Energy-Efficient Resource Allocation for Multi-User

Mobile Edge Computing

Junfeng Guo, Zhaozhe Song, Ying Cui Zhi Liu Yusheng Ji
Shanghai Jiao Tong University, China Shizuoka University, Japan National Institute of Informatics, Japan

Abstract— Designing mobile edge computing (MEC) systems the drastic demands of applications with crucial computation
by jointly optimizing communication and computation resources, and latency requirements, finite battery lifetime and limited
which can help increase mobile batteries’ lifetime and improve communication and computation resources pose challenges for
quality of experience for computation-intensive and latency-
sensitive applications, has received significant interest. In this designing future energy-efficient MEC systems [1].
paper, we consider energy-efficient resource allocation schemes Designing energy-efficient MEC systems requires the joint
for a multi-user mobile edge computing system with inelastic allocation of communication and computation resources a-
computation tasks and non-negligible task execution durations. mong distributed mobiles and MEC servers. Emerging re-
First, we establish a mathematical model to characterize the search toward this direction considers optimal resource al-
offloading of a computation task from a mobile to the base
station (BS) equipped with MEC servers. This computation- location for various types of multi-task MEC systems [2]–
offloading model consists of three stages, i.e., task uploading, [10]. For example, [2]–[4] consider a single-user MEC system
task executing, and computation result downloading, and allows with one BS and multiple elastic tasks, and minimize the
parallel transmissions and executions for different tasks. Then, execution delay of all tasks under the transmission power
we formulate the weighted sum energy consumption minimization constraints. References [5]–[8], [10] study a multi-user MEC
problem to optimally allocate the task operation sequence, the
uploading and downloading time durations as well as the starting system with one BS and one inelastic task for each user,
times for uploading, executing and downloading, which is a and minimize the energy consumption under a hard deadline
challenging mixed discrete-continuous optimization problem and constraint for each task. In [9], the authors investigate a multi-
is NP-hard in general. We propose a method to obtain an optimal user MEC system with one BS and multiple independent
solution and develop a low-complexity algorithm to obtain a sub- elastic tasks for each user, and consider the minimization
optimal solution, by connecting the optimization problem to a
three-stage flow-shop scheduling problem and utilizing Johnson’s of the overall cost of energy, computation, and delay for all
algorithm as well as convex optimization techniques. Finally, users. In particular, the offloading scheduling [2], [3], [6]–[9],
numerical results show that the proposed sub-optimal solution task operation sequence selection [4], [6], storage resource
outperforms existing comparison schemes. allocation [10], and transmission time (or power) allocation
Index Terms— Mobile edge computing, computation offload- [2]–[10] are considered in the optimizations.
ing, resource allocation, optimization, flow-shop scheduling.
Note that [2]–[6], [8] assume that the size of the computa-
tion result of each task is negligible, and fail to take account of
I. I NTRODUCTION
the resource consumption for the BS to transmit computation
With the support of on-device cameras and embedded sen- results to mobiles. In addition, [3], [8], [10] assume that task
sors, new applications with advanced features, e.g., navigation, execution durations are negligible. Although [2], [4]–[7], [9]
augmented reality, interactive online gaming and multi-media consider non-negligible task execution durations, they ignore
transformation like Speech2Text, are emerging. These appli- the fact that the execution of one task can be conducted
cations are both computation-intensive and latency-sensitive. during the transmission of another task (i.e, transmissions and
Mobile edge computing (MEC) is a promising technology executions for different tasks can be conducted in parallel).
providing an IT service environment and cloud-computing Thus, [2], [3], [5], [7]–[10] do not consider the optimization
capabilities at the edge of the mobile network, within the of the task operation sequence, which significantly affects the
Radio Access Network and in close proximity to mobile users opportunities for parallel processing and hence leads to larger
to improve quality of experience. In an MEC system, compu- delay under power constraints or larger energy consumption
tation tasks of mobile users are uploaded to the base station under deadline constraints. Therefore, the obtained solutions
(BS) and executed at the attached MEC servers. Then, the in [2]–[10] are not suitable for the applications involving
computation results are transmitted back to the mobiles. With computation-intensive and latency-sensitive tasks with com-
putation results of large sizes, such as augmented reality,
Junfeng Guo is also with Science and Technology on Communication
Networks Laboratory. The work of Y. Cui was supported by NSFC grant interactive online gaming and multi-media transformation.
61401272 and grant 61521062, National 863 project 2015AA01A710 as In this paper, we shall address the above issues. We consider
well as Science and Technology on Communication Networks Laboratory energy-efficient resource allocation schemes for a multi-user
Foundation. The work of Z. Liu was supported by JSPS Kakenhi Grant
Numbers 16H02817 and 15K21599. The work of Y.Ji was supported in part MEC system with one BS of computing capability and mul-
by JSPS Kakenhi Project JP16H02817. tiple users, with inelastic computation tasks of non-negligible

