A Proximal Algorithm for Joint Resource

Allocation and Minimizing Carbon Footprint in

Geo-distributed Fog Computing
Cuong T. Do, Nguyen H. Tran, Chuan Pham, Md. Golam Rabiul Alam, Jae Hyeok Son and Choong Seon Hong

Department of Computer Enginneering, Kyung Hee University, 446-701, Korea

Email: {dtcuong, nguyenth, pchuan, robi, sonjaehyeok, cshong}@khu.ac.kr

Abstract—Large-scale Internet applications, such as content

distribution networks, are deployed in a geographically dis-
tributed manner and emit massive amounts of carbon footprint 'LVWULEXWHG,QWHOOLJHQFH
at the data center. To provide uniform low access latencies, )2*&RSXWLQJ &HQWUDOL]HG 'LVWULEXWHG,QWHOOLJHQFH
Cisco has introduced Fog computing as a new paradigm which ,QWHOLJHQFH&ORXG )2*&RSXWLQJ
&RPSXWLQJ
can transform the network edge into a distributed computing
infrastructure for applications. Fog nodes are geographically
distributed and the deployment size at each location reflects the
regional demand for the application. Thus, we need to control
the fraction of user traffic to data center to maximize the social
welfare. In this paper, we consider the emerging problem of joint
resource allocation and minimizing carbon footprint problem for
video streaming service in Fog computing. To solve the large-
scale optimization, we develop a distributed algorithm based
on the proximal algorithm and alternating direction method 'LVWULEXWHG,QWHOOLJHQFH 'LVWULEXWHG,QWHOOLJHQFH
)2*&RSXWLQJ )2*&RSXWLQJ
of multipliers (ADMM). The numerical results show that our
algorithm converges to near optimum within fifteen iterations,
and is insensitive to step sizes. )LHOG$UHD1HWZRUN
**/7(:L)L

I. I NTRODUCTION
(PEHGGHG6\VWHPV
Recently, Cisco has introduced Fog computing as a new DQG6HQVRUV

paradigm which can transform the network edge into a dis-

tributed computing infrastructure for applications that take
Fig. 1. Fog Computing architecture.
advantage of the billions of devices already connected to
the Internet of Things (IoT) [1], [2]. In this year, Cisco
continues on delivering its vision for the fog computing, being
a systematic, highly virtual, secure, and network-integrated aggregation can still be efficiently performed on powerful
platform that provides computing, storage, and networking resources in the core of the Cloud. Data center resources may
services between end points and traditional Cloud computing still be used with Fog computing, but they do not dominate
data centers [3]. The Fog is located below the Cloud in a over the entire picture.
widely distributed manner and serves as an optimized transfer Fog computing is one potential approach to dealing with the
medium for services and data within the Cloud. Fig. 1 presents demands of the ever-increasing number of Internet-connected
this idealized information and computing architecture and devices sometimes referred to as the Internet of Things. Many
illustrates an implementation of Fog Computing. It can be advantages of Fog computing are listed in [1] as follows. Fog
called “Fog” simply because fog is a cloud close to the can support the applications with low latency requirements
ground. Since Fog has wide geographical distribution, the (e.g. gaming, video streaming, augmented reality). Due to
Fog paradigm is well positioned for big data and real time low latency at the edge, Fog applications involve real-time
analysis and it supports mobile computing and data streaming. interactions rather than batch processing. In addition, Fog
In Fog Computing model, data, processing and applications can support mobility and location awareness, heterogeneity,
are concentrated in devices at the network edge, rather than interoperability and federation, integration with the Cloud
existing almost entirely in the Cloud, to isolate them from the and support for on-line analysis, predominance of wireless
Cloud systems and place them closer to the end-user. Putting access and dynamic per user optimization. Potential Internet of
computing resource near the edge allows Fog to perform Things (IoT) services and applications of Fog Computing are
low latency processing while latency tolerant and large scope presented in [1] such as Connected Vehicle, Smart Grid, and

