRR, vol.
to every 12 months [15]. The scanner is in range at the abs/1403.2919, 2014. [Online]. Available: http://arxiv.org/abs/1403.2919
[5] K. Cho, W. Park, M. Hong, G. Park, W. Cho, J. Seo, and K. Han,
moment of connection. We base the parameter optimization “Analysis of latency performance of Bluetooth low energy (BLE)
and performance evaluation on Algorithm 1, since the adver- networks,” Sensors, vol. 15, no. 1, pp. 59–78, 2014.
tiser plays the peripheral role at the application level and has [6] W. S. Jeon, M. H. Dwijaksara, and D. G. Jeong, “Performance Analysis
of Neighbor Discovery Process in Bluetooth Low-Energy Networks,”
constrained DL and energy consumption. We suppose ideal IEEE Transactions on Vehicular Technology, vol. 66, no. 2, pp. 1865–
channel conditions. 1871, 2017.
Table VIII shows DL, energy consumption and lifetime [7] Bluetooth SIG. Adopted Bluetooth Profiles, Services, Protocols
and Transports. 2018. [Online]. Available: https://www.bluetooth.com/
results. With our proposition, advertiser lifetime can be 281 specifications/gatt
times higher than with the SIG Profile when using ST Blu- [8] A. Liendo, D. Morche, R. Guizzetti, and F. Rousseau, “Efficient Blue-
eNRG, reaching a lifetime of 7 years and 4 months with an tooth Low Energy Operation for Low Duty Cycle Applications,” in IEEE
International Conference on Communications, 2018.
average advertiser DL of ≈ 4 s, which is considered a good [9] Bluetooth SIG. Proximity Profile specification. 14 July 2015. Version
user experience for this use case. The SIG Profile provides 1.0.1. [Online]. Available: https://www.bluetooth.com/specifications/gatt
a lifetime of 9.5 days. This proves that BLE is a suitable [10] J. Liu and C. Chen, “Energy analysis of neighbor discovery in Bluetooth
Low Energy networks,” Nokia.(nd), 2012.
technology for this kind of applications when using an optimal [11] STMicroelectronics. BlueNRG Current Consumption Estimation
configuration. Lifetime on the scanner side improvement is Tool. 2017. Version 1.4. [Online]. Available: http://www.st.com/
not achieved with the proposed configuration as extending the en/embedded-software/stsw-bnrg001.html
[12] Université Grenoble Alpes, Grenoble INP / UJF. WalT - Wireless
advertiser lifetime is the priority. Testbed. 2013-2014. [Online]. Available: http://walt.forge.imag.fr/
[13] R. Knoblauch, M. Pietrucha, and M. Nitzburg, “Field studies of pedes-
VII. C ONCLUSIONS trian walking speed and start-up time,” Transportation Research Record:
Journal of the Transportation Research Board, no. 1538, pp. 27–38,
We have extended an existing model [4] and proposed 1996.
a novel parameter optimization method to compute optimal [14] R. W. Bohannon, “Comfortable and maximum walking speed of adults
parameters for BLE devices depending on application require- aged 20—79 years: reference values and determinants,” Age and ageing,
vol. 26, no. 1, pp. 15–19, 1997.
ments. This model as well as the parameter optimization are [15] H. Burri and D. Senouf, “Remote monitoring and follow-up of pace-
validated experimentally. We show that BLE is suitable for a makers and implantable cardioverter defibrillators,” Europace, vol. 11,
wide range of IoT applications if parameters are chosen with no. 6, pp. 701–709, 2009.
care. We obtain the appropriate TSI −TSW −TAI configuration
that guarantees a Discovery Latency which satisfies specific
requirements for any use case, such as maximum acceptable