IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 38, NO.

10, OCTOBER 2019 1799

QoS-Adaptive Approximate Real-Time Computation

for Mobility-Aware IoT Lifetime Optimization
Kun Cao , Guo Xu, Junlong Zhou , Member, IEEE, Tongquan Wei , Member, IEEE,
Mingsong Chen , Member, IEEE, and Shiyan Hu , Senior Member, IEEE

Abstract—In recent years, the Internet of Things (IoT) has I. I NTRODUCTION

promoted many battery-powered emerging applications, such as
NTERNET of Things (IoT), providing ubiquitous connec-
smart home, environmental monitoring, and human healthcare
monitoring, where energy management is of particular impor-
tance. Meanwhile, there is an accelerated tendency toward mobil-
I tivity for anyone and anything at any time and any place,
has been envisioned as one of the most promising networking
ity of IoT devices, either being transported by humans or being paradigms for the information society [1]. In the past few
mobile by itself. Existing energy management mechanisms for years, IoT has promoted a variety of emerging applications,
battery-powered IoT fail to consider the two significant character-
such as intelligent transportation, smart home, environmen-
istics of IoT: 1) the approximate real-time computation and 2) the
mobility of IoT devices, resulting in unnecessary energy waste and tal monitoring, etc. According to a report from McKinsey
network lifetime decay. In this paper, we explore mobility-aware Global Institute, the potential economic impact of IoT applica-
network lifetime maximization for battery-powered IoT appli- tions could be as much as $11.1 trillion per year in 2025 [2].
cations that perform approximate real-time computation under Nonetheless, various challenges in IoT remain to be addressed,
the quality-of-service (QoS) constraint. The proposed scheme one of them being the efficient energy management [3]. Due
is composed of offline and online stages. At offline stage, an
optimal mobility-aware task schedule that maximizes network to the considerations of either portability requirements, or
lifetime is derived by using mixed-integer linear programming the low-cost installation, deployment, and maintenance, many
technique. Redundant executions due to mobility-incurred over- IoT devices are powered by batteries that usually provide
lapping of a single task on different IoT devices are avoided only a limited supply of energy. Therefore, efficient energy
for energy savings. At online stage, a performance-guaranteed management is a key concern in battery-powered IoT, which
and time-efficient QoS-adaptive heuristic based on cross-entropy
method is developed to adapt task execution to the fluctuat- aims at prolonging the network lifetime while meeting the
ing QoS requirements. Extensive simulations based on synthetic quality-of-service (QoS) requirement.
applications and real-life benchmarks have been implemented to Several novel energy-efficient management mechanisms
validate the effectiveness of our proposed scheme. Experimental for battery-powered IoT have been proposed in [4]–[7].
results demonstrate that the proposed technique can achieve Colistra et al. [4] presented a consensus protocol for a dis-
up to 169.52% network lifetime improvement compared to
benchmarking solutions.
tributed decision on resource allocation in a simulated IoT sce-
nario. The proposed consensus protocol can improve network
Index Terms—Approximate real-time computation, Internet of lifetime while preserving the required QoS. Luo et al. [5] intro-
Things (IoT), mobility, network lifetime optimization, quality-of-
service (QoS). duced the concept of “equivalent node” to select relay node for
optimal data transmission and energy conservation in WSN-
based IoT. A probabilistic dissemination algorithm is designed
to choose the optimal energy strategy and prolong the lifetime
of whole network. Samie et al. [6] put forward a technique
of computation offloading in a local IoT network for network
lifetime optimization under bandwidth constraints. A hierar-
Manuscript received May 5, 2018; revised August 10, 2018; accepted
September 13, 2018. Date of publication October 1, 2018; date of current chical clustering approach was presented by Li et al. [7] to
version September 18, 2019. This work was supported in part by the Shanghai maximize the network lifetime of battery-powered IoT. This
Municipal NSF under Grant 16ZR1409000, in part by the China HGJ Project approach assigns different duty cycle ratios to the IoT devices
under Grant 2017ZX01038102-002, in part by the National NFSC under
Grant 61802185 and Grant 61872147, in part by Jiangsu NSF under Grant that are located at different layers, resulting in the balance of
BK20180470, and in part by the Fundamental Research Funds for the Central energy consumption among these devices.
Universities. This paper was recommended by Associate Editor Y. P. Liu. All the above works [4]–[7] on IoT energy management
(Corresponding author: Tongquan Wei.)
K. Cao, G. Xu, M. Chen and T. Wei are with the Shanghai Key Laboratory fail to consider two significant characteristics of IoT. One is
of Trustworthy Computing, East China Normal University, Shanghai 200062, the approximate task execution of IoT applications, and the
China (e-mail: [email protected]). other is the mobility of IoT devices. In real-world, many IoT
J. Zhou is with the School of Computer Science and Engineering, Nanjing
University of Science and Technology, Nanjing 210094, China. applications can accept an inaccurate or approximate result in
S. Hu is with the Department of Electrical and Computer Engineering, addition to a guaranteed accurate result [8]. For an approx-
Michigan Technological University, Houghton, MI 49931 USA. imate computation modeled task, its execution cycles are
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. logically decomposed into two parts: one is a mandatory part
Digital Object Identifier 10.1109/TCAD.2018.2873239 and the other is an optional part [9], [10]. The mandatory
0278-0070 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 1800 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 38, NO. 10, OCTOBER 2019

