This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

Virtual Network Recognition and Optimization in

SDN-enabled Cloud Environment
He Li, Kaoru Ota, Mianxiong Dong,
Department of Information and Electronic Engineering, Muroran Institute of Technology,
Muroran, Hokkaido, Japan.
E-mail: {heli, ota, mxdong}@[Link]

Abstract—Cloud computing is a scalable and efficient technology for providing different services. For better reconfigurability and
other purposes, users build virtual networks in cloud environments. Since some applications bring heavy pressure to cloud datacenter
networks, it is necessary to recognize and optimize virtual networks with different applications. In some cloud environments, cloud
providers are not allowed to monitor user private information in cloud instances. Therefore, in this paper, we present a virtual network
recognition and optimization method to improve quality-of-service (QoS) of cloud services. We first introduce a community detection
method to recognize virtual networks from the cloud datacenter network. Then, we design a scheduling strategy by combining SDN-
based network management and instance placement to improve the service-level agreements (SLA) fulfillment. Our experimental result
shows that we can achieve a recognition accuracy as high as 80% to find out the virtual networks, and the scheduling strategy increases
the number of SLA fulfilled virtual networks.

Index Terms—Software Defined Networking, Network Virtualization, Cloud Computing, Datacenter Networks

1 I NTRODUCTION without enough configuration information of virtual

networks for general-purpose applications. Usually, the
Recent years, more and more users deploy different ap- cloud provider is not allowed to access the data or other
plications in the cloud environment with better flexibility information in IaaS instances [8] to understand the virtu-
and scalability. For example, many users purchase big al network configurations. Therefore, the organization of
data cloud services such as Amazon EMR and Google virtual networks in IaaS cloud is agnostic to conventional
Clod Dataproc to build their big data applications. How- scheduling methods.
ever, if users want to use self-configurable platforms for Therefore, in this paper, we first study virtual net-
specific purposes, they have to choose infrastructure- works in the cloud environment. Usually, there are some
as-a-service (IaaS) instead of default big data services. methods to investigate instance relationships from the
For example, for graphics processing units (GPGPU) resource consumption. Software defined networking (S-
accelerated services, users usually need to build the DN) technology is an emerging approach to allow the
platform themselves and purchase GPU resources from cloud provider monitor the end-to-end network traffic
cloud providers [1], [2]. between IaaS instances [9]. Therefore, we choose net-
Many cloud computing scheduling methodologies fo- work traffic records for the virtual network recognition.
cus on virtual network optimizations [3]. One method Since a virtual network in the datacenter network is
is instance placement optimization. The cloud scheduler like a instance community in a complex network, the
adjusts the placement of each instance for differen- virtual network recognition problem can be equivalent
t purposes [4]. Another method is adjusting network to a community detection problem. Thus, we introduce
parameters of each virtual network, such as network community detection model to identify the virtual net-
priority, forwarding path and available bandwidth [5]. works from the end-to-end network traffic records.
Some previous works present combinations of different After recognizing virtual networks, we design a
methods for optimizing the cloud service performance scheduling strategy to optimize the QoS of cloud com-
[6]. In existing virtual network optimization techniques, puting. In cloud computing, for scheduling limited re-
a prerequisite is that the cloud scheduler knows the sources for instances, instance placement is an acceptable
virtual topologies. solution that places instances near to required resource.
Cloud providers usually have efficient scheduling However, instance placement will bring additional over-
tools to optimize the quality of service (QoS) of com- heads in migration and resource reallocation. Because
mercial cloud services [7]. When cloud users build of the flexible management of the SDN-enabled net-
their virtual networks in the cloud platform, the cloud work, it is possible to adjust network resources with
provider can record all configurations then optimize the negligible overhead. Therefore, we combine SDN-based
performance with their scheduling tools. However, it management and instance placement to improve the
is still hard to provide an efficient scheduling method scheduling efficiency. Furthermore, we choose the user

