Galinin A 2017
Galinin A 2017
Galinin A 2017
PII: S1389-1286(17)30182-2
DOI: 10.1016/j.comnet.2017.04.051
Reference: COMPNW 6190
Please cite this article as: Olga Galinina, Sergey Andreev, Mikhail Komarov, Svetlana Maltseva, Lever-
aging Heterogeneous Device Connectivity in a Converged 5G-IoT Ecosystem, Computer Networks
(2017), doi: 10.1016/j.comnet.2017.04.051
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
T
a Dept. of Electronics and Communications Engineering, Tampere University of Technology
b School of Business Informatics, National Research University Higher School of Economics
IP
CR
Abstract
As mobile communications technology is completing its fifth-generation (5G)
cycle, it becomes increasingly capable of supporting novel usage scenarios with
US
stringent performance requirements. Beyond seamless broadband connectivity
for humans, 5G systems are preparing to enable a wide range of machine-type
applications, thus advancing the vision of the Internet of Things (IoT). Fa-
cilitated by the rapidly converging 5G-IoT ecosystem, next-generation indus-
AN
trial IoT services, however, pose unprecedented research problems, primarily
along the lines of providing wireless connectivity with ubiquitous availability,
extreme reliability, and ultra-low latency. To this end, the first part of this
paper1 presents a concise update on the novel requirements and challenges in
the context of the emerging Industrial Internet infrastructure. In the second
M
1 The work of O. Galinina was supported in part with a personal research grant by the
Finnish Cultural Foundation and in part by a Jorma Ollila grant from Nokia Foundation.
The work of S. Andreev was supported in part by a Postdoctoral Researcher grant from the
Academy of Finland and in part by a Jorma Ollila grant from Nokia Foundation.
T
offering improved quality of experience across a diverse range of services, and
(iv) providing better affordability and reduced costs [3].
IP
With these goals in mind, and anticipating the global IMT traffic to grow
on the level of 10 to 100 times from 2020 to 2030 [4], the International Telecom-
munication Union (ITU) has adopted a mission to outline the objectives of the
CR
future development of IMT for 2020 and beyond [5]. In their vision, the core
development of IMT-2020 specifications is expected to complete around 2020
with the subsequent deployment phase to begin [6]. Recall that it has taken
a total of 15 years (1985-2000) to finalize the IMT-2000 work and launch the
US
compliant third-generation (3G) mobile networks, while the follow-up cycle on
IMT-Advanced and fourth-generation (4G) systems shrank to around 12 years
(2000-2012). With further reduced timeframe of only about 8 years (2012-2020),
the emerging fifth-generation (5G) broadband technology is prepared to rapidly
AN
innovate all aspects of its design [7]: from advanced spectrum utilization and
physical layer techniques to flexible configurability and network management.
In light of the above, the core “big three” technology trends comprising the
5G framework [8] are (i) further network densification with small cells of dif-
M
spectral efficiency. The synergistic use of these solutions should allow achiev-
ing peak data rates of 10 (and sometimes up to 20) Gbps as well as enabling
user experienced data rates of 100 Mbps for wide area coverage cases and 1
Gbps in indoor hotspot cases. As of late 2015, the 3rd Generation Partnership
PT
With the help of the 3GPP’s New Radio, and fueled by the rapid uptake of
capable smartphones and tablets, the forthcoming 5G ecosystem will accelerate
the seamless delivery of advanced wireless communication applications to con-
sumers [10], including location-based, cloud, and social services, high-definition
AC
2
ACCEPTED MANUSCRIPT
T
The envisaged 5G-grade M2M usage scenarios comprise the following two cate-
gories:
IP
1. Massive machine-type communications that feature a very large number
of connected objects, from low-complexity devices to more advanced ap-
CR
pliances with enhanced capabilities: monitoring of smart grid and agri-
cultural systems, connected sensors, actuators, wearables, cameras, and
vehicles.
2. Ultra-reliable and low-latency communications that include machine-centric
US
connectivity with real-time constraints: unmanned cars and drones, indus-
trial manufacturing and production processes, real-time traffic control and
optimization of transportation systems, emergency and disaster response.
In the rapidly emerging converged 5G-IoT ecosystem [14], various networked
AN
devices will operate at different carrier frequencies. The latter will embrace
higher frequency ranges and wider channel bandwidths than today. It is presently
a common consensus that in the near future, to satisfy the diversity of IMT-
2020 requirements, multiple RATs will need to work in concert [15]. These will
M
include (i) LTE-Advanced evolution and New Radio solutions by 3GPP, (ii) fast
growing family of protocols by the Institute of Electrical and Electronics Engi-
neers (IEEE), such as the widely deployed WiFi technology, and (iii) numerous
short-range radio platforms. However, such heterogeneous connectivity poses
ED
unprecedented challenges, which are primarily aligned with the need of accom-
modating “high-end” IoT devices having stringent performance guarantees. In
particular, providing with the required levels of over-the-air reliability remains
a cumbersome and insufficiently understood research problem, and we continue
PT
3
ACCEPTED MANUSCRIPT
T
IP
CR
US
AN
M
Figure 1: Example 5G-grade Industrial Internet applications for the factory automation.