978-1-5090-5019-2/17/$31.00 ©2017 IEEE

task execution durations. We adopt a more comprehensive task
model, specifying each task using three parameters, i.e., the
size of the task before execution, workload and size of the
computation result. Based on this task model, we first establish
a mathematical model to characterize the offloading of a task
from a mobile to the BS. This task-offloading model consists
of three stages, i.e., task uploading, task executing and com-
putation result downloading, and allows parallel transmissions (a) Multi-user MEC system. (b) Three stages.
and executions for different tasks. Then, we formulate the
weighted sum energy consumption minimization problem to Fig. 1: System model.
optimally allocate the task operation sequence as well as the
uploading and downloading time durations. The problem is a
challenging mixed discrete-continuous optimization problem size and hence trivial to download, which may not hold for
and is NP-hard in general. We propose a method to obtain an many applications with computation results of large sizes,
optimal solution and develop a low-complexity algorithm to such as augmented reality and multi-media transformation.
obtain a sub-optimal solution. Specifically, for given uploading The adopted task model in this paper properly addresses this
and downloading durations, we connect the problem to a three- limitation. In addition, note that the three parameters of a
stage flow-shop scheduling problem [11] and utilize Johnson’s computation task are determined by the nature of the task
algorithm to obtain an optimal task operation order and the itself, and for some computation tasks, such as html2text and
minimum total completion time under a mild condition. For a x264 video encoding, these three parameters can be estimated
given total uploading and downloading duration, we obtain to certain extent based on some prior offline measurements
the optimal uploading and downloading durations for each (please see the measurement table in [13] for an example).
task, using convex optimization techniques. Finally, numerical
results show that the proposed sub-optimal solution has low- B. Computation-Offloading Model with Non-negligible Task
complexity, close-to-optimal performance and outperforms ex- Execution Durations
isting comparison schemes. Offloading task k to the BS for executing comprises three
II. S YSTEM M ODEL sequential stages: 1) uploading task k of Lu,k bits from mobile
As illustrated in Fig. 1, we consider a multi-user MEC k to the BS; 2) executing task k at the BS (which requires Nk
system consisting of one single-antenna BS and K single- BS CPU-cycles); 3) downloading the computation result of
antenna mobiles, denoted by set K , {1, 2, ..., K}. The Ld,k bits from the BS to mobile k. Let tu,k , te,k and td,k
BS has powerful computing capability by running servers of denote the uploading, execution and downloading durations
a constant CPU-cycle frequency (in number of CPU-cycles (in seconds) in the three stages, respectively, where
per second) at the network edge. Without loss of generality, tu,k ≥ 0, k ∈ K, (1)
we assume that each mobile has one computation-intensive
Nk
and latency-sensitive (computation) task with deadline T (in te,k = , k ∈ K, (2)
seconds), which is offloaded to the BS for executing. Note F
that multiple tasks with the same deadline can be treated td,k ≥ 0, k ∈ K. (3)
as one super task, whose required computational capability Here, F denotes the fixed CPU-cycle frequency at the BS.
(workload) is the sum of the multiple tasks’ workloads.1 Note that te,k is fixed, while tu,k and td,k can be optimized.
A. Task Model Denote tu , (tu,k )k∈K , te , (te,k )k∈K and td , (td,k )k∈K .
We first adopt a more comprehensive computation task Although F is usually large, the BS has to handle multiple
model. The computation task at mobile k ∈ K, referred to computation-intensive and latency-sensitive tasks and the over-
as task k, is characterized by three parameters, i.e., the size all execution duration of all these offloading tasks may not
of the (uploaded) task before computation Lu,k > 0 (in bits), be negligible in practice. To reflect the impact of the overall
workload Nk > 0 (in number of CPU-cycles), and size of execution duration of all tasks, we capture the execution
the (downloaded) computation result Ld,k > 0 (in bits). The duration of each task in the computation-offloading model.
computation of each task k has to be accomplished within T We consider Time Division Multiple Access (TDMA) with
seconds. Time-Division Duplexing (TDD) operation for transmission,
Remark 1 (Task Model): Note that prior work [2], [3], [5] and any uploading duration and downloading duration do
typically assumes computation results to be negligible in not overlap. The execution of one task and the transmission
of another task can be conducted at the same time. These
1 We assume that all tasks have to be executed at the BS due to crucial
make the MEC system more efficient but the computation-
computation and latency requirements. The optimization results obtained in offloading model with non-negligible task execution durations
this paper can be extended to study a more general task scenario, where
some tasks can be executed locally and different tasks may have different is much more complex than the ideal one with negligible task
deadlines [12]. execution durations, as illustrated in Fig. 2(a). In the following,
 Total completion time
completion time, as illustrated in Fig. 2(b). Thus, without loss
UL / DL: tu,1 tu,2 td,2 tu,3 td,1 td,3 of generality, we consider the same sequence for uploading,
executing and downloading the K tasks, denoted by S ∈ S,
Executing: te,2 te,1 te,3 where S denotes the set of the K! different permutations of
time all tasks in K. We let subscript [k] denote the task index
(a) Arbitrary operation sequences and starting times for at position k of sequence S. From Lemma 1 in [15], we
uploading, execution and downloading operations. know that, completing the uploading operations of all K tasks
Total completion time before starting the downloading operation of any task will not
increase the total completion time, as illustrated in Fig. 2(c).
UL / DL: tu,1 tu,2 td,1 tu,3 td,2 td,3
To ensure that at any time, there are at most one task being
Executing: te,1 te,2 te,3 executed and at most one task being transmitted, we have the
following constraints:
time 
(b) Same operation sequence for uploading, execution 
su,[k] ≥ cu,[k−1]
and downloading operation. se,[k] ≥ ce,[k−1] , k = 2, 3, ..., K, (7)