978-1-4799-8342-1/15/$31.00 ©2015 IEEE 324 ICOIN 2015

Wireless Sensor and Actuator Networks. The growing ubiquity decisions can also be misaligned and lead to sub-optimal
of IoT devices is enabling a wide range of novel, large scale, equilibria. In this paper, we study the joint resource allocation
latency sensitive applications that motivates Fog computing. and minimizing carbon footprint problem for streaming service
Fog Computing will enable a totally new breed of applications in Fog computing. We assume that physical devices called
and services that will come to life as hundreds of billion of Fog computing nodes (FCNs) are placed in the network
things [4]. infrastructure to deliver video streaming service from content
Fog computing is an addition which develops the concept of providers. For example, streaming services are hosted at the
Cloud services and does not replace Cloud computing. In [2], network edge such as “smart” routers and switches with more
the authors strive to clarify key characteristics of Cloud and application-level functionality, or even end devices such as set-
Fog computing. While Fog nodes provide localization, there- top-boxes or access points. By doing so, Fog reduces service
fore enabling low latency and context awareness, the Cloud latency, and improves QoS. FCN aggregates video demands
provides global centralization. Many applications require both from nearby end users. As the tenants of a Fog provider,
Fog localization, and Cloud globalization, particularly for the content providers want to delivery content from the data
analysis and big data [1]. In [5], Madsen and Albeanu show the center as much as possible to FCN to increase their utility.
reliability challenges posed by Fog computing platforms in- Because the applications and end users are heterogeneous,
corporating networks of smart devices communicating among the utility varies significantly depending on the geographical
them and also with Cloud. The authors in [6] present a distribution of end-users. Thus, we need to control the fraction
high level programming model for future Internet applications of traffic (from huge number of end-users to data center) to
that are geographically distributed, large scale, and latency maximize the utility of content providers and minimize the
sensitive. The proposed model is a PaaS programming model, carbon footprint at the data center.
called Mobile Fog, that provides a simplified programming In a large scale systems, fast distributed resource alloca-
abstraction and supports applications dynamically scaling at tion and social welfare maximization are critical problems.
runtime. They also analyze use cases for the programming Traditional solutions to such problems rely on primal/dual
model with the camera network and connected vehicle applica- decomposition and gradient methods [20]–[23], however these
tions to show the efficiency of Mobile Fog. In [7], the authors methods have slow convergence speed and sensitivity to step
exploit the knowledge that is only available at the edge (Fog) sizes and require strict convex assumptions. Related work
nodes to improve user’s webpage rendering performance. has considered geo-distributed and large scale system such
In industry, many companies are ready for adopting Fog as work in [16] for video streaming in cloud computing. In
computing [8]. Any company that delivers content can start [16], the authors propose a distributed algorithm with fast
using Fog computing [2]. A good example is Netflix who is convergence speed but it requires strict convex assumptions
able to reach a large number of globally distributed customers. like primal/dual decomposition and gradient methods. Here,
The delivery of video-on-demand service would not be effi- the joint resource allocation and minimizing carbon footprint
cient enough if it is based on the data management in one or problem is a very large-scale convex optimization due to
two central data centers. Fog computing thus allows providing large number of FCN (dozen of thousand devices in [3]).
very large amounts of streamed data by delivering the data Thus, for reasons of performance and scalability, we introduce
directly into the vicinity of the customer [9]. a distributed solution for the joint resource allocation and
Almost all of these streaming servers are built on top of minimizing carbon footprint problem. Our algorithm is based
geographically distributed infrastructure to provide uniform on proximal algorithms [24], a powerful algorithm that re-
reliability and performance. Consequently, an application’s cently has been applied in many large scale distributed convex
environmental impact can vary significantly depending on the optimization problems. Comparing to conventional methods
geographical distribution of end-users. They need an effective such as gradient methods, proximal algorithms have faster
algorithm to direct demand across the wide area to data convergence speed with modest accuracy, insensitivity to step
center. Many previous works exist in this area such as [10]– sizes, and robustness without strong assumptions such as strict
[18]. The problem can be cast as an large-scale optimization convexity of the objective function [19], [24].
that maximizes the total utility or minimize the total cost. The remainder of this paper is organized as follows. Section
In such a large-scale system, fast distributed algorithms for II introduces the system model. In Section III, we present
resource allocation are becoming increasingly important. A joint resource allocation and minimizing carbon footprint
new algorithms for large-scale distributed resource allocation problem. Section IV shows the numerical results. We draw
with coupled objectives is presented in [16]. Inspiring by the conclusions in Section V.
Alternating Direction Method of Multipliers (ADMM) [19],
the authors have introduced an unified framework for studying
various cloud traffic management problems, ranging from ge- II. S YSTEM MODEL
ographical resource allocation to backbone traffic engineering.
Today, geographical resource allocation and energy cost In this section, we introduce our model first and then
are managed independently, leading to poor performance and formulate joint resource allocation and minimizing carbon
high costs in many cases [12]. The objectives of the two footprint problem.