ED
modularity [21], which are believed to produce altogether the much needed
technology-agnostic user experience. Driven by the example of industrial au-
tomation – that focuses primarily on process automation in manufacturing,
quality control, and material handling – a multi-year strategic initiative has
CE
been launched in Europe, named “Industrie 4.0” [22]. Sponsored by the Ger-
man government, it brings together the public and private sectors as well as the
academia to construct a holistic vision and a harmonized action plan for deliver-
ing wireless communications technology to the industrial sector. In response to
AC
this, China has recently proposed its “Made in China 2025” strategy to advance
the nation-wide integration of digital technologies and boost industrialization.
As global research on Industrial Internet is decisively gearing up, it is com-
monly anticipated that 5G technology will provide support for the needs of
Industrie 4.0 and similar initiatives [23], [24]. A major challenge in this context
is bringing higher levels of network and service availability to distant and chal-
lenging locations, which may be demanded by the advanced factory automation
4
ACCEPTED MANUSCRIPT
and remote control applications [25]. Combining mobile edge cloud technology
with virtual or augmented reality for sensory and haptic control [26], future
Industrial Internet solutions gain momentum to facilitate a dramatic paradigm
shift from opportunistic content delivery to ultra-reliable skill-set delivery net-
works [27], with seconds of outage per year. However, the current Internet
infrastructure largely follows the “best effort” principles and significant efforts
T
are necessary to support such systems [28], most of which require reliable com-
munications, while some real-time control applications also demand low-latency
IP
data transmissions [29].
To this end, the entire protocol stack of a haptic Industrial Internet sys-
tem will need to be reworked [30] to achieve carrier-grade access reliability.
CR
Broadly, reliability is related here to the probability of guaranteeing a required
performance over a certain time interval and for particular operating condi-
tions. More specifically, to improve link-level performance it may be desirable
to refrain from utilizing retransmission-based mechanisms, such as automatic
US
repeat request (ARQ) and hybrid ARQ (HARQ), since they typically incur
additional unwanted latency [27]. An alternative approach to offering higher
reliability without relying on the (H)ARQ schemes is to employ diversity in the
frequency or spatial domain by leveraging multiple uncorrelated wireless links.
AN
Therefore, simultaneous connectivity on several radio channels (often named
multi-connectivity [31]) becomes an important technique to approach carrier-
grade reliability for future tactile applications.
Given that cellular systems are heavily regulated and natively coordinate
M
wireless interference, their more reliable communication links are often preferred
for baseline connectivity and control. However, the emerging Industrial Inter-
net applications will require a tighter interworking between the current 3GPP
LTE technologies and other RATs across both public and private networks.
ED
5
ACCEPTED MANUSCRIPT
T
communications may disrupt such systems and ultimately render them useless
or even harmful for people. To effectively combat the unpredictable random-
IP
ness of a radio link when delivering reliability to these advanced use cases, it is
important to employ and combine alternative transmission channels across var-
ious RATs. By flexibly leveraging such heterogeneous multi-connectivity, the
CR
5G ecosystem will become equipped with efficient means to support stringent
Industrial Internet applications, far beyond the conventional mobile Internet.
However, the question of how to combine these diverse radio technologies re-
mains essentially open for most of the target IoT scenarios.
US
Recent research in [33] outlines a novel high-level framework that allows a
running application to only utilize a wireless link if the underlying communication-
related conditions are sufficiently favorable. It is thus deemed possible to deter-
mine boundary criteria for the transmission success, while formally expressing
AN
availability as the expected presence or absence of link reliability at the time
of initiating a data session. Along these lines, link reliability is connected to
its ability of transmitting or receiving a particular volume of data successfully
within a particular latency budget. Further, system reliability characterizes its
capability to detect and indicate the existence of link reliability to the applica-
M
tion as well as guarantee its presence as often as possible. With these mecha-
nisms, a multi-connectivity application may be offered sufficient and predictable
success rates within the necessary constraints, as well as be timely warned about
ED
Practical use cases may include video surveillance, urban monitoring, object
tracking, wireless healthcare, wearables, and connected cars, among others. The
“high-end” device in question has access to a set of RATs (potential options
include 3GPP and IEEE protocols: the 802.11 family, LTE-M, EC-GSM, NB-
CE
IoT, ZigBee, LoRa, Bluetooth, 6LoWPAN, etc.), each of which has its unique
characteristics in terms of transmission rate and link reliability. In the simplest
setup, the multi-connectivity IoT device may select the most appropriate radio
technology by comparing its application’s data rate requirement against the
AC
6
ACCEPTED MANUSCRIPT
T
signaling protocols). Based on the actual situation, the device may select ei-
ther one channel that is “the best” in some sense, or choose several links if this
IP
combination is more beneficial.