to be executed once on IoT device ψ1 in addition to being

executed once on IoT device ψ2 . This results in overlapping
task τ3 being executed twice. In fact, overlapping task τ3 only
needs to be executed once. Obviously, redundant execution of
overlapping tasks leads to unnecessary energy waste, thereby
decaying the network lifetime. Meanwhile, since network
QoS requirement may fluctuate at runtime, task execution
should quickly adapt to the fluctuating QoS requirement when
(a) optimizing network lifetime. In this paper, we propose a QoS-
adaptive mobility-aware scheme to optimize network lifetime
of battery-powered IoT applications with approximate real-
time computation requirements. To the best of our knowledge,
this is the first attempt to optimize network lifetime with
joint considerations of IoT device mobility and approximate
real-time computation. As illustrated in Fig. 2(b) and (c), our
developed solution can not only avoid redundant execution of
(b) overlapping tasks for energy saving but also meet the fluctu-
ating QoS requirement at runtime. The major contributions of
Fig. 1. Illustration to demonstrate IoT device mobility. (a) Initial distribution
of device locations. (b) Current distribution of device locations due to device this paper are summarized as follows.
mobility. 1) We investigate the problem of network lifetime
optimization for battery-powered IoT applications that
perform approximate real-time computation under the
part must execute to completion before the deadline and QoS constraint. In particular, the mobility of IoT devices
generate an acceptable result. The optional part refines and is taken into consideration.
improves the result produced by the mandatory part. Some 2) We present a mixed-integer linear programming (MILP)-
works [11]–[13] have utilized the concept of approximate com- based mobility-aware offline solution to the QoS-
putation to design task scheduling schemes for system QoS constrained network lifetime optimization, and pro-
maximization. However, these works are tailored to maximize pose a cross-entropy method-based QoS-adaptive online
system QoS for an individual IoT device, and thus they cannot heuristic that can quickly adapt task execution to the
be directly applied to the lifetime optimization of the network fluctuating QoS requirement at runtime.
consisting of multiple IoT devices. 3) We conduct extensive simulation experiments to validate
Mobility of IoT devices is another important characteristic the effectiveness of our proposed algorithms. Simulation
of IoT. There is an accelerated tendency toward mobility of IoT results demonstrate that the proposed algorithms can
devices, either being transported by humans, like smartphones, achieve up to 169.52% network lifetime improvement
or being mobile by itself, like a robot or an unmanned aerial compared to benchmarking schemes.
vehicle. Fig. 1 illustrates an example of device mobility. As The remainder of this paper is organized as follows.
shown in Fig. 1(a), tasks τ3 , τ5 , τ8 , τ10 , τ11 , and τ13 appear in Section II introduces network architecture and models.
more than one IoT device in the initial distribution of device Section III describes problem definition and the overall frame-
locations. For example, task τ3 appears not only in device work of our optimization solution. Section IV presents the
ψ1 but also in device ψ2 . In this situation, device ψ1 and offline approach based on MILP technique while Section V
device ψ2 need to process task τ3 separately because task τ3 shows the proposed online scheme based on cross-entropy
has different collected input data on different devices. Due method. The effectiveness of the proposed solution is verified
to device mobility, the initial distribution of device locations in Section VI and concluding remarks are given in Section VII.
may change, as demonstrated in Fig. 1(b). In Fig. 1(b), any
one of tasks τ3 , τ5 , τ8 , τ10 , τ11 , and τ13 has the same collected II. N ETWORK A RCHITECTURE AND M ODELS
input data on different devices. We divide these tasks into two
In this section, we briefly introduce our network architecture
categories: one for exclusive task and one for overlapping task.
and models including task model, energy model, and network
An exclusive task is the task that can only be executed by the
lifetime and QoS model.
specific IoT device, and an overlapping task refers to the task
that can be handled by any one of the IoT devices containing
this task. For instance, as shown in Fig. 1(b), task τ1 and A. Network Architecture
task τ2 are exclusive tasks whereas task τ3 is an overlapping We consider a local network composed of N IoT devices
task. This is because tasks τ1 and τ2 can only be executed by ψ = {ψ1 , ψ2 , . . . , ψN } and a gateway, as shown in Fig. 3. All
device ψ1 , while task τ3 can be executed either by device ψ1 IoT devices are connected to the gateway by using a specific
or by device ψ2 . IoT wireless technology (e.g., HaLow, LPWAN, and BLE [6]).
However, as demonstrated in Fig. 2(a), existing energy Fig. 4 presents the architecture diagram of any IoT device in
management mechanisms fail to take into consideration the the local network. As illustrated in the figure, each IoT device
mobility of IoT devices, and they allow overlapping task τ3 ψi (1 ≤ i ≤ N) is equipped with two major modules: one

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 CAO et al.: QoS-ADAPTIVE APPROXIMATE REAL-TIME COMPUTATION FOR MOBILITY-AWARE IoT LIFETIME OPTIMIZATION 1801

(a) (b) (c)

Fig. 2. Illustration of existing mechanisms and our proposed solution to deal with the mobility of IoT devices shown in Fig. 1. Overlapping tasks are marked
with red rectangles, while exclusive tasks are marked with rectangles of other colors that are consistent with the colors used in Fig. 1. (a) Existing energy
management mechanisms perform redundant execution of overlapping tasks τ3 , τ5 , τ8 , τ10 , τ11 , and τ13 . It is clear that these mechanisms lead to unnecessary
energy waste. (b) Our proposed MILP-based offline method avoids redundant execution of overlapping tasks τ3 , τ5 , τ8 , τ10 , τ11 , and τ13 . Any overlapping
task is only executed once for energy saving. (c) According to the fluctuating QoS requirement at runtime, our proposed online method quickly reallocates
overlapping task τ3 from IoT device ψ1 to ψ2 for achieving further energy saving.

Fig. 4. Architecture diagram of any IoT device.

Fig. 3. Local network with N IoT devices and a gateway. Otherwise task τm,i is an exclusive task and thus ci [m][j] is
set to 0. Tasks are assumed to be heterogeneous and therefore
the activity factor of a task, denoted by μ (ranging in (0,1]),
is an energy source module and the other is an energy dissi- is introduced to capture how intensively functional units are
pation module. The energy source module is typically in the being utilized by the task [15].
form of a battery (e.g., Li-ion). The energy dissipation mod- In this paper, we consider approximate computation real-
ule drains energy from the energy source module and consists time tasks [9], [10]. Every task τm,i is logically decomposed
of three parts: 1) sensors; 2) a heterogeneous multiprocessor into two parts: 1) a mandatory part with execution cycles
system-on-chip (MPSoC); and 3) a signal transceiver. The sen- Mm,i , which must execute to completion before the dead-
sors sense physical phenomena and acquire input data required line and generate an acceptable result and 2) an optional
for tasks. The MPSoC system performs task execution, and part with maximum execution cycles Om,i , which refines and
the signal transceiver implements communication between the improves the result produced by the mandatory part.1 We use
IoT device and the gateway. Assume that the MPSoC system a tuple τm,i : {μm,i , Rm,i , Vm,i , Mm,i , Om,i , αm,i } to character-
on IoT device ψi is equipped with Li heterogeneous proces- ize the approximate computation modeled task τm,i . In the
sors, denoted by i = {i,1 , i,2 , . . . , i,Li }, where every tuple, μm,i is the activity factor of task τm,i . Rm,i represents
processor i,l (1 ≤ l ≤ Li ) is characterized by a given supply the amount of input data needed for task τm,i to start its execu-
voltage/frequency pair (vi,l , fi,l ). tion. Vm,i denotes the amount of data acquired by IoT device
ψi for task τm,i in a sampling cycle. The αm,i (ranging in [0,1])
B. Application Model is the optional execution factor of task τm,i , which denotes the
proportion of executed optional cycles to maximum optional
Assume that real-time periodic bag-of-tasks applica- cycles of task τm,i . Therefore, the actual length Wm,i of task
tions [14] are to be executed in the local network. In such τm,i , measured by the total execution cycles, can be given by
a periodic application, tasks are independent, activated with
a common period, and share a common deadline. Tasks are Wm,i = Mm,i + αm,i × Om,i . (1)
independent in the sense that there is no precedence or com-
munication among tasks. Let T and D represent the period C. Energy Model
and deadline of the application, respectively. Suppose that ξi
The overall power consumption Pcon of any IoT device in
tasks of the application can be performed by IoT device ψi ,
the local network includes the power consumption Psen of
denoted by i = {τ1,i , τ2,i , . . . , τm,i , . . . , τξi ,i }. Let Ci be a
sensors, the power consumption Pexe of the MPSoC system,
ξi × N matrix, where ci [m][j] ∈ Ci (1 ≤ j ≤ N, j = i) gives
and the power consumption Pcom of the signal transceiver.
the information whether task τm,i ∈ i (1 ≤ m ≤ ξi ) is an
overlapping task or not. If τm,i ∈ j holds, then task τm,i is 1 For the sake of generality, we assume the weight of task τ
m,i that indicates
an overlapping task and hence ci [m][j] takes the value of 1. the relative importance of the task has been integrated into Om,i .