325
 TABLE I
C ARBON DIOXIDE EMISSION PER KILOWATT- HOUR FOR THE MOST
'DWD COMMON FUEL TYPES

&HQWHU Nuclear Coal Gas Oil Hydro Wind

CO2 15 968 440 890 13.5 22.5
gkWh

; ; ;1 function C(·) considered in many existing works [11], [12],

[14] is given as follows
C(y) = c · r · PUE · P(y), (2)
where c denotes the carbon footprint cost in term of $/g at
)&1 )&1 ˎ )&11
the data center, r is the average carbon emission rate g/KWh,
PUE is the power usage effectiveness and P(y) represents the
server power at the data center. The authors in [11] present a
approximate estimation of the average carbon emission rate r
Fig. 2. Multiple Fog Computing Nodes with a Data Center.
of a data center. By summing the weighted contribution from
each fuel type, we can calculate the average carbon emission
A. System Model rate of the data center as follows
We consider a content provider run Fog computing services ∑j ej ·rj
r= , (3)
over a data center and N Fog Computing Nodes (FCNs) are ∑j ej
located at the edge of network in distinct geographical regions where e j is the electricity generated from fuel type j and
to serve video streaming for end users as illustrated in Fig. r j is the carbon emission rate of fuel type j which is used
2. We assume that FCNs send a request to the data center. in the data center. The carbon emission rate of the most
After the data center have finished serving request, it sends the common fuel types was reported in [11], [26] and shown
response video streaming back to FCNs. We use xn to denote in Table I. The power usage effectiveness PUE is the ratio
the amount of video streaming to FCN n from the data center. between total infrastructure power and server power. Since
We assume that the reservation of egress network bandwidth total infrastructure power mainly consists of server power and
from the data center to FCNs has already been enabled by the cooling power, PUE is commonly used as a measure of data
detailed engineering techniques proposed in [10], [25]. Thus, center energy efficiency. We assume that the requested video
the data center can guarantee the bandwidth to serve video streaming corresponds to the workload in the data center. P(y)
streaming of clients at FCNs. To model the social welfare, we represents the server power at the data center, which is a
consider both the carbon footprint cost at the data center and function of the total of requested video streaming y and can be
utility of the content provider, which are detailed below. obtained empirically. From a measurement study by Google
B. Utility and Cost Function [27], a commonly used server power function is given by
Utility, or usefulness, is the (perceived) ability of something P(y) = Ç · Pidle + (Ppeak − Pidle ) · y · κ, (4)
to satisfy needs or wants. Utility is an important concept in where κ is a conversion factor that translates requested video
economics and optimization theory, because it represents sat- streaming into workload, Ç is workload capacity of the data
isfaction experienced by the consumer of a goods (e.g, amount center, Pidle is server idle power and Ppeak peak power.
of video streaming). Some studies have applied complicated
utility functions with fairness considerations [17] or a utility III. J OINT R ESOURCE A LLOCATION AND M INIMIZING
loss for the demand from end users that are not guaranteed C ARBON F OOTPRINT P ROBLEM
[16]. However, we consider an affine utility function that is In this section we formulate joint resource allocation and
the de facto utility function widely used in the literature [18]. minimizing carbon footprint problem and introduce a new dis-
An affine utility function at FCN n has the following form: tributed optimization algorithm based on proximal algorithms.
Un (xn ) = αn xn , (1) A. Problem Formulation
We now formulate joint resource allocation and minimizing
where xn is the amount of response video streaming at FCN n
carbon footprint problem. Putting the utility and cost function,
and αn > 0 is a conversion factor that translates user-perceived
the social welfare maximization problem can be written as
request video streaming into utility (e.g., revenue). Because
N
the applications and end users are heterogeneous, αn varies
significantly depending on the geographical distribution of
max ∑ Un (xn ) −C(y)
xn ≥0,y≥0 n=1
(5)
end-users. N
For the environmental cost, the carbon footprint of energy s.t. ∑ xn = y ≤ Ç/κ,
at the data center can also be taken into account. The cost n=1