As a result, our approach in this work is novel and significantly differs from
that in [33], as it inherently relies on the existence of multiple RATs and explores
CR
a possibility to utilize them at the same time in order to improve communication
reliability. While taking into account the target data rate as required by the
running application, we implicitly consider the variance in the achieved data
rate. Our mathematical problem is generally formulated in terms of convex
US
quadratic optimization that can be solved efficiently to provide with the weights
(shares), according to which the application data could be distributed optimally
across the available links. For independent channels, we derive an algorithm
based on closed-form expressions, which results in exact calculation of the said
AN
weights. Our proposed approach allows to sufficiently decrease the variance in
the total achieved data rate as well as outperforms intuitive heuristic strategies
of preferring “the best” transmission channel.
In our proposed model, the optimal link selection procedure of a multi-
M
connectivity IoT device is based on two criteria: (i) the expected rate has to
exceed a certain target bitrate threshold dictated by the application-level re-
quirements and (ii) the resulting channel reliability should be the highest. In
the sequel, we understand the term “channel reliability” in the sense of the
ED
minimum rate variance to offer e.g., a better rate-centric performance, but our
formulation may be extended to other interpretations of this parameter. In what
follows, we first outline the proposed optimization framework on the conceptual
level, that is, without relying on specific details of an underlying air interface or
PT
method.
7
ACCEPTED MANUSCRIPT
across the financial community, the classical portfolio theory has been seldom
applied for wireless communication related formulations. However, we found it
to be naturally suited to our envisaged system model (see Fig. 2) and thus a
powerful tool to utilize.
T
IP
CR
US
AN
Figure 2: Envisioned device-centric multi-connectivity system model.
Let us assume that for N available radio channels, certain predictors provide
information on the expected achievable data rate r ∈ RN (equivalent to a port-
M
folio return) as well as the matrix of covariance between the channels V ∈ RNxN
(equivalent to portfolio volatility). We note that the said predictors are adopted
in the “black box” setting as they are only assumed to provide information on
ED
target optimization problem to minimize the variance of the total device data
rate under the constraint on the minimum expected rate r 0 ∈ [r N ,r 1 ] as:
minimize wT V w
w
CE
subject to rT w ≥ r 0 , (1)
1T w = 1, w 0,
where w is the sought vector of channel weights (that is, to what extent each of
AC
the available RATs is used to transmit a part of the overall data stream) and
1 = (1, ..., 1)T . Hereinafter, we refer to a particular channel i,i = 1, .., N by its
corresponding index i, while index 0 has a more general meaning and is employed
to denote the bitrate constraint r 0 . The notation implies that for all the vector
components the condition ≥ holds. Our described formulation is that of convex
quadratic programming due to the fact that the covariance matrix is always
symmetric and positive semidefinite. Hence, a numerical solution is feasible for
8
ACCEPTED MANUSCRIPT
T
PN 2 2
the initial objective function as f (w) = i= 1 σ i wi . Reformulating our original
optimization problem by transforming the inequality constraint with the intro-
IP
duction of the slack variable x ≥ 0 (in the sense of the data rate surplus over
the required minimum), we arrive at the following equivalent formulation:
CR
PN 2 2
minimize i=
w 1 σi wi
PN
subject to i=1 r i wi − x = r 0 , (2)
PN
w
i=1 i = 1, w i ≥ 0, x ≥ 0.
US
Lemma 1. Accounting for the condition r 0 ∈ [r N ,r 1 ], there exists at least one
feasible solution point to the target optimization problem.
Proof. For example, consider a point w = (1, 0, ..., 0).
AN
In addition, we may assume that r i > r i +1 and σi > σi +1 for any channel
number i. Otherwise, if for some index i 0 , r i = r i +1 or σi ≤ σi +1 , then the
channel i 0 is preferred to (dominates) i 0 + 1 in any conditions and we may
exclude the channel i 0 + 1 from further consideration.