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

service-level agreements (SLA) fulfillment [10], a typical cloud services in a Service-as-a-Service (SaaS) environ-
QoS measurement of cloud computing, as the scheduling ment. From their work, the network management is an
object instead of other performance metrics. Finally, we important issue for improving QoS of cloud jobs. The
evaluate our work by extensive simulations and the authors also introduced the task management system
results show our method performs better than other into a hybrid environment consisting of local and remote
optimizations. clouds [12]. Moreover, Erol et al. [13] implemented a task
The main contributions of this paper are summarized management system into EU FP7 PANACEA project,
as follows. and the experimental results show that the task manage-
• We first study the virtual network recognition in ment system improved the job execution efficiency in a
SDN-enabled cloud computing environment. We real-world testbed. However, for an IaaS cloud, the cloud
propose a community detection method to recog- provider is not allowed to manage user tasks directly.
nize virtual networks from end-to-end network traf- A traditional solution is network overprovisioning
fic information recorded in the SDN controller. which brings an unacceptable cost to large scale dat-
• We propose a QoS-aware scheduling algorithm to acenter networks. Moreover, since the detailed traffic
improve the SLA fulfillment of cloud services based models are hard to be known, it is not appropriate for
on the recognized virtual networks. We combine the the cloud environment. QoS-policies-based service dif-
network management and instance placement in the ferentiation is an emerging method that segregates traffic
scheduling algorithm. for isolating performance to permit traffic engineering
• We evaluate the recognition and optimization [14]. An important solution is network virtualization
method by extensive simulations with settings and that introduces virtual networks for deployment of user
traffic trace data from real-world cloud computing customized networks [15]. Thus, in this paper, we focus
networks. We also compare our work with the other on user virtual networks in the cloud environment.
optimizations. Instance placement is a major solution for scheduling
The rest of this paper is summarized as follows. Sec- resources in the cloud environment. Meng et al. [16]
tion 2 reviews the related work. Our network scenario proposed a heuristic method to optimize the aggregate
and motivations are introduced in Section 3. Section 4 traffic rates in the cloud datacenter network by plac-
presents the recognition model. A QoS-aware scheduling ing cloud instances together with large mutual traffic.
is proposed in Section 5. Section 6 gives the simulation Jayasinghe et al. [17], Shrivastra et al. [18] and Wen
results. Finally, Section 7 concludes this paper and give et al. [19] also introduced similar approaches to mini-
the future work. mize cloud datacenter traffic. From the production traces
based experiments, these methods show a significant im-
2 R ELATED W ORK provement for the network performance. Biran et al. [20]
proposed a heuristic solution that jointly optimizes the
In this section, we first discuss the network scheduling routing between physical servers and instance placemen-
strategies in the cloud environment. Then, we introduce t, which brings better performance on network schedul-
some related SDN technologies for cloud computing. ing. Alicherry et al. [21] focus on the geo-distributed
cloud environment and place instances across different
2.1 Network Scheduling in Cloud Environment datacenter networks. Hu et al. [22] presented vBundle
There are some previous works focusing on network system focusing on instance placement in user groups to
scheduling in the cloud environment. The network re- optimize traffic between physical servers. Since instance
source is a critical resource since all physical servers placement brings additional overhead to the cloud sys-
are interconnected by datacenter networks, and com- tem, we combined network management and instance
munication overhead constrains overall performance. placement together in our scheduling strategy.
There are mainly two aspects on improving network
performance in the cloud environment. One is to design 2.2 SDN-enabled Cloud Computing
an appropriate network topology, which is determined SDN is a key network technology for service provision-
before the deployment of cloud systems. Usually, most ing of network services in the cloud environment [23].
cloud datacenters choose tree-like topologies such as fat SDN can provide flexible and efficient network resource
trees or hypercubes. These topologies focus on increasing management for supporting on-demand cloud services.
the number of network ports, which linearly increases For example, Frederic et al. [24] proposed a SDN-based
the network bandwidth. However, changing the network cognitive packet network algorithm to optimize the
topology in an existing cloud environment is difficult. routing management, which is a potential solution for
Thus, another aspect that scheduling network resource cloud datacenter network optimization. Furthermore, the
is more feasible in cloud computing. authors introduced the cognitive routing engine into the
Job scheduling is an efficient method for optimizing cloud environment to improve the QoS of the tenant
network performance by adjusting the Job arrangement network performance [25].
between different cloud servers. Wang et al. [11] pro- Cloud data center networks need several fundamental
posed a task management system to improve the QoS of requirements, including scalable deployment, dynamic