326
where the constraint is the capacity constraint at data center. where f : ℜn → ℜ ∪ {+∞} is a closed proper convex function,
If Problem (5) is small then it would be easy to solve. k is the iteration counter, and xk denotes the kth iteration of
However, in Fog, Problem (5) is an extremely large-scale the algorithm. The proximal operator proxλ f : ℜn → ℜn of f
optimization problem. For example, Akamai is the largest is defined by
content delivery network in the world. It delivers 15 − 20
1
 
percent of worldwide Internet traffic and it has more than k
proxλ f (x ) = arg min f (x) + k 2
� x − x �2 , (9)
95, 000 servers in 1, 900 networks across 71 countries [28]. x 2λ
Then, the number of variables xn is Θ(104 ). We need to update
where � · �22 is the usual Euclidean norm.
the resource allocation decision dynamically as demand varies
2) ADMM and Sharing problem: : ADMM was created in
on a hourly. Thus, we need a distributed algorithm to solve
the 1970s and has recently received renewed interest in solv-
such a large scale problem.
ing large-scale distributed convex optimization in statistics,
Problem (5) often arises in networking systems, where each
machine learning, signal processing [19]. We first consider
user n is associated with an utility Un (xn ) by using xn units of
a problem
some resource, while the users incur a total cost C(y), which
presents coupled constraint arbitrarily among users, at the min f (x) + g(z) (10)
resource provider. Conventional distributed manners to solve
Problem (5) are dual decomposition and gradient methods s.t. x − z = 0,
[19]. We can obtain the Lagrangian dual problem of Problem where f , g : ℜn → ℜ ∪ {+∞} are closed proper convex func-
(5) as follows tions. The augmented Lagrangian associated with the problem
N (10) is
min q(u) =
u
∑ max
xn
{Un (xn ) − uxn } + max{uy −C(y)},
y
(6)
n=1 Lρ (x, z, y) = f (x) + g(z) + yT (x − z) + (ρ/2) � x − z �22 , (11)
where u is the price set by the resource provider. If Problem
(5) is convex, the dual gap is zero then we only need to solve where ρ is a penalty parameter and y ∈ ℜn is a dual variable
the dual problem. We have a gradient algorithm to the dual associated with the consensus constraint. ADMM solves the
problem as follows: dual problem with the iterations:

xnk+1 := arg max Un (xn ) − uk xn , ∀n (7) xk+1 := arg min Lρ (x, zk , yk ), (12)
xn
k+1 k+1 k
z := arg min Lρ (x , z, y ),
yk+1 := arg max uk y −C(y),
y k+1 k k+1 k+1
y := y + ρ(x −z ).
N
uk+1 := uk − γ(yk+1 − ∑ xnk+1 ), Then the ADMM is presented as the proximal algorithm as
n=1
follows
where γ is a sufficiently small positive step size. The conven-
tional dual decomposition approach bases on the assumption xk+1 := proxλ f (zk − uk ), (13)
that each Un is strictly concave and monotonically increasing, z k+1
:= proxλg (x k+1 k
+ u ),
and C(y) is strictly convex. However, this assumption is not k+1 k k+1
always applied to the cloud system [16], because the joint re- u := u + x − zk+1 ,
source allocation and minimizing carbon footprint problem has
with uk = (1/ρ)yk and λ = 1/ρ.
non strictly convex cost function C(y) in (2). Furthermore, the
ADMM can be used to solve the sharing problem given in
conventional dual decomposition approach suffers from many
the form:
performance issues for solving such a large-scale problem
[16], [19]. We apply proximal algorithms [24] to make new N N
distributed optimization algorithms. We present preliminaries min ∑ fi (xi ) − g( ∑ xi ) (14)
i=1 i=1
of proximal algorithms in the following section.
Sharing can be written in ADMM form by copying all the
B. Preliminaries on Proximal Algorithm and ADMM variables:
1) Proximal Algorithm: : There is a wide range of liter- N N
ature on applying various proximal algorithms to particular min ∑ fi (xi ) − g( ∑ zi ) (15)
problems or problem domains, such as loss minimization in i=1 i=1
machine learning, optimal control, energy management, and s.t. xi − zi = 0, i = 1, ..., N,
signal processing [24]. The proximal minimization algorithm,
also called proximal iteration or the proximal point algorithm, with variable xi ∈ ℜn , i = 1, ..., N, and f : ℜn → ℜ ∪ {+∞}, g :
is ℜn → ℜ ∪ {+∞} are closed proper convex functions. Remark
that f (·) and/or g(·) are not assumed to be strictly convex as
xk+1 := proxλ f (xk ), (8) in [20] or needs some special characteristics like those in [16].