Total completion time
sd,[k] ≥ cd,[k−1]
UL / DL: tu,1 tu,2 tu,3 td,1 td,2 td,3 sd,[1] ≥ cu,[K] . (8)
Executing: te,1 te,2 te,3 Remark 2 (Non-negligible Task Execution Durations):
time
Note that task execution durations are not considered in [3],
[8], and durations for downloading computation results are
(c) Completing the uploading operations of all tasks
before starting the downloading operation of any task. not captured in [2]–[6], [8], resulting in much simpler task
operation models in [2]–[6], [8]. The proposed computation-
offloading model in this paper successfully addresses these
Fig. 2: Illustration example of three operations of all tasks at the BS limitations. Specifically, under this model, the uploading or
(K = 3). For each task, the duration of each operation is represented
by the length of the corresponding rectangle. downloading of a task can be conducted in parallel with the
execution of another task, and the task operation sequence
greatly affects the total completion time (and hence the
we introduce new notations and constraints to mathematically energy consumption which will be seen in Section II-C).
specify this model. C. Energy Consumption Model
Let su,k , se,k and sd,k denote the starting times for up-
loading, executing and downloading task k, respectively. Let Similar to [5] and [16], we consider low CPU voltage at the
cu,k , ce,k and cd,k denote the completion times for uploading, BS, and model the energy consumption for executing a task
executing and downloading task k, respectively. As each of the as follows.2 At the BS, the amount of energy consumption
three stages cannot be interrupted, we first have the following for computation in a single CPU-cycle with frequency F
constraints: is µF 2 , where µ is a constant factor determined by the
 switched capacitance at the MEC servers [5]. Then, the energy

su,k + tu,k = cu,k consumption for executing task k at the BS is given by:
se,k + te,k = ce,k , k ∈ K. (4)

 Ee,k , µNk F 2 . (9)
sd,k + td,k = cd,k
We now introduce the energy consumption model for task
To ensure that the uploading, execution and downloading
uploading and downloading operations. We consider the block
operations of task k are conducted sequentially, we require:
 fading channel model. Let hk denote the channel power gain

su,k ≥ 0 for mobile k which is assumed to be constant during the T
se,k ≥ cu,k , k ∈ K. (5) seconds [2]–[9]. Let pk denote the transmission power of mo-

 bile k for uploading task k. Then, the achievable transmission
sd,k ≥ ce,k
rate (in bit/s) for uploading task k is given by:
By (4) and (5), we know that the total completion time for ( )
pk |hk |2
processing all K tasks is given by maxk∈K {cd,k } [14]. To rk = B log2 1 + , (10)
n0
guarantee the deadline constraint, we have:
where B and n0 are the bandwidth (in Hz) and the complex
max {cd,k } ≤ T. (6) additive white Gaussian channel noise, respectively. On the
k∈K
other hand, the transmission rate for uploading task k is fixed
We can have three sequences (orders) for uploading, execu-
as rk = Lu,k /tu,k , since this is the most energy-efficient
tion, and downloading of the K tasks, respectively. Following
the proof of Lemma 3 in [14], we can show that the three 2 The circuit power is omitted here for simplicity but can be accounted for
sequences can be made the same without increasing the total by adding a constant [5], [16].
 3UREOHP
(QHUJ\0LQLPL]DWLRQ