M
In what follows, we solve the above problem formulation for two cases of
separate interest: (i) when the number of channels N = 2 and the solution is
w = (w1 , w2 ), and (ii) when N ≥ 3 (a more general case), so that we may remove
two components of the vector via the equality constraints. Due to the convexity
ED
of our objective function, the minimum for the unconstrained problem has to
be sought at the stationary point, i.e., where the gradient is zero. However, the
formulation in both cases requires non-negative solutions for all the components
wi and x. While for the former the non-negativity of the stationary point may
PT
be proven explicitly, the variable x may turn out to have either sign. Hence,
the solution for both considered cases is further split into two parts: for non-
negative x the obtained stationary point would be a solution, while for negative
x we continue calculations by seeking the minimum point at the border of the
CE
9
ACCEPTED MANUSCRIPT
T
IP
CR
Figure 3: Illustration of our approach for N = 2: if stationary point x (1) is feasible, then it is a
US
solution; otherwise, we seek for the minimum at the border of feasible set (x (2) is a solution).
Since f (x) is a convex function, we may establish the minimum point from
the expression f 0 (x) = 0 as:
f 0 (x) = 2σ12 (r 0 − r 2 + x) + 2σ22 (x + r 0 − r 1 ) , (6)
and from f 0 (x) = 0 we obtain:
PT
−σ 12 (r 0 −r 2 ) +σ 22 (r 1 −r 0 )
x= , (7)
(σ12 +σ22 )
where the corresponding weights are:
CE
−r 2 σ 22 +σ 22 r 1 σ 22
w1 = (σ 12 +σ 22 )(r 1 −r 2 )
= σ 12 +σ 22
> 0,
−σ 12 r 2 +r 1 σ 12 σ 12 (8)
w2 = (r 1 −r 2 )(σ 12 +σ 22 )
= σ 12 +σ 22
> 0.
AC
10
ACCEPTED MANUSCRIPT
T
notation for the resultant summations:
IP
PN PN PN
Sσ = i=3 σi2 wi2 , Sr = i=3 r i wi , Sw = i=3 wi , (10)
and continue by expressing the first two weights from the rate constraint and
CR
PN
the normalization condition i= 1 wi = 1:
w1 = r0 + x −Sr1r −r
−r 2 +r 2 S w
,
2
−r 0 − x + S r +r 1 −r 1 S w (11)
w2 = 1 − w1 − Sw = .
US r 1 −r 2
Substituting the expressions (11) into the objective function and considering
the constraints wi ≥ 0, we may write down the function of N − 2 variables as
follows:
AN
f (w0 ; x) = Sσ + σ12 (r0 + x −S r −r 2 +r 2 S w )
+ σ22 (−r0 − x +(rS1r−r
+r 1 −r 1 S w )
2 2
(r 1 −r 2 ) 2 2)
2
s.t. r 0 + x − Sr − r 2 + r 2 Sw ≥ 0,
− r 0 − x + Sr + r 1 − r 1 Sw ≥ 0, wi ≥ 0, x ≥ 0,
M
where w0 = (w3 , ..., w N ; x) and r 1 > r 2 . We also note that the following holds:
∂S w ∂S r ∂Sσ
wi = 1, wi = ri , wi = 2σi2 wi .
ED
Then, the stationary points may be defined as zeros of the gradient as:
∂ f (w0 ; x)
wi = 2σi2 wi + 2σ22 (r(r1i−r
−r 1 )
2)
2 (−r 0 − x + Sr + r 1 − r 1 Sw )
(−r i +r 2 )
(12)
+2σ12 (r1 −r 2 (r 0 + x − Sr − r 2 + r 2 Sw ),
PT
2)
and
∂ f (w0 ; x) 1
= 2σ22 (r1 −r 2 (−r 0 − x + Sr + r 1 − r 1 Sw )
x 2)
CE
1
−2σ1 (r1 −r2 )2 (r 0 + x − Sr − r 2 + r 2 Sw ).
2
σ 22 +σ 12
We then substitute the above expression into (12) and also equate it to zero
as:
(1 −S w )(σ 22 r 1 +σ 12 r 2 )
0 = σi2 wi + σ12 (−r i +r 2 )
(r 1 −r 2 ) 2 2
σ 2 +σ 12 − r 2 + r 2 Sw
(1 −S w )(σ 22 r 1 +σ 12 r 2 )
+σ22 (r(r11−r
−r i )
2)
2
σ 2 +σ 2
− r 1 + r S
1 w .
2 1
11
ACCEPTED MANUSCRIPT
T
P
N 2 2
1 σ1 σ2
Sw = σ 2 σ 2 +σ 2
(1 − Sw ),
IP
i=3 i 2 1
CR
2σ2
σ1
R0 2
2 +σ 2
σ2 R 0 σ 12 σ 22
Sw = 1
2σ2
σ1
= σ 22 +σ 12 + R 0 σ 12 σ 22
, (15)
1+ R 0 2
2 +σ 2
σ2 1
P
N
US
1
where R0 = σ i2
.
i=3
Finally, we may find the optimal weights from (14) as follows:
1 σ 12 σ 22
wi = > 0. (16)
AN
σ i2 σ 12 +σ 22 + R 0 σ 12 σ 22
We remark that the latter holds for any i = 1, ..., N, which for i = 1, 2 may
be easily verified by substituting (15) and (13) into (11).