327
 The proximal algorithm associated with the problem (15) is 6
x 10
9
given by
8
xik+1 := proxλ fi (zki − uki ), ∀i (16)

Objective value
7
zk+1 := proxλg (xk+1 + uk ),
6
uk+1
i := uki + xk+1 − zk+1
i , ∀i
5

where z = [z1 , ..., zN ] ∈ ℜNn . The first and last steps can be 4
carried out independently in parallel for each i = 1, ..., N at
3
the subsystems. 0 5 10 15 20 25 30
iter (k)

C. Our Optimal Distributed Algorithm

Fig. 3. Objective value (αn is generated by uniform distribution).
We rewrite Problem (5) as follows
N N 6

max ∑ Un (xn ) −C( ∑ zn )

x 10
(17) 8

n=1 n=1 7

s.t. xn = zn , ∀n

Objective value
6

N 5

∑ zn ≤ Ç/κ. 4

n=1 3

Since Problem (17) is a sharing problem, we have the iterations 2

0 5 10 15 20 25 30
iter (k)
based on proximal algorithms as follows
Fig. 4. Objective value (αn is generated by normal distribution).
xnk+1 := proxλUn (zkn − ukn ), ∀n (18)
k+1
z := proxλC (xk+1 + uk ), is 2 x the average capacity Ç/N. The carbon emission rate at
uk+1
n := ukn + xk+1 − zk+1
n , ∀n data center r is set as much as the average in the U.S. (562
g/kWh) and the cost c is equal to 19.10−6 $/g.
Thus, we have a distributed algorithm (Algorithm 1) to solve
We evaluate the convergence of Algorithm 1 under the
the joint resource allocation and minimizing carbon footprint
above setup. Fig. 3 and Fig. 4 plot the convergence of objective
problem.
values. Since the number of iterations is small while the
Algorithm 1 Optimal Distributed Solution for (17) number of user N is 100, it suggests that our algorithm can
solve a large-scale problem effectively.
1: Choose randomly a initial value u0 , x0 , z0
Fig. 5 and Fig. 6 show the trajectory of the primal and
2: while not convergence:
dual residuals. The primary residual � r �2 is defined as
3: Each FCN n updates request video streaming xnk+1 by
∑Nn=1 � xn − zn �. The dual residual � s �2 is defined as
using x-update in (18). Send its request xnk+1 and ukn to
∑Nn=1 � ρ · (zk+1
n − zkn ) �. The residuals measure how well the
the data center.
iteration of Algorithm 1 satisfy the convergent conditions. The
4: The data center collects request video streaming xnk+1 , ukn
size of these residuals indicates how far the iterates are from a
from all FCN and updates zk+1 by using z-update in (18).
solution. The primary residual reflects how well the constraints
Then, the data center sends response video streaming xnk+1
{xn − zn } are satisfied, and is sometimes called the primal
with zk+1 to FCN n for all n.
n feasibility gap. For example, if the primal residual is 102 ,
5: Each FCN n updates dual value uk+1 by using u-update
n which is already small enough since xn is in the order of 102 .
in (18).
Hence, we conclude that the constraints are well satisfied after
6: end while
20 iterations.