2SWLPDOVROXWLRQ (TXLYDOHQWIRUPXODWLRQ (TXLYDOHQWIRUPXODWLRQ 5HODWHGSUREOHPV 6XERSWLPDOVROXWLRQ

([KDXVWLYHVHDUFK 2SWLPDO 3UREOHP 3UREOHP 3UREOHP 2SWLPDO &ORVHGIRUPVROXWLRQ

'XUDWLRQ$OORFDWLRQ
RYHU6 VROXWLRQ 6HTXHQFH6HOHFWLRQ 'XUDWLRQ$OORFDWLRQ VROXWLRQ LQDQG
IRU*LYHQ7BX

6XESUREOHP
6WDQGDUGFRQYH[
RSWLPL]DWLRQ
2SWLPDO 3UREOHP 6XESUREOHP
VROXWLRQ 'XUDWLRQ$OORFDWLRQ
DOJRULWKPV
6XESUREOHP
&ORVHGIRUPPLQLPXP 3UREOHP 3UREOHP
2SWLPDO 3UREOHP 2SWLPDO
WRWDOFRPSOHWLRQWLPH 6HTXHQFH6HOHFWLRQDQG 7KUHH6WDJH)ORZ6KRS -RKQVRQÿVDOJRULWKP
VROXWLRQ 6WDUWLQJ7LPH$OORFDWLRQ VROXWLRQ
LQ 6WDUWLQJ7LPH$OORFDWLRQ 3UREOHP

Fig. 3: Proposed optimal and sub-optimal solutions to Problem 1.

transmission method for transmitting (Lu,k bits )in tu,k seconds duration allocation constraints. Specifically, we have the fol-
rk
(due to the fact that pk = |hnk0|2 2 B − 1 is a convex lowing optimization problem.
( x ) Problem 1 (Energy Minimization):
function of rk ) [5]. ) g(x) , n0 2 − 1 . Then, we
( Define
B

Lu,k
have pk = |h1k |2 g tu,k . Thus, the energy consumption at E∗ , min E(tu , td )
S∈S,su ,se ,sd ,tu ,td
mobile k for uploading task k to the BS is given by: s.t. (1), (3), (4), (5), (6), (7), (8),
( )
tu,k Lu,k
Eu,k (tu,k ) , pk tu,k = g . (11) where su , (su,k )k∈K , se , (se,k )k∈K and sd , (sd,k )k∈K .
|hk |2 tu,k Problem 1 is a mixed discrete-continuous optimization
Similarly, the energy consumption at the BS for transmitting problem with three types of variables, i.e., the task opera-
the computation result of task k to mobile k is given by: tion sequence (discrete variable), uploading and downloading
( ) durations (continuous variables), as well as starting times for
td,k Ld,k
Ed,k (td,k ) , g . (12) uploading, execution and downloading operations (continuous
|hk |2 td,k variables). It is NP-hard in general, which will be seen in the
Thus, the energy consumption at the BS for executing task k discussion in Section IV. By exploiting structural properties of
and transmitting the computation result of task k is given by: Problem 1, we propose an equivalent formulation, as shown in
Fig. 3. This formulation separates the three types of variables
EBS,k (td,k ) = Ee,k + Ed,k (td,k ). (13) and facilitates the optimization.
Problem 2 (Sequence Selection):
The weighted sum energy consumption corresponding to task
k is given by: E ∗ = min Eseq
∗
(S)
S
Ek (tu,k , td,k ) = Eu,k (tu,k ) + βEBS,k (td,k ), (14) s.t. S ∈ S.

where β ≥ 0 is the corresponding weight factor. Therefore, Let S∗ denote the optimal solution. Eseq
∗
(S) is given by the
the weighted sum energy consumption is given by: following sub-problem.
∑ Problem 3 (Duration Allocation for Given S): For all S ∈
E(tu , td ) = Ek (tu,k , td,k ). (15) S, we have
k∈K ∗
Eseq (S) , min E(tu , td )
Note that the weighted sum energy consumption is a convex tu ,td