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 1802 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 38, NO. 10, OCTOBER 2019

Therefore, Pcon is expressed as

Pcon = Psen + Pexe + Pcom . (2)

The power consumption Pexe of an MPSoC system can be

modeled as the sum of static power consumption Psta and
dynamic power consumption Pdyn , that is, [16]

Pexe = Psta + Pdyn . (3)

Psta is independent of switching activity and can be regarded

as a processor-dependent constant while Pdyn is related to
charging and discharging of gates in the circuits. The dynamic Fig. 5. Overview of our proposed solution.
power consumption of processor i,l when executing task τm,i
at the supply voltage/frequency pair (vi,l , fi,l ) is given by [16]
of optional parts of all the approximate computation real-time
Pdyn (m, i, l) = μm,i Ci,l
eff 2
vi,l fi,l (4) tasks in the network. It is denoted by Q and is expressed as
eff is the effective switching capacitance of processor 
N  ξi
N 
where Ci,l
i,l . Therefore, the overall energy consumption of IoT device Q= Qi = αm,i Om,i (9)
ψi during scheduling horizon T is hence calculated as i=1 i=1 m=1

where Qi is the QoS achieved by IoT device ψi , αm,i is the

energy

Li
optional execution factor of task τm,i , and Om,i is the maximum
Econ (i) = Pcom (i) × T + Psta (i, l) × T
optional cycles of task τm,i .
l=1

Li   Rm,i

+ μm,i Ci,l
eff 2
vi,l Wm,i + Psen (i) III. P ROBLEM D EFINITION AND OVERALL F RAMEWORK
Vm,i
l=1 τm,i ∈i,l In this section, we first give a definition of the network
(5) lifetime optimization problem, and then show the overall
framework of our proposed solution.
where Psen (i) and Pcom (i) are the power consumption of sen-
sors and the signal transceiver on IoT device ψi , respectively. A. Problem Definition
Psta (i, l) is the static power consumption of processor i,l .
i,l ⊆ i is the task subset consisting of these tasks that are In this paper, we concentrate on the scenario where the
executed on processor i,l . available energy supply of each IoT device during the schedul-
ing horizon is sufficient to complete the execution of the
mandatory parts of all approximate tasks on the IoT device,
D. Network Lifetime and QoS Model but sufficient or insufficient to perform the execution of the
The lifetime of an IoT device can be defined as the ratio of optional parts of all approximate tasks on the IoT device.
its remaining energy to its power consumption [6]. Therefore, The studied problem is formally defined as follows. Given
the lifetime Ti of IoT device ψi is given by a mobility-aware local network, the target application, and
energy the network QoS requirement, design a task allocation and
Erem (i)
Ti = energy (6) scheduling scheme to maximize network lifetime under the
Econ (i)/T constraints that: 1) the mandatory parts of tasks must be com-
energy pleted before the deadline; 2) the network QoS requirement
where Erem (i) is the remaining energy of IoT device ψi and
is satisfied; and 3) for each IoT device, its energy consump-
it is expressed as the difference between the available energy
energy tion cannot exceed its available energy supply during the
supply Esup (i) provided by the energy source module and
energy scheduling horizon.
energy consumption Econ (i)
energy energy energy
Erem (i) = Esup (i) − Econ (i). (7) B. Overview of Our Proposed Solution
Fig. 5 presents an overview of our proposed solution to
The network lifetime, denoted by T, can be thus represented
QoS-constrained network lifetime optimization. As shown in
as the minimal lifetime of the IoT device in the local network
the figure, our solution is composed of two stages: 1) a
T = min{T1 , T2 , . . . , Ti , . . . , TN }. (8) mobility-aware offline stage and 2) a QoS-adaptive online
stage. At offline stage, an optimal mobility-aware task sched-
It has been shown that the QoS of a task can be repre- ule that can maximize network lifetime while satisfying
sented as a linear or concave function of optional cycles of all design constraints is derived by using MILP technique.
the task [9]. The more cycles the optional part of the task exe- Redundant executions due to mobility-incurred overlapping of
cutes, the higher QoS the task generates. Thus, the network a single task on different IoT devices are avoided for energy
QoS can be defined as the sum of the executed CPU cycles savings. Since the network QoS requirement may fluctuate

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 CAO et al.: QoS-ADAPTIVE APPROXIMATE REAL-TIME COMPUTATION FOR MOBILITY-AWARE IoT LIFETIME OPTIMIZATION 1803

at runtime, the offline task schedule needs to be adjusted at 4) For each IoT device, its energy consumption cannot
runtime at the minimum cost. Therefore, we provide a low- exceed its available energy supply during the scheduling
cost yet high-performance QoS-adaptive online scheme. In the horizon. This constraint is equivalent to the condition
proposed online scheme, if the gateway monitors no change in that the lifetime of any IoT device should not be less
network QoS requirement2 at runtime during the current time than zero during the scheduling horizon
interval [t, t + T] for executing new arriving task instances, 
the optimal task schedule produced at offline stage will be Ti Am,i,l , Sm , αm ≥ 0, ∀τm ∈ , i ∈ Mψm , l ∈ M
i
.
directly adopted without any modification to schedule tasks; (17)
else an online heuristic will be invoked. The main idea of
5) The offline QoS requirement Qoff
goal should be met
the heuristic is to redetermine the optimal processors for the

execution of overlapping tasks by using cross-entropy method αm × Om ≥ Qoffgoal . (18)
such that the network lifetime is optimized and task execution τm ∈
quickly adapts to the fluctuating QoS requirement.
6) Tasks have no overlapping executions in the same pro-
cessor. Let Sp and Sq denote the start time of task τp
IV. MILP-BASED M OBILITY-AWARE O FFLINE
and task τq when the two tasks are executed on proces-
N ETWORK L IFETIME O PTIMIZATION
sor i,l , respectively, and the variable S min(p,q) denote
In this section, we present an MILP-based mobility-aware the minimum of the Sp and Sq . Thus, the equation
offline approach to solve the QoS-constrained network lifetime S min(p,q) = min(Sp , Sq ) holds. The bp,q is an auxil-
optimization problem. iary binary decision variable indicating the relationship
of S min(p,q) , Sp , and Sq . If Sp < Sq holds, that is,
A. MILP Formulation S min(p,q) = Sp , then bp,q = 1; else bp,q = 0. H is
For ease of presentation, let  = {1 ∪· · ·∪i ∪· · ·∪N } be a large enough constant number and is set to 10 000
the task set consisting of all exclusive tasks and overlapping in the experiment. Similar to S min(p,q) , the variable
tasks on N IoT devices. An overlapping task on different IoT S max(p,q) is introduced to indicate the maximum of
devices only appears once in task set . Let Mψm be the set of the Sp and Sq . That is, S max(p,q) = max(Sp , Sq ) holds.
the indexes of these IoT devices that can execute task τm ∈ . Two auxiliary variables hp,q and gp,q are also introduced
Let M i be the set of the indexes of these processors on IoT as pseudo-linear constraints to facilitate the formula-
device ψi . We define the following variables: tion. The following constraints must be satisfied to avoid
⎧ overlapping executions:
⎨ 1 if task τm ∈  is assigned to
Am,i,l = processor i,l , i ∈ Mψm , l ∈ M
i
(10) S min ≤ Sp (19)
⎩ (p,q)
0 otherwise S min ≤ Sq (20)
Sm : the start time of task τm (11) (p,q)
S min ≥ Sp − H × (1 − bp,q ) (21)
αm : the optional execution factor of task τm . (12) (p,q)
S min ≥ Sq − H × bp,q (22)
1) Objective: The objective function is expressed as (p,q)
 bp,q = 1, 0 (23)
maximize : T Am,i,l , Sm , αm , ∀τm ∈ , i ∈ Mψm , l ∈ M
i
.
S max = Sp + Sq − S min (24)
(13) (p,q) (p,q)
 