IV. N UMERICAL R ESULTS V. C ONCLUSION

The numerical parameters are set as follows. The servers We study the joint resource allocation and minimizing car-
have peak power Ppeak = 200 W, and consume a half power bon footprint problem for streaming service in Fog computing.
at idle. The PUE is 1.5. These numbers represent state-of-the- We formulated the problem as a general convex optimization,
art data center hardware. αn > 0, a utility-revenue conversion where the location diversity of requested video streaming
factor, is randomly generated by uniform distribution [0,1000] utility and costs are modeled. We developed an efficient
and the normal distribution 1000 x N(0.5, 0.2), respectively. distributed algorithm based on the proximal algorithm to
The initial request video streaming xn0 is set randomly in decompose the large scale global problem into many subprob-
[50,100]. Then the capacity Ç is 1.4 x the total initial request lems, each of which can be quickly solved. The numerical
video streaming ∑Nn=1 xn0 . The upper bound of bandwidth xn0 results are conducted to evaluate the algorithm’s performance.

328
 4
10 [8] Fog computing is a new concept of data distribution. [On-
2
line]. Available: http://www.cloudtweaks.com/2013/12/fogcomputing-
10
is-a-new-concept-of-datadistribution/
||r||2

10
0 [9] Fog computing: Data, information, application and services
need to be delivered more efficiently to the end user.
10
−2
[Online]. Available: http://clouduser.de/en/analysis/fogcomputing-
0 10 20 30 40 50 60
data-information-applicationand-services-needs-to-be-deliveredmore-
10
4
efficient-to-the-enduser-22362
[10] D. Xie, N. Ding, Y. C. Hu, and R. Kompella, “The only constant
2
10 is change: incorporating time-varying network reservations in data
||s||2

0 centers,” ACM SIGCOMM Computer Communication Review, vol. 42,

10
no. 4, pp. 199–210, 2012.
10
−2 [11] P. X. Gao, A. R. Curtis, B. Wong, and S. Keshav, “It’s not easy being
0 10 20 30 40 50 60
iter (k) green,” ACM SIGCOMM Computer Communication Review, vol. 42,
no. 4, pp. 211–222, 2012.
[12] Z. Liu, M. Lin, A. Wierman, S. H. Low, and L. L. Andrew, “Greening
Fig. 5. Residual (αn is generated by uniform distribution). geographical load balancing,” in Proceedings of the ACM SIGMETRICS
joint international conference on Measurement and modeling of com-
10
4 puter systems. ACM, 2011, pp. 233–244.
[13] C. Feng, H. Xu, and B. Li, “An alternating direction method approach
10
2
to cloud traffic management.”
||r||2

0
[14] H. Xu, C. Feng, and B. Li, “Temperature aware workload management
10 in geo-distributed datacenters,” in Proceedings of the ACM SIGMET-
−2
RICS/international conference on Measurement and modeling of com-
10
0 10 20 30 40 50 60 puter systems. ACM, 2013, pp. 373–374.
4
[15] H. Xu and B. Li, “Joint request mapping and response routing for
10
geo-distributed cloud services,” in INFOCOM, 2013 Proceedings IEEE.
IEEE, 2013, pp. 854–862.
[16] D. Niu and B. Li, “An efficient distributed algorithm for resource allo-
||s||2

2
10
cation in large-scale coupled systems,” in INFOCOM, 2013 Proceedings
IEEE. IEEE, 2013, pp. 1501–1509.
10
0