function of the uploading and downloading time durations for s.t. TF (S, tu , td ) ≤ T,
all K tasks. (1), (3).
III. P ROBLEM F ORMULATION AND O PTIMAL S OLUTION Let (t∗u (S), t∗d (S)) denote the optimal solution. TF (S, tu , td )
is given by the following sub-problem.
In this section, we first formulate the energy minimization
Problem 4 (Starting Time Allocation for Given S, tu , td ):
problem for the multi-user MEC system with non-negligible
For given S ∈ S, tu and td , the minimum total completion
task execution durations. Then, we propose a method to obtain
time is given by:
an optimal solution.
We would like to minimize the weighted sum energy TF (S, tu , td ) , min cd,[K]
su ,se ,sd
consumption for the multi-user MEC with non-negligible task
execution durations under the uploading and downloading s.t. (4), (5), (7), (8).
 { {( ) ( )} }
∑
j
∑
j−1
∑
i ∑
i−1 ∑
K ∑
K
TF (S, tu , td ) = max max te,[k] − td,[k] + tu,[k] − te,[k] , tu,[k] + td,[k] . (16)
1≤i≤j≤K
k=1 k=1 k=1 k=1 k=1 k=1
{( ) ( )}
∑
j
∑
j−1
∑
i ∑
i−1 ∑
K
TeF (S, tu , td ) = max te,[k] − td,[k] + tu,[k] − te,[k] + td,[k] . (17)
1≤i≤j≤K
k=1 k=1 k=1 k=1 k=1