2) Constraints: The constraints are summarized below. hp,q = H × Sp − S max + Sq (25)
(p,q)
1) Every task τm is assigned to exactly one processor
  Mq + αq Oq Rq
Sp − hp,q ≥ + (26)
Am,i,l = 1, ∀τm ∈ . (14) fi,l Vq,i
m
i∈Mψ i  
l∈M
gp,q = H × Sq − S max + Sp (27)
2) The optional execution factor αm of task τm takes the (p,q)
value from interval [0, 1] Mp + αp Op Rp
Sq − gp,q ≥ + . (28)
fi,l Vp,i
αm ∈ [0, 1], ∀τm ∈ . (15)
3) The mandatory part of every task τm meets its deadline B. Algorithm of MILP-Based Offline Approach
    Algorithm 1, performed by the gateway, demonstrates the
Rm Mm
Sm + Am,i,l + ≤ D, ∀τm ∈ . pseudo-code of our proposed MILP-based offline lifetime
m i
Vm,i fi,l optimization scheme. Inputs to the algorithm are the task set,
i∈Mψ l∈M

device set, and offline QoS requirement. The optimal offline
(16)
task scheduling that can maximize network lifetime under the
2 We assume that the time granularity of changes in QoS requirements at network QoS and real-time constraints is derived by solving
runtime cannot be smaller than the application period. the MILP formulated in Section IV-A.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 1804 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 38, NO. 10, OCTOBER 2019

Algorithm 1: Mobility-Aware MILP Offline Scheme X = (X1 , . . . , Xz , . . . , XZ ) is a sample vector consisting of

Input: 1) Task set , 2) Device set ψ, Z random samples produced by the corresponding proba-
off bility P(x, u). P(x, u) is in parametric class of probabilities
3) Offline QoS requirement Qgoal ;
Output: Offline schedule table off ; {P(x, ι), ι ∈ } with parameter ι being set to u. γ is a threshold
1 Generate schedule table 
off for N IoT devices by or level parameter, and Pu (F(X) ≥ γ ) represents the proba-
solving the MILP formulated in Section IV-A; bility of F(X) ≥ γ . Eu ( {F(X)≥γ } ) denotes the expected value
2 Return Schedule table  .
off of {F(X)≥γ } , and {F(X)≥γ } is the indicator function, that is,
1 F(x) ≥ γ
{F(x)≥γ } = (31)
0 otherwise.

V. C ROSS -E NTROPY M ETHOD -BASED Q O S-A DAPTIVE The cross-entropy method aims to find the minimal γ such
O NLINE L IFETIME O PTIMIZATION that λ(γ ) (i.e., the probability of F(X) ≥ γ ) approaches 0. It is
then that the probability of F(X) < γ approaches 1, indicating
The proposed MILP-based offline lifetime optimization
γ is the minimal upper bound on F(x) for ∀x ∈ 0 and thus
scheme generates an optimal task schedule that maximizes
the optimal solution to (29). The main steps of cross-entropy
network lifetime while satisfying all design constraints.
method are summarized and described as follows.
However, due to the fact that the network QoS requirement
1) Set iteration counter t ← 1 and initialize probability
may fluctuate at runtime, the offline task schedule needs to
vector P0 .
be adjusted at runtime. Therefore, we provide a low-cost yet
2) Generate Z samples {X1 , . . . , Xz , . . . , XZ } based on prob-
high-performance QoS-adaptive online task scheduling algo-
ability vector Pt−1 , and calculate sample performance
rithm based on cross-entropy method. Our online scheme is
{F(X1 ), . . . , F(Xz ), . . . , F(XZ )}.
designed to redetermine the optimal processors for overlap-
3) Select Belite samples with best performance, and let
ping task execution such that not only is the network lifetime
ϑ denote the set of indices of best samples. Derive
optimized but also task execution can quickly adapts to the
threshold γt using the average performance of Belite best
fluctuating QoS requirement. In this section, we first briefly
samples
introduce the theoretic basis of cross-entropy method, and then
describe our proposed online scheme based on cross-entropy 1 
γt ← elite F(X ). (32)
method. B ∈ϑ
4) Obtain probability vector Pt using
A. Theoretical Basis of Cross-Entropy Method Z
z=1 {F(Xz )≥γt } {xz,a =b}
The basic idea of cross-entropy method is to transform Pt,a,b = Z
z=1 {F(Xz )≥γt }
the deterministic optimization problem into its corresponding
stochastic optimization problem [17]. By using an iterative a = 1, 2, . . . , n; b = 1, 2, . . . , n∗ (33)
sampling algorithm, the stochastic optimization problem is where xz,a is the ath element in sample Xz , and Pt,a,b
then addressed. In each iteration of the sampling algorithm, indicates the probability of xz,a being mapped to b (i.e.,
multiple random samples representing solutions to the deter- xz,a = b) at tth iteration.
ministic optimization problem are generated, and these random 5) If predefined stop criterion is met, exit; otherwise, set
samples converge to the optimal or near-optimal solution in t ← t + 1, and return to 2).
a probabilistic way. Compared to conventional optimization
techniques, such as genetic algorithm [18] or particle swarm B. Our Online Scheme Based on Cross-Entropy Method
optimization algorithm [19], cross-entropy method has a com-
The objective of online scheduling includes not only gener-
plete theoretical basis, and has been proven to be an effective
ating a high quality task schedule but also minimizing the
solution to optimization problems [17]. For more details on
runtime scheduling overheads. To this end, the main idea
the theoretical basis of cross-entropy method, readers are
of our QoS-adaptive online scheme is to reallocate over-
suggested to refer to literature [17].
lapping tasks to optimal processors by using cross-entropy
Now, let us consider a discrete or combinatorial
method such that the network lifetime is optimized while
optimization problem that the goal is to find the optimal map-
∗ task execution can be quickly adapted to the fluctuating QoS
ping β ∗ : Rn → Rn such that the value of function F(x) about
requirement.
variable x in the state space 0 is maximized when x = β ∗ ,
Algorithm 2, executed by the gateway, shows the pseudo-
that is,
code of our proposed cross-entropy method-based online
 scheme. Line 1 of the algorithm initializes the online task
F β ∗ = γ ∗ = max F(x). (29)
x∈0 schedule on to the offline task schedule off . If no change in
network QoS requirement at runtime is detected, the optimal
This optimization problem is associated with the following task schedule produced at offline stage will be directly adopted
estimation of the probability [17]: without any modification (lines 2 and 3); else an online
 heuristic based on cross-entropy method will be invoked
λ(γ ) = Pu (F(X) ≥ γ ) = Eu {F(X)≥γ } . (30) (lines 4–24). Lines 5 and 6 of the algorithm construct a task set