0 10 20 30 40 50 60
[17] H. Xu and B. Li, “A general and practical datacenter selection framework
iter (k) for cloud services,” in Cloud Computing (CLOUD), 2012 IEEE 5th
International Conference on. IEEE, 2012, pp. 9–16.
Fig. 6. Residual (αn is generated by normal distribution). [18] S. Narayana, J. W. Jiang, J. Rexford, and M. Chiang, “To coordinate
or not to coordinate? wide-area traffic management for data centers,”
Technical report, Princeton University, Tech. Rep., 2012.
[19] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
As future work, we plan to more thoroughly study the effect optimization and statistical learning via the alternating direction method
of the multiple data centers with traffic engineering schemes. of multipliers,” Foundations and Trends R in Machine Learning, vol. 3,
no. 1, pp. 1–122, 2011.
VI. ACKNOWLEDGEMENT [20] S. Boyd, L. Xiao, and A. Mutapcic, “Subgradient methods,” lecture notes
of EE392o, Stanford University, Autumn Quarter, vol. 2004, 2003.
This work was supported by the ICT R&D program of [21] N. H. Tran, C. S. Hong, and S. Lee, “Joint Congestion Control and
MSIP/IITP. [14-000-05-001, Smart Networking Core Technol- Power Control with Outage Constraint in Wireless Multihop Networks,”
Vehicular Technology, IEEE Transactions on, vol. 61, no. 2, pp. 889–
ogy Development]. Dr. CS Hong is the corresponding author. 894, 2012.
[22] ——, “Cross-Layer Design of Congestion Control and Power Control
R EFERENCES in Fast-Fading Wireless Networks,” Parallel and Distributed Systems,
[1] F. Bonomi, R. Milito, J. Zhu, and S. Addepalli, “Fog computing and its IEEE Transactions on, vol. 24, no. 2, pp. 260–274, 2013.
role in the internet of things,” in Proceedings of the first edition of the [23] C. Do, N. Tran, Z. Han, L. Le, S. Lee, and C. S. Hong, “Optimal
MCC workshop on Mobile cloud computing. ACM, 2012, pp. 13–16. Pricing for Duopoly in Cognitive Radio Networks: Cooperate or Not
[2] M. Hajibaba and S. Gorgin, “A review on modern distributed computing Cooperate?” Wireless Communications, IEEE Transactions on, 2014, in
paradigms: Cloud computing, jungle computing and fog computing,” press.
CIT. Journal of Computing and Information Technology, vol. 22, no. 2, [24] N. Parikh and S. Boyd, “Proximal algorithms,” Foundations and Trends
pp. 69–84, 2014. in Optimization, vol. 1, no. 3, pp. 123–231, 2013.
[3] Cisco technology radar trends: Fog computing. [Online]. [25] H. Ballani, P. Costa, T. Karagiannis, and A. Rowstron, “Towards
Available: http://www.cisco.com/web/solutions/trends/tech-radar/fog- predictable datacenter networks,” in ACM SIGCOMM Computer Com-
computing.html munication Review, vol. 41, no. 4. ACM, 2011, pp. 242–253.
[4] Fog computing happens when big data analytics marries internet [26] J. V. Spadaro, L. Langlois, and B. Hamilton, “Greenhouse gas emissions
of things. [Online]. Available: http://tarrysingh.com/2014/07/fog- of electricity generation chains: Assessing the difference,” IAEA bulletin,
computing-happens-when-big-data-analytics-marries-internet-of-things/ vol. 42, no. 2, pp. 19–28, 2000.
[5] H. Madsen, B. Burtschy, G. Albeanu, and F. Popentiu-Vladicescu, [27] X. Fan, W.-D. Weber, and L. A. Barroso, “Power provisioning for a
“Reliability in the utility computing era: Towards reliable fog comput- warehouse-sized computer,” in ACM SIGARCH Computer Architecture
ing,” in Systems, Signals and Image Processing (IWSSIP), 2013 20th News, vol. 35, no. 2. ACM, 2007, pp. 13–23.
International Conference on. IEEE, 2013, pp. 43–46. [28] E. Nygren, R. K. Sitaraman, and J. Sun, “The akamai network: a
[6] K. Hong, D. Lillethun, U. Ramachandran, B. Ottenwälder, and B. Kold- platform for high-performance internet applications,” ACM SIGOPS
ehofe, “Mobile fog: a programming model for large-scale applications Operating Systems Review, vol. 44, no. 3, pp. 2–19, 2010.
on the internet of things,” in Proceedings of the second ACM SIGCOMM
workshop on Mobile cloud computing. ACM, 2013, pp. 15–20.
[7] J. Zhu, D. S. Chan, M. S. Prabhu, P. Natarajan, H. Hu, and F. Bonomi,
“Improving web sites performance using edge servers in fog computing
architecture,” in Service Oriented System Engineering (SOSE), 2013
IEEE 7th International Symposium on. IEEE, 2013, pp. 320–323.

329