We can show the following result.3 Problem 6: (Sequence Selection and Starting Time Alloca-
Theorem 1: (Relationship between Problem 1 and Problem- tion): For any tu and td , we have
s 2-4): Problem 1 and Problems 2-4 are equivalent. TF∗ (tu , td ) , min cd,[K]
Based on Theorem 1, we propose a method to obtain an S∈S,su ,se ,sd
optimal solution to Problem 1, by solving Problems 2-4. This s.t. (4), (5), (7), (8).
method is illustrated in Fig. 3. First, we solve Problem 4 for In the following, we discuss Problem 6 and Problem 5, re-
given S ∈ S, tu and td . spectively, based on which we propose a low-complexity sub-
Lemma 1: (Minimum Total Completion Time): For given optimal solution, as illustrated in Fig. 3. Before interpreting
S ∈ S, tu and td , the minimum total completion time is given Problem 6, we introduce some background on M -stage flow-
by (16) (at the top of the this page). shop scheduling problems. In an M -stage flow-shop schedul-
Then, substituting TF (S, tu , td ) in (16) into Problem 3, ing problem, all tasks have to be processed on M machines
we have a convex optimization problem with 2K variables, following the same machine order. Each task requires certain
as TF (S, tu , td ) in (16) is convex (note that it is a positive fixed processing time on a machine. The objective is to find an
weighted sum of convex piecewise linear functions). This operation sequence for processing all tasks on each machine
problem can be solved using standard convex optimization so that a given criterion is optimal. The most commonly
techniques. Finally, we can solve Problem 2 by evaluating all studied criterion in literature is the total completion time.
possible choices for S ∈ S using exhaustive search. When M ≥ 3, an M -stage flow-shop scheduling problem is
NP-hard in general [14]. When M = 3, the three sequences
for processing the tasks on the three machines can be set to be
IV. S UB - OPTIMAL S OLUTION
the same without losing optimality, and the optimal sequence
Note that obtaining an optimal solution to Problem 2 can be efficiently obtained by Johnson’s algorithm in a special
requires solving Problem 3 and Problem 4 up to K! times. The case [14].
complexity is not acceptable when K is large. In this section, By treating tu , te (given in (2)) and td as the processing
we first propose an equivalent formulation for Problem 1 times for three separate machines (i.e., uploading machine,
related to a standard three-stage flow-shop scheduling problem executing machine and downloading machine), Problem 6 can
and a convex optimization problem. Then, we develop a low- be regarded as a three-stage flow-shop scheduling problem
complexity algorithm to obtain a sub-optimal solution by with an additional constraint in (8) (i.e., the uploading machine
utilizing Johnson’s algorithm [14] and convex optimization and the downloading machine cannot operate at the same
techniques. time). By relaxing the additional constraint in (8) and using
First, we propose an equivalent formulation for Problem 1, the expression of the minimum total completion time (without
i.e., Problem 5 and Problem 6, as illustrated in Fig. 3. Different the additional constraint in (8)), we can transform Problem 6
from Problems 2-4, this equivalent formulation divides the into a standard three-stage flow-shop scheduling problem [14].
three types of variables into two groups. One includes the up- Problem 7 (Three-Stage Flow-Shop Scheduling Problem):
loading and downloading durations, and the other includes the For any tu and td , we have
task operation sequence and the starting times for uploading, TeF∗ (tu , td ) , min TeF (S, tu , td ),
S∈S
execution and downloading operations.
Problem 5 (Duration Allocation): where the minimum total completion time (without the ad-
ditional constrain in (8)) TeF (S, tu , td ) is given by (17) (at
E ∗ = min E(tu , td ) the top of this page), and is obtained by optimizing su , se , sd
tu ,td under the constraints in (4), (5), (7). Let S∗ (tu , td ) denote an
s.t. TF∗ (tu , td ) ≤ T, optimal solution.
(1), (3). It can be easily verified that Problem 7 is a standard
three-stage flow-shop scheduling problem [11], [14]. We now
Let (t∗u , t∗d ) denote the optimal solution. TF∗ (tu , td ) is the establish the relationship between Problem 6 and Problem 7.
optimal value of the following sub-problem. Lemma 2: (Relationship between Problem 6 and Prob-
lem 7): Given tu and td , an optimal solution S∗ (tu , td ) to
3 We omit all the proofs due to page limitation. Please refer to [12] for Problem 7 is also optimal for Problem 6, i.e., TF∗ (tu , td )
details. = TF (S∗ (tu , td ), tu , td ).
 By Lemma 2, instead of solving Problem 6, we can focus on efficiently. In addition, it can be easily verified that, t†u,k (Tu,d )
solving Problem 7. Johnson’s algorithm [14] (of complexity and t†d,k (Tu,d ) increase with Tu,d for all k ∈ K.
O(K log K)) can guarantee to find an optimal sequence for Based on the above analysis, we can obtain a sub-optimal
the three-stage flow-shop scheduling problem in the special solution to Problem 1. The details are summarized in Algo-
case where rithm 1.
min {tu,k } ≥ max {te,k }. (18)
k∈K k∈K
Algorithm 1 : Sub-optimal Solution to Problem 5
Note that (18) usually holds, as the execution duration of each ∑
1: Set Tu,d = T − k∈K te,k .
task at the BS is usually small due to the strong computing 2: repeat
capability at the attached MEC servers. Thus, we can use 3: Solve Problem 8 by calculating (t†u (Tu,d ), t†d (Tu,d )) accord-
Johnson’s algorithm to solve Problem 7 approximately. If (18) ing to (19) and (20).
holds, the obtained solution is optimal; otherwise, it is usually 4: Solve Problem 7 using Johnson’s algorithm to ob-
a sub-optimal solution with close-to-optimal performance. tain S∗ (t†u (Tu,d ), t†d (Tu,d )) and then use (16) to obtain
Next, we focus on solving Problem 5. Note that TF (S∗ (t†u (Tu,d ),∑t†d (Tu,d )), t†u (Tu,d ), t†d (Tu,d )).
† †
TeF (S, tu , td ) in (17) is convex, implying that TeF∗ (S, tu , td ) is 5: Set Tu,d = k∈K (tu,k (Tu,d ) + td,k (Tu,d )) + (T −
∗ † † † †
convex. Thus, Problem 5 is convex and can be solved using s- TF (S (tu (Tu,d ), td (Tu,d )), tu (Tu,d ), td (Tu,d ))).
† †
6: until TF (S∗ (t†u (Tu,d ), td (Tu,d )), t†u (Tu,d ), td (Tu,d )) = T
tandard convex optimization techniques. As TeF∗ (S, tu , td ) does
not have a simple form, the complexity for solving Problem 5
may still be high. Let Tu,d denote the total uploading and Note that the time complexity for Step 3 (with λ∗ (Tu,d ) ob-
downloading duration. To reduce the complexity for solving tained using bisection search) is polynomial. The complexity
Problem 5, we propose a related problem. for Step 4 is O(K log K). ∑ We can show that if (18) holds for
Problem 8 (Duration Allocation for Given Tu,d ): For all t†u (Tu,d ) where Tu,d = T − k∈K te,k , Algorithm 1 converges.
Tu,d > 0, we have By structural properties of Problem 7 and Problem 8, we
min E(tu , td ) can show the following result.
tu ,td † †
∑ ∑ for (tu (Tu,d ), td (Tu,d ))
Lemma 3: If constraint (18) holds
s.t. (tu,k + td,k ) ≤ Tu,d , given in (19), where Tu,d = T − k∈K te,k , then E(tu , td ) de-
k∈K creases as the number of iterations increases and Algorithm 1
(1), (3). stops in a finite number of iterations.