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

resource provisioning, QoS differentiation, and flexible

network control [26]. Thus, many works introduce SDN s1 s6
to meet these requirements and there are two important
issues for SDN-enabled cloud computing. The first issue s2 v1 s3 v s4 s5 v3
2
is namely virtual switching that builds communication
paths between instances in the same physical server.
Original virtual switching solutions can not provide e-
nough visibility and management on the communication
in the same physical server. OpenvSwitch, a software im-
plementation of the SDN-enabled switch, is introduced
as edge switching to provide visible and manageable
instance communication in the same physical server. p1 p2 p3 p4 p5 p6 p7 p8
Cloud providers can collect network status and traffic
information from these software SDN devices [27], [28]. Fig. 1. Virtual networks in a cloud datacenter network
Moshref et al. [29] proposed a virtual cloud rule informa-
tion base (vCRIB) to improve OpenvSwitch performance
in the cloud environment that optimizes rule placement We choose several instances to describe the scheduling
among physical and virtual switches. problem, which are denoted by s1 to s6 , and assign s1
Instance migration is another application of SDN- and s2 to virtual network v1 , s3 and s4 to v2 , and s5
enable cloud computing [30]. Since the IP protocol limits and s6 to v3 . The cloud administrator puts instance s1
the broadcast domain in datacenter networks, it is hard on physical service p1 , s2 on p7 , s3 on p5 , s4 and s5
to migrate instance across different broadcast domains. on p4 , and s6 on p8 . Thus, when users execute applica-
Address resolution protocol (ARP) also limits the in- tions in these virtual networks, the data flows between
stance addressing in the broadcast domain. SDN-enabled nearby instances in virtual networks will be forwarded
cloud computing provides new opportunity to solve this across multiple switches and routers in the data center
problem by introducing mobile IP and locator/identifier networks.
separation protocol (LISP) into instance migration [31], Since most data center networks choose tree-like
[32], [33]. OpenDaylight, the largest open source SDN topologies, the traffic across different aggregation
controller, has added a LISP-based mapping service for switches is heavy. In traditional clusters or private cloud-
support instance migration in the cloud environment s, administrators are able to adjust instance placement
[34]. After inserting forwarding rules in SDN devices, to resolve this problem with virtual network configu-
it is possible to support intra- and inter- datacenter rations. A possible adjustment is to place all instances
migration [35]. Mann et al. [36] proposed an instance from the same virtual network into a subnet under the
migration method named CrossRoads that migrates in- aggregation switch. In our example, placing instance s2
stances across cloud datacenters using pseudo address to server p1 , s4 to p5 , and s5 to p8 could solve the
mapping in the SDN controller. For maintaining in- problem.
stance connections after instance migration, CrossRoads However, the administrator is not allowed to access
changes the forwarding rules and then sets the correct the private information in instances. Furthermore, an
destination. Due to the opportunities brought by SDN appropriate scheduling strategy for different applica-
technologies, we propose the SDN-based virtual network tions is another challenge. Therefore, in the rest of this
recognition and bandwidth adjustment in this paper. paper, we will discuss a new methodology focusing
on recognizing and optimizing virtual networks in the
cloud environment.
3 BACKGROUND AND M OTIVATION
In this section, we first discuss the virtual networks in 3.2 SDN-enabled Cloud Datacenter Network
the cloud environment and then introduce the SDN-
Since we chose network traffic information for the vir-
enabled cloud datacenter network for the recognition
tual network recognition, a convenient way to monitor
and optimization methodology.
network information is using SDN. Thus, we introduce
the SDN-enabled cloud datacenter network as the main
3.1 Virtual Network in Cloud Computing scenario of our methodology.
We focus on the network virtualization in an IaaS envi- As shown in Figure 2, there is a small cloud data center
ronment. As shown in Fig. 1, there is a cloud datacenter network with a SDN controller. Cloud servers connected
network with eight physical servers denoted by p1 to p8 . by a tree topology network and all network devices are
Instances are executed on these servers and connected connected to the SDN controller. In the management
by the tree network. In user configurations, there are server, the cloud controller can migrate instances for load
three virtual networks denoted by v1 , v2 , and v3 . We balance and scheduling strategies.
assume the cloud administrator places these instances In the SDN-enabled cloud datacenter network, we
accordingly to a default scheduling strategy. assume that all network devices including virtual and