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 CAO et al.: QoS-ADAPTIVE APPROXIMATE REAL-TIME COMPUTATION FOR MOBILITY-AWARE IoT LIFETIME OPTIMIZATION 1805

Algorithm 2: QoS-Adaptive Online Scheme according to current probability vector and then select feasi-
Input: 1) Task set , 2) Device set ψ, 3) Online QoS ble samples from the generated samples. If no feasible sample
is found and the current iteration counter is no less than 2,
goal , 4) Offline schedule table 
requirement Qon off

Output: Online schedule table  ; on lines 18 and 19 terminate the iterative process. Lines 20 and 21
1 Initialize 
on ← off ; derive the performance of feasible samples and accordingly
off
2 if Qgoal == Qgoal then
on update the threshold. Lines 22 and 23 update probability vec-
tor and iteration counter used for next iteration. Line 24 returns
3 Return  . on
the online schedule table generated based on the sample with
4 else best performance at the (t − 1)th iteration.
5 Construct task set  that stores all overlapping tasks;
6 Create processor set  where each element m ∈  VI. E VALUATION
indicates the set of these processors that overlapping
A. Experimental Settings
task τm ∈  can be assigned to;
7 for each IoT device ψi do Two sets of simulations have been implemented to val-
8 Delete all overlapping tasks from oni ;
idate the effectiveness of our proposed network lifetime
9 Update start times of all exclusive tasks in on
i ;
optimization solution. The first set of simulations is based on
10 Derive its available energy supply using Eq. (7); synthetic applications while the second set of simulations is
off based on real-life benchmarks. In each set of simulations, two
11 Calculate QoS difference Q = Qon goal − Qgoal ; local networks that consist of, respectively, 5 (N = 5) and
12 Assign Q equally to overlapping tasks in task set 10 (N = 10) IoT devices are adopted. In addition to compar-
 while satisfying Eq. (15); ing the performance achieved by our proposed offline method
13 Set t = 1, and initialize maximal iteration number with that of three benchmarking algorithms RAN, CTF [22],
tmax ; and HWG [23], we also compare the performance achieved
14 Initialize probability vector P0 : for each overlapping by the proposed online approach with that of three bench-
task τm ∈  , set the probability of allocating τm to marking algorithms GEN, PSO, and GCS [12]. The mentioned
the processors in m to 1/sizeof (m ) and the benchmarking algorithms are described below.
processors not in m to 0; 1) RAN is an algorithm that randomly selects tasks whose
15 while t ≤ tmax do optional parts are to be completed for the purpose of
16 Generate J samples according to Pt−1 using Latin meeting the network QoS requirement.
hypercube importance sampling [20]; 2) CTF [22] is a method that assigns QoS-critical jobs
17 Select Z feasible samples meeting Eq. (16) and higher priorities such that their optional cycles can be
Eq. (17) from J samples using completed first. The QoS-critical jobs are defined as
acceptance-rejection method [21]; tasks with larger maximum optional cycles.
18 if find no feasible sample and t ≥ 2 then 3) HWG [23] is a state-of-the-art approach that integrates a
19 break; worst-fit-based partitioning heuristic with genetic algo-
20 Calculate network lifetime for each feasible rithm to generate a task allocation that reduces energy
sample using Eq. (8); consumption while satisfying all design constraints.
21 Obtain threshold γt using Eq. (32); 4) GEN is a method that replaces cross-entropy method
22 Derive probability vector Pt using Eq. (33); (lines 13–24 of Algorithm 2) with genetic algorithm [18]
23 t ← t + 1; to perform network lifetime optimization.
5) PSO, similar to GEN, is an approach that uses the par-
24 Return on that is the feasible sample with optimal
ticle swarm optimization algorithm [19] to replace the
network lifetime at (t − 1)th iteration.
cross-entropy method adopted in Algorithm 2.
6) GCS [12] is a dynamic scheduling method that deter-
mines the best allocation of slack cycles for maximizing
and a processor set. For each IoT device, lines 7–9 first delete QoS. For fair comparison, we stop the running of GCS
all overlapping tasks from its schedule table and then updates when the network QoS requirement is satisfied.
the start times of all exclusive tasks in the schedule table. All the algorithms above are implemented in C++, and the
The available energy supply that can be used for overlapping simulations are performed on a machine with Intel i7 Dual-
task execution for each IoT device is calculated in line 10. Core 3.5 GHz processor, and 16 GB memory. We perform
Lines 11 and 12 first derive network QoS difference and 1000 experiments to obtain the average of the simulation data.
then equally assign the network QoS difference to overlapping
tasks. Lines 13–23 perform the reallocation process for over- B. Simulation for Synthetic Applications
lapping tasks by using cross-entropy method. Lines 13 and 14 In this set of simulations, a set of synthetic MPSoC
initialize the iteration counter, maximal iteration number, and systems is adopted to execute synthetic applications. The sup-
probability vector. Line 15 determines whether the algorithm ply voltage v, operating frequency f , and effective switching
continues the iteration process or exits the optimization. In capacitance Ceff of ten processors built on 65 nm technology
each iteration, lines 16 and 17 first generate multiple samples are listed in Table I. The number of processors of an MPSoC

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 1806 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 38, NO. 10, OCTOBER 2019

(a) (b) (c)

Fig. 6. Network lifetime when running 10 synthetic applications under N = 5 and varying offline QoS requirements. (a) Qoff off
goal = 0.5. (b) Qgoal = 0.6.
(c) Qoff
goal = 0.7.

(a) (b) (c)

Fig. 7. Network lifetime when running 10 synthetic applications under N = 10 and varying offline QoS requirements. (a) Qoff off
goal = 0.5. (b) Qgoal = 0.6.
(c) Qoff
goal = 0.7.