Let (t†u (Tu,d ), t†d (T∑u,d )) denote the optimal solution. V. S IMULATION R ESULTS
∗ ∗ ∗ ∗ ∗
Let Tu,d , k∈K (t u,k + t d,k ), where (tu , td ) is the In the simulation, we consider the following settings [4],
optimal solution to Problem 5. As any uploading duration [5]. We let β = 0.1, T = 80ms, µ = 10−29 , and F = 6 × 109 .
and downloading duration do not overlap, we can easily show Channel power gain hk for mobile k is modeled as Rayleigh
that (t†u (Tu,d
∗
), t†d (Tu,d
∗
)) = (t∗u , t∗d ). That is, for optimal fading with average power loss 10−3 . The complex additive
∗
Tu,d obtained by solving Problem 5, the optimal solution to white Gaussian channel noise is n0 = 10−9 W. For each task
∗
Problem 8 with Tu,d is the same as that to Problem 5. Later, k, Lu,k and Ld,k follow the uniform distribution over [1 ×
the relationship between Problem 5 and Problem 6 will be 105 , 5 × 105 ] (bits), and Nk follows the uniform distribution
used to reduce the complexity for solving Problem 5. We can over [0.5 × 107 , 1.5 × 107 ] (CPU-cycles). All random variables
easily verify that Problem 8 is convex and Slater’s condition is are independent.
satisfied, implying that strong duality holds. Thus, Problem 8
can be solved using KKT conditions. The optimal solution A. Comparison Between Optimal and Sub-optimal Solutions
(t†u (Tu,d ), t†d (Tu,d )) to Problem 8 is given by: In this part, we use a numerical example to compare the
†

L
t (T ) = ( ( λ∗ (T u,k)|h |2 −n0 ) )
ln 2 optimal solution and the proposed sub-optimal solution in

 u,k u,d u,d k
B W n0 e +1 terms of the weighted sum energy consumption and com-
† , k ∈ K, putational complexity. From Fig. 4(a), we can see that the

 ( (
βLd,k ln 2
) )
td,k (Tu,d ) = B W λ∗ (Tu,d )|hk |2 −n0 +1 performance of the proposed sub-optimal solution is very
n0 e

(19) close to that of the optimal solution. From Fig. 4(b), we

can see that the computation time for computing the sub-
where W (·) denotes the Lambert function, e is a mathematical optimal solution grows at a much smaller rate than the optimal
constant that is the base of the natural logarithm, and λ∗ (Tu,d ) solution with respect to the number of users. This numerical
satisfies: example demonstrates the applicability and efficiency of the
∑ (L + βLd,k ) ln 2 sub-optimal solution.
( ( u,k ) ) = Tu,d . (20)
λ∗ (Tu,d )|hk |2 −n0
k∈K B W n0 e +1 B. Comparisons with Existing Schemes
∗
Note that, λ (Tu,d ) in (20) can be easily obtained using the In this part, we compare the proposed sub-optimal solution
bisection method. Thus, we can compute (t†u (Tu,d ), t†d (Tu,d )) (using Algorithm 1) with three existing comparison schemes.
 10 -2 20 10 3 10 4
Baseline 1 Baseline 1
Optimal Optimal Baseline 2 Baseline 2
10 2 10 3 Baseline 3
Sub-optimal
Energy consumption (J)

Sub-optimal Baseline 3

Energy consumption (J)

Energy consumption (J)

Computation time (s)
15 Sub-optimal
Sub-optimal
10 1 10 2

-3 10 0
10 10 10 1

-1
10 10 0
5
10 -2
10 -1

10 -4 0 10 -3
1 2 3 4 5 1 2 3 4 5 10 -2
1 4 7 10 13 16 19 0.05 0.055 0.06 0.065 0.07 0.075 0.08
Number of users K Number of users K Number of users K Time slot duration (s)

(a) K vs energy consumption. (b) K vs computation time. (a) K users at T = 80ms. (b) Deadline T at K = 10.