M
N
P ri
σ 22 r 1 +σ 12 r 2 +σ 12 σ 22 2
i=3 σ i
x= N
P
− r 0 < 0, (17)
σ 22 +σ 12 +σ 12 σ 22 2
1
i=3 σ i
PT
then we should further seek for a point, when the minimum bitrate constraint
becomes active, that is, x = 0.
Rewriting the above objective function together with its gradient for the
case with no slack variable x added, we obtain:
CE
1 −r 2 ) )2
2
s.t. r 0 − Sr − r 2 + r 2 Sw ≥ 0,
− r 0 + Sr + r 1 − r 1 Sw ≥ 0, wi ≥ 0, x ≥ 0,
AC
and
∂ f˜(w0 ; x)
wi = 2σi2 wi + 2σ12 (−r i +r 2 )
(r 1 −r 2 ) 2 (r 0 − Sr − r 2 + r 2 Sw )
2 (−r i +r 1 )
+2σ2 (r1 −r2 )2 (r 0 − Sr − r 1 + r 1 Sw ).
For the sake of exposition, we omit here several less important technical
details, which are described as appropriate in Appendix. As a result of solving
12
ACCEPTED MANUSCRIPT
∂ f˜(w0 ; x)
wi = 0, we obtain the corresponding vector w:
σ 22 (− R 1 (r 1 +r 0 ) + R 2 +r 0 r 1 R 0 ) + (r 1 −r 2 )(r 0 −r 2 )
w1 = (r 1 − r 2 ) 2 D0 ,
σ 12 (− R 1 (r 0 +r 2 ) + R 2 +r 0 r 2 R 0 )−(r 1 −r 2 )(r 0 −r 1 )
w2 = (r 1 − r 2 ) 2 D0 , (18)
T
1 σ 12 (r 2 −r 0 )(r 2 −r i ) +σ 22 (r 1 −r 0 )(r 1 −r i )
wi = (r − r2 )2 ,
IP
σ i2 1 D0
where
PN 1 PN ri PN r i2
R0 = i=3 σ 2 , R1 = i=3 σ 2 , R2 = i=3 σ 2 .
CR
i i i
US
In this section, we aim at constructing a representative practical example
of how our proposed multi-connectivity link selection scheme applies to an ad-
vanced IoT scenario and helps manage several alternative RATs in an optimal
AN
manner to improve communications reliability subject to a particular bitrate
constraint. We emphasize that here we interpret reliability (as a system perfor-
mance indicator) in the sense of decreased variance in the achieved date rate.
To this end, let us consider a “high-end” IoT device (e.g., an industrial cam-
era), which may utilize several alternative independent RATs, that is, switch
M
between them in e.g., time-division mode. The share of time when the channel
i is exploited directly corresponds in our solution to the respective weight wi .
Note that frequency-division regime may also be assumed here, i.e., simulta-
ED
bitrate application, which requires its data rate to be at least r 0 . This parameter
is varied below to capture a range of IoT application requirements. In practice,
r 0 should be selected with a certain margin on top of the actual requirements
by the target IoT application to compensate for any unpredictable throughput
CE
in the operator’s cloud. The data communicated on all the channels is collected
and accurately combined [36] at the receiver side, where the final data rate
variance is subject to minimization.
From the potential variety of available radio technologies, we select those
that adhere to our outlined system model and enumerate them in the order of
their decreasing expected data rate (as discussed in Section 4). Let us assume
here for instance that the final set of candidate RATs comprises, e.g., three
13
ACCEPTED MANUSCRIPT
technologies (e.g., WiFi, NB-IoT, and BLE, see Fig. 4) with the vector of average
data rates given by the typical values of r = [20, 5, 1] Mbps. For convenience,
instead of the standard deviation σ, we further exploit the coefficient of variation
(i.e., the variance-to-mean ratio) γi = σr ii . The latter is a standard measure of
relative dispersion, which for the selected parameters is assumed to be γ =
[0.3, 0.3, 0.3]. We remind that the set of σi has to be sorted in the descending
T
order; otherwise, if σi > σi − 1 , the technology i should be removed from the list.
IP
CR
US
Figure 4: Considered practical multi-connectivity example.
AN
For the sake of better exposition, here and below we focus on the class of
multi-radio solutions where no additional redundancy is introduced across the
multiple RATs as well as disregard any channel switching overheads, which are
heavily technology-dependent. All these real-world factors may be taken into
account as needed with the corresponding modifications of our outlined system
M
PN
model and, more specifically, by altering the constraint i= 1 wi , which then
readily translates into scaling of w and r 0 .