TABLE I
PARAMETERS OF THE S IMULATED P ROCESSOR M ODEL [24] sensors on IoT device ψi to acquire the input data required for
task τm,i is arbitrarily chosen from the interval [1, 5] Mb/s [31].
The power consumption of sensors and that of the signal
transceiver on each IoT device are indiscriminately picked
from the ranges [0.5, 2] W and [1, 5] W, respectively, [32].
The maximum iteration number tmax is set to 10.
1) Performance Comparison for Offline Approach: We uti-
lize a normalized network QoS requirement in the comparative
study, the maximum of which is set to 1. To be specific, the
network QoS requirement is normalized against the sum of
the maximum optional cycles of all tasks on N IoT devices.
Fig. 6 presents the normalized network lifetime under varying
offline QoS requirements when the local network consists of
system is randomly selected in the range [2, 10]. Task activity five IoT devices executing synthetic applications. The results
factors are uniformly distributed in the interval [0.4, 1] [15]. given in the figure clearly show that our proposed offline
The worst case execution cycles (WCECs) of tasks are scheme achieves better performance in terms of network life-
assumed to be in the range of [4 × 109 , 6 × 1010 ] [25]–[27]. time compared to the three benchmarking algorithms. For
Each task τm,i is instantiated by haphazardly picking two example, in the case of Qoff goal = 0.5, the network lifetime
WCECs from the range. One is for its mandatory part achieved by the proposed approach is 96.25%, 41.86%, and
Mm,i and the other is for its maximum optional part Om,i . 9.76% higher on average than that of RAN, CTF, and HWG,
The number of tasks on individual IoT device is randomly respectively. In the case of Qoffgoal = 0.6, the network lifetime
selected from the range [10, 30]. The amount of input data achieved by the proposed approach is 94.57%, 38.18%, and
needed for each task is randomly picked from the interval 9.24% higher on average than that of RAN, CTF, and HWG,
[1, 20] Mb [28]. The information matrix Ci is also generated respectively. In the case of Qoff goal = 0.7, the network life-
in a random way. The common deadline D of tasks is assumed time achieved by the proposed approach is 92.8%, 36.6%, and
to be 1.5 × max{ τm,i ∈i Mm,i /fi,max | i = 1, 2, . . . , N}, where 8.18% higher on average than that of RAN, CTF, and HWG,
fi,max is the maximum frequency supported by IoT device respectively.
ψi [24], [29]. The application period T is set to be the same Fig. 7 plots the normalized network lifetime under varying
as the common deadline D. A set of Li-lion battery banks offline QoS requirements when the network consists of ten IoT
equipped with [4, 10] Ah normal capacity and [15, 20] V devices executing synthetic applications. Similar to the results
terminal voltage [30] is served as energy supply modules. For shown in Fig. 6, our proposed offline algorithm significantly
the battery bank on each IoT device, its normal capacity and improves the network lifetime. For instance, in the scenario
terminal voltage are generated at random from corresponding goal = 0.5, the network lifetime achieved by the proposed
of Qoff
intervals. The amount of data per sampling cycle sensed by approach is 122.22%, 59.02%, and 27.14% higher on average

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 CAO et al.: QoS-ADAPTIVE APPROXIMATE REAL-TIME COMPUTATION FOR MOBILITY-AWARE IoT LIFETIME OPTIMIZATION 1807

(a) (b) (c)

Fig. 8. Network lifetime when running 10 synthetic applications under N = 5 and varying online QoS requirements. (a) Qon on
goal = 0.55. (b) Qgoal = 0.65.
(c) Qon
goal = 0.75.

(a) (b) (c)

Fig. 9. Network lifetime when running 10 synthetic applications under N = 10 and varying online QoS requirements. (a) Qon on
goal = 0.55. (b) Qgoal = 0.65.
on
(c) Qgoal = 0.75.

TABLE II
than that of RAN, CTF, and HWG, respectively. In the scenario RUNTIME (U NIT: s) ON AVERAGE FOR D ERIVING THE S CHEDULING
goal = 0.6, the network lifetime achieved by the proposed
of Qoff OF 10 S YNTHETIC A PPLICATIONS U NDER N = 5
approach is 115.47%, 55.5%, and 23.05% higher on average
than that of RAN, CTF, and HWG, respectively. In the scenario
goal = 0.7, the network lifetime achieved by the proposed
of Qoff
approach is 111.08%, 50.21%, and 20.60% higher on average
than that of RAN, CTF, and HWG, respectively.
2) Performance Comparison for Online Approach: Fig. 8 TABLE III
depicts the normalized network lifetime under varying online RUNTIME (U NIT: s) ON AVERAGE FOR D ERIVING THE S CHEDULING
OF 10 S YNTHETIC A PPLICATIONS U NDER N = 10
QoS requirements when the network consists of five IoT
devices. As demonstrated in the figure, our proposed online
scheme achieves remarkable network lifetime improvement
compared to the three benchmarking algorithms. To be spe-
goal = 0.55, the network lifetime achieved
cific, in the case of Qon
by the proposed approach is 21.54%, 13.14%, and 11.74%
higher on average than that of GCS, GEN, and PSO, respec-
tively. In the case of Qon goal = 0.65, the network lifetime scenario of Qongoal = 0.75, the network lifetime achieved by the
achieved by the proposed approach is 19.58%, 10.87%, and proposed approach is 26.99%, 10.49%, and 12.85% higher on
10.42% higher on average than that of GCS, GEN, and PSO, average than that of GCS, GEN, and PSO, respectively.
respectively. In the case of Qon goal = 0.75, the network lifetime Table II demonstrates the runtime on average for deriving
achieved by the proposed approach is 17.13%, 9.55%, and the scheduling of synthetic applications using the proposed
9.90% higher on average than that of GCS, GEN, and PSO, online scheme and benchmarking schemes when the network
respectively. consists of five IoT devices. The metric Sp denotes the speedup
Fig. 9 plots the normalized network lifetime under vary- achieved by the proposed method when compared to the base-
ing online QoS requirements when the network consists of line approach in terms of average runtime. It can be easily
ten IoT devices. As shown in the figure, our proposed online seen from the table that our proposed online scheme greatly
algorithm achieves better lifetime improvement compared to reduces the runtime for deriving task scheduling. For exam-
the three benchmarking algorithms. Specifically, in the sce- ple, when online QoS requirement Qon goal is set to 0.55, our
nario of Qon goal = 0.55, the network lifetime achieved by the proposed online scheme can achieve up to 13.44 times of
proposed approach is 30.39%, 12.72%, and 16.28% higher on speedup. Table III presents the runtime on average for deriv-
average than that of GCS, GEN, and PSO, respectively. In the ing the scheduling of synthetic applications using the proposed
scenario of Qon goal = 0.65, the network lifetime achieved by the online scheme and benchmarking schemes when the network
proposed approach is 28.89%, 11.88%, and 15.06% higher on consists of ten IoT devices. Similar to the results presented
average than that of GCS, GEN, and PSO, respectively. In the in Table II, our proposed online scheme also achieves better

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 1808 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 38, NO. 10, OCTOBER 2019

(a) (b) (c)

Fig. 10. Network lifetime when running 10 real-life applications under N = 5 and varying offline QoS requirements. (a) Qoff off
goal = 0.5. (b) Qgoal = 0.6.
(c) Qoff
goal = 0.7.

(a) (b) (c)

Fig. 11. Network lifetime when running 10 real-life applications under N = 10 and varying offline QoS requirements. (a) Qoff off
goal = 0.5. (b) Qgoal = 0.6.
(c) Qoff
goal = 0.7.