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

In the recognition problem, we consider the network

Cloud Management Cloud Controller
Server traffic in each link as the weight of an edge. We for-
SDN Controller
mulate the problem as a community identification in a
weighted network. Let G denote the entire network of
all instances, V denote the instance set, and E denote
SDN Switches
the edge set between instances, in which G = (V, E).
For two instance a, b ∈ V, we use a two-tuples group
(a, b) ∈ E to denote the edge, and wab and wba to
denote the network traffic between instance a and b,
respectively. The set of virtual networks is denoted by O
and a virtual network is denoted by oi , i ∈ [1, |O|]. For
Servers an instance a ∈ V, we use ka to denote the degree and
Instances
Na to denote
P neighbor instance set of instance a, where
ka = b∈Na (wab + wba ). We use a belonging degree
Fig. 2. SDN-enabled Cloud Datacenter Network Struc- B(a, oi ) to denote the relationship between instance a
ture and virtual network oi , given by
P
TABLE 1 (wab + wba )
B(a, oi ) = b∈oi (1)
Example of the end-to-end traffic information ka
where all neighbor instances of node a are included in
10/11/2016, 3:43 AM (UTC)
- s1 s2 s3 s4 s5 s6
virtual network oi , B(a, oi ) = 1.
s1 - 3738 0 0 0 0 Since each virtual network has heavier network traffic
s2 4670 - 0 0 0 0 between member instances, it is necessary to identify
s3 0 0 - 1570 0 0 the community which has more internal traffic. Thus,
s4 0 0 1066 - 0 0
s5 0 0 0 0 - 5312 we adopt the normalized cut metric, which is a notion
s6 0 0 0 0 6230 - in community goodness [38]. Let Φ(oi ) to denote the
normalized cut metric of virtual network oi ∈ O, given
by
physical network devices support SDN protocols. Thus, cut(oi , G\oi )
Φ(oi ) = (2)
the SDN controller can monitor end-to-end traffic in- woi
formation of all network links. The traffic information where cut(oi , G\oi ) is the traffic of the cut edges of oi
is stored in the cloud management server for further and woi is the traffic of all edges in community oi .
scheduling strategies. For protecting user privacy, we Since the identification accuracy is better with more
assume that the controller has no authority to get private internal traffic and lower normalized cut metric, we
data in network flows. Only the number of packets optimize the Φ(oi ) in our virtual network recognition.
transferred in network flows is transferred to the SDN Unfortunately, the normalized cut metric optimization is
controller. NP-hard in community detection [38]. Thus, we design
As shown in Table 1, we can get the end-to-end traf- a heuristic algorithm to identify virtual networks.
fic information from the SDN-enabled cloud datacenter As shown in Algorithm 1, the algorithm first sets O as
network. The value in each cell is the number of packets empty. Then, the algorithm begins a loop to find all vir-
transfered in end-to-end flows. There are only a few tual networks in network G. In the loop, an initial edge
packets across different virtual networks because of the (a, b) with the heaviest traffic is found and instance a
network isolation. Thus, it is possible to recognize virtual and b on this edge are put into virtual network oi . Based
networks by analyzing the end-to-end traffic informa- on the initial virtual network, the algorithm searches
tion. the set of adjacent instances denoted by Noi . Then, the
algorithm traverses all adjacent instances to extend the
4 V IRTUAL N ETWORK R ECOGNITION virtual network. If adding an adjacent instance d into
the virtual network can bring lower normalized cut
In this section, we discuss the method for recognizing metric, the algorithm adds the new instance into virtual
virtual networks in the cloud environment. network oi and updates the adjacent instances of oi .
There are some solutions for identifying virtual net- After identifying a new virtual network oi , the algorithm
works with the network traffic. Since the virtual network deletes all edges in virtual network oi and adds the
can be considered as a group of instances, the virtual net- virtual network into the result.
work recognition is similar to the community detection
problems in complex networks [37]. Thus, we formulate
the virtual network recognition as a community iden- 5 V IRTUAL N ETWORK O PTIMIZATION
tification problem and design an efficient algorithm to In this section, we describe our scheduling strategy for
solve it. optimizing QoS of the cloud system. We first statement