Fig. 4: Comparison between optimal and sub-optimal solutions. Fig. 5: The weighted sum energy consumption versus the number
of users K and the deadline T for the multi-user MEC system with
non-negligible task execution durations.
All the three baseline schemes assume that the transmission
∑Kexecution durations cannot be paralleled, and consider T −
and
stage flow-shop scheduling problem and utilizing Johnson’s
k=1 te,k as the total uploading and downloading duration.
In particular, Baseline 1 ∑ allocates the total uploading and algorithm as well as convex optimization techniques. Finally,
K numerical results show that the proposed sub-optimal solution
downloading duration T − k=1 te,k equally to the uploading
and∑downloading operations of all tasks, i.e., tu,k = tu,k = outperforms existing comparison schemes significantly.
T− K k=1 te,k
2K for all k ∈ K [5]. Baseline 2 optimally allocates R EFERENCES
the total uploading and downloading duration to uploading [1] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “Mobile
and downloading operations to minimize the weighted sum edge computing: Survey and research outlook,” CoRR, 2017. [Online].
energy consumption, using the optimal solution in (19) and Available: http://arxiv.org/abs/1701.01090
[2] Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloading
(20). Baseline 3 allocates the total uploading and downloading for mobile-edge computing with energy harvesting devices,” IEEE J.
duration to all users according to their task sizes (here we Select. Areas Commun., vol. 34, no. 12, pp. 3590–3605, Dec. 2016.
ignore the size of each computation result) to minimize the [3] J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, “Delay-optimal computation
task scheduling for mobile-edge computing systems,” in Proc. IEEE
weighted sum energy consumption, using the optimal solution ISIT, July 2016, pp. 1451–1455.
in (19) and (20), and then for each user, allocates the time [4] Y. Mao, J. Zhang, and K. B. Letaief, “Joint task offloading scheduling
duration to its uploading and downloading duartion operations and transmit power allocation for mobile-edge computing systems,” in
in Proc. IEEE WCNC, Mar. 2017, pp. 1–6.
to minimize the energy consumption corresponding to its task. [5] C. You, K. Huang, H. Chae, and B.-H. Kim, “Energy-efficient resource
Fig. 5(a) and Fig. 5(b) illustrate the weighted sum energy allocation for mobile-edge computation offloading,” IEEE Trans. Wire-
consumption versus the number of users K and the deadline T , less Commun., vol. 16, no. 3, pp. 1397–1411, Mar. 2017.
[6] Y. Yu, J. Zhang, and K. B. Letaief, “Joint subcarrier and cpu time
for the sub-optimal solution and the baseline schemes. From allocation for mobile edge computing,” in Proc. IEEE GLOBECOM,
Fig. 5(a) and Fig. 5(b), we can observe that as the number of Dec. 2016, pp. 1–6.
users K increases or the deadline T decreases, the weighted [7] J. Cheng, Y. Shi, B. Bai, and W. Chen, “Computation offloading in
cloud-ran based mobile cloud computing system,” in Proc. IEEE ICC,
sum energy consumption increases. Baseline 3 achieves the May 2016, pp. 1–6.
maximum weighted energy consumption among all the three [8] F. Wang, J. Xu, X. Wang, and S. Cui, “Joint offloading and computing
schemes, as it does not properly consider the size of the optimization in wireless powered mobile-edge computing systems,” in
Proc. IEEE ICC, May 2017, pp. 1–6.
computation result of each task. The sub-optimal solution [9] M.-H. Chen, B. Liang, and M. Dong, “Joint offloading decision and
greatly outperforms the three baselines, as it more efficiently resource allocation for multi-user multi-task mobile cloud,” in Proc.
utilizes the time duration over which the transmission and IEEE ICC, May 2016, pp. 1–6.
[10] Y. Cui, W. He, C. Ni, C. Guo, and Z. Liu, “Energy-efficient resource
execution operations are conducted in parallel (and hence allocation for cache-assisted mobile edge computing,” in Proc. IEEE
maximizes the total transmission time). LCN, Oct. 2017.
[11] H. Emmons and G. Vairaktarakis, Flow shop scheduling: theoretical
VI. C ONCLUSION results, algorithms, and applications. Springer Science & Business
Media, 2012.
In this paper, we consider energy-efficient resource al- [12] J. Guo, Z. Song, Y. Cui, Z. Liu, and Y. Ji, “Energy-efficient
location schemes for a multi-user mobile edge computing resource allocation for multi-user mobile edge computing,” CoRR,
2016. [Online]. Available: http://arxiv.org/abs/1611.01786
system with inelastic computation tasks and non-negligible [13] A. P. Miettinen and J. K. Nurminen, “Energy efficiency of mobile clients
task execution durations. We first establish a computation- in cloud computing.” HotCloud’10, vol. 10, pp. 4–4, 2010.
offloading model with non-negligible task execution durations. [14] S. M. Johnson, “Optimal two-and three-stage production schedules with
setup times included,” Naval research logistics quarterly, vol. 1, no. 1,
Then, we formulate the weighted sum energy consumption pp. 61–68, 1954.
minimization problem by optimally allocating communication [15] H. D. Mathes, “A 2-machine sequencing problem with machine repe-
and computation resources, which is NP-hard in general. tition and overlapping processing times,” OR-Spektrum, vol. 21, no. 4,
pp. 477–492, 1999.
We propose a method to obtain an optimal solution and [16] A. P. Chandrakasan, S. Sheng, and R. W. Brodersen, “Low-power cmos
develop a low-complexity algorithm to obtain a sub-optimal digital design,” IEICE Trans. Electronics, vol. 75, no. 4, pp. 371–382,
solution, by connecting the optimization problem to a three- 1992.