(a) (b) (c)

Fig. 12. Network lifetime when running 10 real-life applications under N = 5 and varying online QoS requirements. (a) Qon on
goal = 0.55. (b) Qgoal = 0.65.
on
(c) Qgoal = 0.75.

performance in terms of runtime compared to the three bench- higher network lifetime compared to the three benchmark-
marking schemes. For instance, when online QoS requirement ing algorithms. For example, in the case of Qoff goal = 0.5,
Qon
goal is set to 0.55, our proposed online scheme can achieve the network lifetime achieved by the proposed approach is
up to 18.76 times of speedup. 155.62%, 70.30%, and 25.18% higher on average than that of
RAN, CTF, and HWG, respectively. Fig. 11 plots the normal-
ized network lifetime under varying offline QoS requirements
C. Simulation for Real-Life Applications when the network consists of ten IoT devices executing real-
In this set of simulations, a set of practical MPSoC systems life tasks. Similar to the results shown in Fig. 10, our proposed
that are constructed on Intel Core Duo processor, Intel Xeon offline algorithm significantly improves the network lifetime.
processor, AMD Athlon processor, TI DSP processor, and 2) Performance Comparison for Online Approach: Fig. 12
SPARC64 processor is adopted for simulation. All the proces- compares the normalized network lifetime under varying
sor parameters of these MPSoC systems can be found in [14]. online QoS requirements when the network consists of five
The tool MEGA [33] that incorporates approximate computa- IoT devices executing real-life tasks. As demonstrated in the
tion is utilized to generate real-life benchmarks. The settings figure, our proposed online scheme achieves higher network
of task number, sensors, signal transceiver, and Li-lion battery lifetime compared to the three benchmarking algorithms. Take
the case of Qon goal = 0.55 as an example, the network life-
bank on each IoT device are the same as that of simulation
for synthetic applications. time achieved by the proposed approach is 19.75%, 10.87%,
1) Performance Comparison for Offline Approach: Fig. 10 and 6.94% higher on average than that of GCS, GEN, and
demonstrates the normalized network lifetime under varying PSO, respectively. Fig. 13 plots the normalized network life-
offline QoS requirements when the local network consists of time under varying online QoS requirements when the network
five IoT devices executing real-life tasks. The results given in consists of ten IoT devices executing real-life tasks. As shown
the figure clearly present that our proposed scheme achieves in the figure, our proposed online algorithm achieves striking

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 CAO et al.: QoS-ADAPTIVE APPROXIMATE REAL-TIME COMPUTATION FOR MOBILITY-AWARE IoT LIFETIME OPTIMIZATION 1809

(a) (b) (c)

Fig. 13. Network lifetime when running 10 real-life applications under N = 10 and varying online QoS requirements. (a) Qon on
goal = 0.55. (b) Qgoal = 0.65.
(c) Qon
goal = 0.75.

TABLE IV
RUNTIME (U NIT: s) ON AVERAGE FOR D ERIVING THE S CHEDULING results reveal that our proposed solution achieves remarkable
OF 10 R EAL -L IFE A PPLICATIONS U NDER N = 5 network lifetime improvement.

R EFERENCES
[1] P. Semasinghe, S. Maghsudi, and E. Hossain, “Game theoretic mecha-
nisms for resource management in massive wireless IoT systems,” IEEE
Commun. Mag., vol. 55, no. 2, pp. 121–127, Feb. 2017.
TABLE V [2] J. Manyika, The Internet of Things: Mapping the Value Beyond the Hype,
RUNTIME (U NIT: s) ON AVERAGE FOR D ERIVING THE S CHEDULING McKinsey Glob. Inst., New York, NY, USA, 2015.
OF 10 R EAL -L IFE A PPLICATIONS U NDER N = 10 [3] J. Henkel, S. Pagani, H. Amrouch, L. Bauer, and F. Samie, “Ultra-
low power and dependability for IoT devices (Invited paper for IoT
technologies),” in Proc. IEEE Design Autom. Test Europe Conf. Exhibit.,
2017, pp. 954–959.
[4] G. Colistra, V. Pilloni, and L. Atzori, “Objects that agree on task
frequency in the IoT: A lifetime-oriented consensus based approach,”
in Proc. IEEE World Forum Internet Things, 2014, pp. 383–387.
[5] J. Luo, D. Wu, C. Pan, and J. Zha, “Optimal energy strategy for node
selection and data relay in WSN-based IoT,” Mobile Netw. Appl., vol. 20,
no. 2, pp. 169–180, 2015.
lifetime improvement compared to the three benchmarking [6] F. Samie et al., “Computation offloading and resource allocation for low-
algorithms. For example, in the scenario of Qon goal = 0.55,
power IoT edge devices,” in Proc. IEEE World Forum Internet Things,
2016, pp. 7–12.
the network lifetime achieved by the proposed approach is
[7] Q. Li, S. P. Gochhayat, M. Conti, and F. Liu, “EnergIoT: A solution to
24.06%, 16.49%, and 11.13% higher on average than that of improve network lifetime of IoT devices,” Pervasive Mobile Comput.,
GCS, GEN, and PSO, respectively. vol. 42, pp. 124–133, Dec. 2017.
Table IV lists the runtime on average for deriving the [8] S. Mittal, “A survey of techniques for approximate computing,” ACM
Comput. Surveys, vol. 48, no. 4, p. 62, 2016.
scheduling of real-life tasks when the network consists of [9] J. W. S. Liu et al., Algorithms for Scheduling Imprecise Computations,
five IoT devices. It can be easily seen from the table that Springer, 1991, pp. 203–249.
our proposed online scheme greatly reduces the runtime for [10] M. Amirijoo, J. Hansson, and S. H. Son, “Specification and management
of QoS in real-time databases supporting imprecise computations,” IEEE
producing task scheduling. For example, when online QoS Trans. Comput., vol. 55, no. 3, pp. 304–319, Mar. 2006.
requirement Qongoal is set to 0.55, our proposed online scheme [11] L. A. Cortes, P. Eles, and Z. Peng, “Quasi-static assignment of voltages
can achieve up to 13.51 times of speedup. Table V presents and optional cycles in imprecise-computation systems with energy con-
siderations,” IEEE Trans. Very Large Scale Integr. (VLSI) Syst., vol. 14,
the runtime on average for deriving the scheduling of real-life no. 10, pp. 1117–1129, Oct. 2006.
tasks when the network consists of ten IoT devices. Similar to [12] H. Yu, B. Veeravalli, and Y. Ha, “Dynamic scheduling of imprecise-
the results presented in Table IV, our proposed online scheme computation tasks in maximizing QoS under energy constraints for
embedded systems,” in Proc. IEEE Asia South Pac. Design Autom. Conf.,
also achieves better performance in terms of runtime compared 2008, pp. 452–455.
to the three benchmarking schemes. [13] T. Wei et al., “Cost-constrained QoS optimization for approxi-
mate computation real-time tasks in heterogeneous MPSoCs,” IEEE
Trans. Comput.-Aided Design Integr. Circuits Syst., vol. 37, no. 9,
pp. 1733–1746, Sep. 2018.
VII. C ONCLUSION [14] K. Li, X. Tang, and K. Li, “Energy-efficient stochastic task scheduling
on heterogeneous computing systems,” IEEE Trans. Parallel Distrib.
In this paper, we tackle the problem of QoS-constrained Syst., vol. 25, no. 11, pp. 2867–2876, Nov. 2014.
network lifetime optimization for approximate computation [15] H. Huang, V. Chaturvedi, G. Quan, J. Fan, and M. Qiu, “Throughput
real-time tasks in battery-powered IoT. Particularly, the mobil- maximization for periodic real-time systems under the maximal temper-
ity of IoT devices is taken into consideration to avoid redun- ature constraint,” ACM Trans. Embedded Comput. Syst., vol. 13, no. 2s,
p. 70, 2014.
dant executions of the same task on different IoT devices [16] M. A. Haque, H. Aydin, and D. Zhu, “On reliability management of
for energy reduction. Our proposed solution consists of a energy-aware real-time systems through task replication,” IEEE Trans.
mobility-aware offline scheme based on MILP technique and Parallel Distrib. Syst., vol. 28, no. 3, pp. 813–825, Mar. 2017.
[17] L. Deng, The Cross-Entropy Method: A Unified Approach to
a QoS-adaptive online scheme based on cross-entropy method. Combinatorial Optimization, Monte-Carlo Simulation, and Machine
Extensive simulations are conducted, and the experimental Learning, Taylor & Francis, 2006.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.
 1810 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 38, NO. 10, OCTOBER 2019