ED
fact that N = 3 requires utilizing our general procedure for an arbitrary value
of N, and yet this case is still easy to visualize in the plots.
Along these lines, we implement the framework to search for our target opti-
mal solution in a home-grown MATLAB-based modeler. In order to confirm the
CE
ing timeshare of RATs by varying the data rate requirement r 0 , which reflects
a wide range of IoT application demands. Finally, we compare our proposed
solution with a simpler heuristic approach of selecting a single radio technology
based on the application-level requirements, which confirms the benefits of our
weighted solution that leverages multiple RATs.
To begin with, Fig. 5 demonstrates the resulting standard deviation sur-
face in (w1 , w2 , w3 ) coordinates as well as indicates the optimum point that is
14
ACCEPTED MANUSCRIPT
T
IP
CR
US
Figure 5: Standard deviation surface vs. traffic distribution (w1, w2, w3 ) and optimal point.
AN
obtained by applying the proposed analytical procedure for the target bitate
of r 0 = 1 Mbps. As it may be observed in the figure, the optimal solution is
an inner point within the interval of 0 ≤ w2 ≤ 1, which means that the rate
M
constraint is not critical in this case and leaves more flexibility with respect to
the rate distribution across the channels. Interestingly, since w2 > 0, w3 < 1,
the optimal solution provides a slight improvement as compared to assigning all
of the data load to the most capable third channel. Moreover, the difference
ED
between the observed standard deviation values is considerable (e.g., from 0.3
up to 6), which confirms the practical importance of solving our aimed mini-
mization problem. In the top-right corner of the figure, we also demonstrate
the corresponding surface of feasible solutions for a more demanding application
PT
types of IoT applications. Here, we illustrate the evolution of the set (w1 , w2 , w3 )
by following two distinct examples: (i) equal coefficients of variation or variance-
to-mean ratio γ = [0.3, 0.3, 0.3] (the same relative dispersion for all the RATs)
AC
15
ACCEPTED MANUSCRIPT
T
IP
CR
US
AN
Figure 6: Optimal set of (w1, w2, w3 ) vs. target rate r 0 : for equal (upper) and decreasing
(lower) coefficient of variation γ (variance-to-mean ratio).
M
values are shown in Fig. 7. For the regions of low rate demand, naturally,
there is almost no benefit in using the optimal selection, and single-channel
approach might seem to be a meaningful solution. However, as the data rate
demand grows higher than the minimal available rate (e.g., within the interval
CE
r 0 ∈ (1, 5) Mbps), it becomes reasonable to not only exploit another RAT, but
also continue offloading a part of the data to the previously employed one. This
“gradual offloading” of excess data rate in case of increased bitrate demand
(as opposed to a discrete selection of another channel) leads to the provably
AC
6. Conclusions
This paper has reviewed in detail the ongoing convergence between the 5G
and the IoT visions as well as discussed some of the enabling technologies.
16
ACCEPTED MANUSCRIPT
T
IP
CR
US
Figure 7: Standard deviation for two link selection schemes: use of a single channel that
AN
supports r 0 (green) and utilization of our optimal solution (blue).
Among other scenarios, we emphasized the industrial IoT applications and stud-
ied the anticipated impact on these, which results from such convergence. As a
M
17
ACCEPTED MANUSCRIPT
Appendix
This Appendix contains the auxiliary derivations that clarify the optimiza-
tion conducted in Section 4.
First, let us establish a solution to the following system:
∂ f˜(w0 ; x)
T
wi = 2σi2 wi + 2σ12 (−r i +r 2 )
(r 1 −r 2 ) 2 (r 0 − Sr − r 2 + r 2 Sw )
2 (−r i +r 1 )
+2σ2 (r1 −r2 )2 (r 0 − Sr − r 1 + r 1 Sw ).
IP
Denoting 1 − Sw as y and r 0 − Sr as z, we may write down the following:
CR
0 = σi2 wi (r 1 − r 2 ) 2 + σ12 (r 2 − r i )(z − r 2 y)
+σ22 (r 1 − r i )(z − r 1 y), z ≥ r 2 y, z ≤ r 1 y.
US
σi2 wi (r 1 − r 2 ) 2 = z −σ12 (r 2 − r i ) − σ22 (r 1 − r i )
+y σ12 (r 2 − r i )r 2 + σ22 (r 1 − r i )r 1 , i = 3, ..., N,
(19)
AN
σi2 wi (r 1 − r 2 ) 2 = z −σ12 (r 2 − r i ) − σ22 (r 1 − r i )
(20)
+y σ12 (r 2 − r i )r 2 + σ22 (r 1 − r i )r 1 , i = 3, ..., N.