[18] M. Qiu, Z. Ming, J. Li, K. Gai, and Z. Zong, “Phase-change memory Junlong Zhou (S’15–M’17) received the Ph.D.
optimization for green cloud with genetic algorithm,” IEEE Trans. degree in computer science from East China Normal
Comput., vol. 64, no. 12, pp. 3528–3540, Dec. 2015. University, Shanghai, China, in 2017.
[19] J. Kennedy, “Particle swarm optimization,” in Encyclopedia of Machine He was a Visiting Scholar with the University of
Learning. New York, NY, USA: Springer, 2011, pp. 760–766. Notre Dame, Notre Dame, IN, USA, from 2014 to
[20] R. L. Iman, Latin Hypercube Sampling. Chichester, U.K.: Wiley, 2008. 2015. He is currently an Assistant Professor with
[21] S. Lucidl and M. Piccioni, “Random tunneling by means of acceptance- the School of Computer Science and Engineering,
rejection sampling for global optimization,” J. Optim. Theory Appl., Nanjing University of Science and Technology,
vol. 62, no. 2, pp. 255–277, 1989. Nanjing, China. His current research interests
[22] H. Kooti, D. Dang, D. Mishra, and E. Bozorgzadeh, “Energy budget include real-time embedded systems, cloud comput-
management for energy harvesting embedded systems,” in Proc. IEEE ing, and cyber physical systems.
Int. Conf. Embedded Real Time Comput. Syst. Appl., 2012, pp. 320–329. Dr. Zhou has been an Associate Editor of the Journal of Circuits, Systems,
[23] S. Saha, Y. Lu, and J. S. Deogun, “Thermal-constrained energy- and Computers since 2017.
aware partitioning for heterogeneous multi-core multiprocessor real-time
systems,” in Proc. IEEE Int. Conf. Embedded Real Time Comput. Syst.
Appl., 2012, pp. 41–50.
[24] J. Zhou et al., “Thermal-aware task scheduling for energy minimization
in heterogeneous real-time MPSoC systems,” IEEE Trans. Comput.-
Aided Design Integr. Circuits Syst., vol. 35, no. 8, pp. 1269–1282,
Aug. 2016.
[25] R. Jayaseelan and T. Mitra, “Temperature aware task sequencing and
voltage scaling,” in Proc. IEEE/ACM Int. Conf. Comput.-Aided Design, Tongquan Wei (M’11) received the Ph.D. degree in
2008, pp. 618–623. electrical engineering from Michigan Technological
[26] K. Cao et al., “Affinity-driven modeling and scheduling for makespan University, Houghton, MI, USA, in 2009.
optimization in heterogeneous multiprocessor systems,” IEEE Trans. He is currently an Associate Professor with the
Comput.-Aided Design Integr. Circuits Syst., to be published. Department of Computer Science and Technology,
[27] K. Cao, J. Zhou, M. Yin, T. Wei, and M. Chen, “Static thermal-aware East China Normal University, Shanghai, China.
task assignment and scheduling for makespan minimization in hetero- His current research interests include Internet of
geneous real-time MPSoCs,” in Proc. IEEE Int. Symp. Syst. Softw. Rel., Things, edge computing, cloud computing, and
2016, pp. 111–118. design automation of intelligent and CPS systems.
[28] Y. Cheng, L. Zhang, Y. Han, and X. Li, “Thermal-constrained task allo- Dr. Wei has been serving as a Regional Editor
cation for interconnect energy reduction in 3-D homogeneous MPSoCs,” for the Journal of Circuits, Systems, and Computers
IEEE Trans. Very Large Scale Integr. (VLSI) Syst., vol. 21, no. 2, since 2012.
pp. 239–249, Feb. 2013.
[29] J. Zhou et al., “Reliability and temperature constrained task scheduling
for makespan minimization on heterogeneous multi-core platforms,” J.
Syst. Softw., vol. 133, pp. 1–16, Nov. 2017.
[30] Y. Wang, X. Lin, Q. Xie, N. Chang, and M. Pedram, “Minimizing state-
of-health degradation in hybrid electrical energy storage systems with
arbitrary source and load profiles,” in Proc. IEEE Design Autom. Test
Europe Conf. Exhibit., 2014, pp. 1–4. Mingsong Chen (S’08–M’11) received the B.S. and
[31] B. Martinez, M. Monton, I. Vilajosana, and J. D. Prades, “The power of M.E. degrees from the Department of Computer
models: Modeling power consumption for IoT devices,” IEEE Sensors Science and Technology, Nanjing University,
J., vol. 15, no. 10, pp. 5777–5789, Oct. 2015. Nanjing, China, in 2003 and 2006, respectively,
[32] C. J. Deepu, C.-H. Heng, and Y. Lian, “A hybrid data compression and the Ph.D. degree in computer engineering from
scheme for power reduction in wireless sensors for IoT,” IEEE Trans. the University of Florida, Gainesville, FL, USA,
Biomed. Circuits Syst., vol. 11, no. 2, pp. 245–254, Apr. 2017. in 2010.
[33] Y. G. Tirat-Gefen, D. C. Silva, and A. C. Parker, “Incorporating He is currently a Full Professor with the
imprecise computation into system-level design of application-specific Department of Embedded Software and Systems,
heterogeneous multiprocessors,” in Proc. ACM Design Autom. Conf., East China Normal University, Shanghai, China. His
1997, pp. 58–63. current research interests include design automation
of cyber-physical systems, formal verification techniques, and mobile cloud
computing.
Kun Cao is currently pursuing the Ph.D. degree
with the Department of Computer Science and
Technology, East China Normal University,
Shanghai, China.
His current research interests include high
performance computing, heterogeneous multipro-
cessor systems, and cyber physical systems.
Dr. Cao was a recipient of the Reviewer Shiyan Hu (SM’10) received the Ph.D. degree in
Award from the Journal of Circuits, Systems, and computer engineering from Texas A&M University,
Computers in 2016. College Station, TX, USA, in 2008.
He is an Associate Professor with Michigan
Technological University, Houghton, MI, USA,
where he is the Director of the Center for Cyber-
Guo Xu is currently pursuing the master’s
Physical Systems and the Associate Director of
degree with the Department of Computer Science
the Institute of Computer and Cyber systems.
and Technology, East China Normal University,
He has been a Visiting Professor with IBM
Shanghai, China.
Research, Austin, TX, USA, since 2010, and
His current research interest includes power man-
a Visiting Associate Professor with Stanford
agement in mobile devices.
University, Stanford, CA, USA, from 2015 to 2016. he has published over 100
refereed papers. His current research interests include cyber-physical systems,
cybersecurity, computer-aided design of very large scale integration circuits,
and embedded systems.
Dr. Hu is a fellow of IET.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on August 27,2021 at 06:47:58 UTC from IEEE Xplore. Restrictions apply.