Introducing the following notation for the sake of brevity
M
we arrive at:
ED
σi2 wi (r 1 − r 2 ) 2 = z (r i A0 − A1 ) + y ( A2 − r i A1 ) . (22)
PN 1 PN ri PN r i2
R0 = i=3 σ 2 , R1 = i=3 σ 2 , R2 = i=3 σ 2 , (23)
i i i
(
(1 − y)(r 1 − r 2 ) 2 = z (R1 A0 − R0 A1 ) + y (R0 A2 − R1 A1 ) ,
(r 0 − z)(r 1 − r 2 ) 2 = z (R2 A0 − R1 A1 ) + y (R1 A2 − R2 A1 ) .
AC
18
ACCEPTED MANUSCRIPT
T
wi = σ12 R R − R 2 ( A A − A2 )+(r −r )41+(r −r )2 (R A − 2 R A + R A )
h i ( 0 2 1) 0 2 1 1 2 1 2 0 2 1 1 2 0
−R1 A2 + R2 A1 + r 0 R0 A2 − r 0 R1 A1 + r 0 (r 1 − r 2 ) 2 (r i A0 − A1 ) (25)
IP
i
+ −r 0 R1 A0 + r 0 R0 A1 + R2 A0 − R1 A1 + (r 1 − r 2 ) 2 ( A2 − r i A1 ) ,
CR
or, then,
h
wi = σ12 D10 −R1 A0 A2 + r 0 R0 A0 A2 − r 0 R0 A21 + R1 A21 r i
i
+ −R2 A21 + r 0 R1 A21 − r 0 R1 A0 A2 + R2 A0 A2 (26)
i
and the i-th element of the corresponding summation in the first part is:
1
σ i2
−r i A0 A2 + r 0 A0 A2 − r 0 A21 + r i A21 r i
ED
1
σ2
−r i2 A21 + r 0 r i A21 − r 0 r i A0 A2 + r i2 A0 A2 = 0.
i
Consequently, we have:
h i
PT
N0 = (r 1 − r 2 ) 2 σ12 r 22 + r 0 r i − r i r 2 − r 0 r 2
h i
+(r 1 − r 2 ) 2 σ22 r 12 + r 0 r i − r i r 1 − r 0 r 1
h i
= (r 1 − r 2 ) 2 σ12 (r 2 − r 0 )(r 2 − r i ) + σ22 (r 1 − r 0 )(r 1 − r i ) .
CE
19
ACCEPTED MANUSCRIPT
σ 12 (− R 1 (r 0 +r 2 ) + R 2 +r 0 r 2 R 0 )−(r 1 −r 2 )(r 0 −r 1 )
w2 = (r 1 − r 2 ) 2 D0 ,
T
1 σ 12 (r 2 −r 0 )(r 2 −r i ) +σ 22 (r 1 −r 0 )(r 1 −r i )
wi = (r
σ i2 1
− r2 )2 D0 .
IP
Let us in conclusion prove that the denominator D0 is greater than zero.
Correspondingly, consider all the elements of D0 as:
CR
P
N 1 P N r22 P N r22 2
D1 = R0 R2 − R12 = i= 3 σ2 · i=3 σ 2 − i=3 σ 2 ≥ 0, (29)
i i i
US
Consider now the term D3 = R0 A2 − 2R1 A1 + R2 A0 and, in particular, its
i-th element in the form:
{D3 }i = {R0 A2 − 2R1 A1 + R2 A0 }i = σ12
AN
i
× σ12 r 22 − 2r i r 2 + r i2 + σ22 r 12 − 2r i r 1 + r i2
= σ12 σ12 (r 2 − r i ) 2 + σ12 σ22 (r 1 − r i ) 2 > 0.
i i
D0 = D1 D2 + (r 1 − r 2 ) 4 + (r 1 − r 2 ) 2 D3 > 0.
References
ED
[4] Report ITU-R M.2370-0, IMT traffic estimates for the years 2020 to 2030
(2015).
[5] Recommendation ITU-R M.2083-0, IMT Vision – Framework and overall
AC
objectives of the future development of IMT for 2020 and beyond (2015).
[6] A. Gupta, R. K. Jha, A survey of 5G network: Architecture and emerging
technologies, IEEE Access 3 (2015) 1206–1232.
[7] K. Lin, W. Wang, X. Wang, W. Ji, J. Wan, QoE-driven spectrum assign-
ment for 5G wireless networks using SDR, IEEE Wireless Communications
22 (6) (2015) 48–55.
20
ACCEPTED MANUSCRIPT
T
fulfill 5G requirements, in: Proc. of the VTC-Spring, 2015.
IP
[10] S. Chen, J. Zhao, The requirements, challenges, and technologies for 5G
of terrestrial mobile telecommunication, IEEE Communications Magazine
52 (5) (2014) 36–43.
CR
[11] M. Condoluci, M. Dohler, G. Araniti, A. Molinaro, J. Sachs, Enhanced ra-
dio access and data transmission procedures facilitating industry-compliant
machine-type communications over LTE-based 5G networks, IEEE Wire-
less Communications 23 (1) (2016) 56–63.
US
[12] E. Dahlman, G. Mildh, S. Parkvall, J. Peisa, J. Sachs, Y. Selen, J. Skold,
5G wireless access: Requirements and realization, IEEE Communications
Magazine 52 (12) (2014) 42–47.
AN
[13] S. Andreev, O. Galinina, A. Pyattaev, M. Gerasimenko, T. Tirronen,
J. Torsner, J. Sachs, M. Dohler, Y. Koucheryavy, Understanding the IoT
connectivity landscape: A contemporary M2M radio technology roadmap,
IEEE Communications Magazine 53 (9) (2015) 32–40.
M
(2016) 510–527.
21
ACCEPTED MANUSCRIPT
[21] M. Hermann, T. Pentek, B. Otto, Design Principles for Industrie 4.0 Sce-
narios: A Literature Review (2015).
[22] A. Varghese, D. Tandur, Wireless requirements and challenges in Industry
4.0, in: Proc. of the IC3I, 2014, pp. 634–638.
[23] S. Wang, J. Wan, D. Li, C. Zhang, Implementing smart factory of Industrie
T
4.0: An outlook, International Journal of Distributed Sensor Networks.
IP
[24] S. Wang, J. Wan, D. Zhang, D. Li, C. Zhang, Towards smart factory for
Industry 4.0, Computer Networks 101 (C) (2016) 158–168.
[25] 5G-PPP, 5G and the Factories of the Future (2015).
CR
[26] Engineering & Technology, Tactile Internet: 5G and the Cloud on steroids
(2015).
[27] M. Simsek, A. Aijaz, M. Dohler, J. Sachs, G. Fettweis, 5G-enabled Tactile
460–473. US
Internet, IEEE Journal on Selected Areas in Communications 34 (3) (2016)
22
ACCEPTED MANUSCRIPT
T
IP
CR
US
AN
M
ED
PT
CE
AC
ACCEPTED MANUSCRIPT
Olga Galinina is a Research Scientist in the Department of Electronics and Communications Engineering
at Tampere University of Technology, Finland. She received her B.Sc. and M.Sc. degrees in Applied
Mathematics from Department of Applied Mathematics, Faculty of Mechanics and Physics, St.
Petersburg State Polytechnical University, Russia as well as the Ph.D. degree from Tampere University of
Technology. Her research interests include applied mathematics and statistics, queueing theory and its
applications; wireless networking and energy efficient systems, machine-to-machine and device-to-
device communication.
T
IP
Sergey Andreev is a Senior Research Scientist in the Department of Electronics and Communications
Engineering at Tampere University of Technology, Finland. He received the Specialist degree (2006) and
CR
the Cand.Sc. degree (2009) both from St. Petersburg State University of Aerospace Instrumentation, St.
Petersburg, Russia, as well as the Ph.D. degree (2012) from Tampere University of Technology. Sergey
(co-)authored more than 90 published research works on wireless communications, energy efficiency,
heterogeneous networking, cooperative communications, and machine-to-machine applications.
US
Mikhail Komarov is currently Associate Professor at the Department of innovations and business in IT
AN
(School of Business Informatics, Faculty of business and management, National Research University
Higher School of Economics). He received his M.Sc. and Cand.Sc. degrees from Moscow State Institute of
Electronics and Mathematics (Technical University) in 2010 and 2012, respectively. His research
M
interests include energy efficiency of wireless sensor networks, heterogeneous networks and network
challenges with the connection to big data area, e-business and e-commerce, social web of things,
internet of things and mobile applications networks. He authored more than 30 publications.
ED
Svetlana Maltseva has many years of experience in e-commerce and e-business. Main research areas are
PT
social networks and social computing, e-commerce and e-business, Web 3.0, Smart Commerce,
innovations in IT. Dr. Maltseva is a member of the Focus Group: Bridging the Gap from innovation to
standard of the ITU, one of the organizers of the annual Big Data Science Forum (2012), co-chair of the
CE
organizing committee of the Workshops on technological and legal aspects of the internet governance
at the Internet governance forum by the UN (IGF-2012,IGF-2013,IGF-2014). Dr. Maltseva was awarded
with IBM Academic Award in 2013. Dr. Svetlana Maltseva is a programme director (MSc in Big Data
AC
Systems).