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

Algorithm 1 Virtual network recognition algorithm bandwidth between server mi and m0i . We also define a
1: O ← ∅; set Sj to denote all instance in virtual network vj and
2: while E 6= ∅ do sjk , k ∈ [1, |Sj |] denote an instance in set Sj .
3: (a, b) ← arg max(a,b)∈E (wab + wba ); In the instance placement, since most cloud data-
4: oi ← {a, b}; centers choose shared storage solutions for storing in-
5: Noi ← ∅; stance data, the cost of instance migration is mainly
6: for instance h ∈ oi do incurred in transferring memory data. We use a function
7: Noi ← Noi ∪ Nh l(sjk , mi , m0i ) to denote the cost for placing instance
8: end for sjk on server mi , where m0i is the original position of
9: while Noi 6= ∅ do instance sjk . We also use a value Lijk to denote the
10: c0i ← oi ∪ arg maxf ∈Noi B(f, oi ); placement strategy, given by
11: if Φ(c0i ) ≤ Φ(oi ) then (
12: oi ← c0i ; 1, if instance sjk is placed on mi ,
lijk = (3)
13: end if 0, ohterwise.
14: Noi ← ∅
Then, we study QoS measurement in the cloud envi-
15: for instance h ∈ oi do
ronment. In cloud computing, the cloud provider will
16: Noi ← Noi ∪ Nh
use SLA fulfillment to measure the QoS. Usually, SLA
17: end for
is designed for performance guarantee or resource con-
18: end while
sumption of applications. In our scenario, we choose
19: E ← E\Eoi ;
SLA as an agreement on performance guarantee of the
20: O ← O ∪ oi ;
cloud system. Thus, the SLA fulfillment is an appropriate
21: end while
way to measure the QoS of cloud applications. Actually,
the SLA fulfillment is a satisfaction ratio from each user
TABLE 2
according to the task completion time. If the task is
Notations in the virtual network optimization problem
accomplished before the time in the SLA, the fulfillment
Notation Description is 100% otherwise less than 100%.
M Set of physical servers Therefore, the QoS can be measured by the task com-
mi Server in set M pletion time. We assume the default completion time
V Set of virtual networks in the recognition result
vj Virtual network in set V is confirmed by the original configuration. We also as-
C Available bandwidth matrix of all links between sume the required bandwidth is predictable after virtual
servers network recognition. We use rjk d
to denote the required
cii0 Available bandwidth between server mi and m0i
Sj Set of instances in virtual network vj resource of instance sjk , and cjkk0 , k 0 ∈ [1, |Sj |] to denote
d

sjk Instance in set Sj the predict required bandwidth between instance sjk and
Lj Placement matrix of all instances in set Sj
lijk Relationship between instance sjk and server mi
sjk0 . The assigned resource of instance sjk is denoted by
a
d
rjk Required resource of instance sjk rjk , and assigned bandwidth between instance sjk and
cdjkk0 Required bandwidth between instance sjk and sjk0 sjk0 is denoted by cajkk0 . The resource capacity of server
a
rjk Assigned resource for instance sjk mi is denoted by Ric .
a Assigned bandwidth for the link between instance
cjkk0 Thus, the placement strategy should satisfy the con-
sjk and sjk0 straints as
Ric Resource capacity of server mi |V | |Sj |
Qj SLA fulfillment ratio of virtual network vj
X X
a
F (·) Monotone increasing function of all rjka and ca rjk · lijk ≤ Ric , and (4)
jkk0
j=1 k=1

|V | |Sj | |Sj |
X XX
the virtual network optimization problem as a non- cajkk0 lijk li0 jk0 ≤ cii0 (5)
cooperative game and then give a Pareto efficient so- j=1 k=1 k0 =1
lution.
where i 6= i0 and k 6= k 0 .
Then, we calculate the SLA fulfillment ratio of virtual
5.1 Problem Statement network vj . We use Qj to denote the SLA fulfillment,
We use M to denote the set of physical servers, and given by
mi , i ∈ [1, |M |] to denote a server in set M . Then, we  a d
100%, if ∀k ∈ [1, |Sj |], rjk ≥ rjk
use V to denote the set of all virtual networks in the 

recognition result, and vj , j ∈ [1, |V |] to denote one Qj = and ∀k 0 ∈ [1, |Sj |], cajkk0 < crjkk0
 a a
virtual network in set V . 
F (rj∗ , cj∗∗0 ), otherwise,
Since there are different topologies and settings of dat- (6)
acenter networks, we use an elastic |M |×|M | matrix C where 0 ≤ F (·) < 1 is an elastic function for measuring
a
to denote the available bandwidth between each server. SLA fulfillment ratio from rjk and cajkk0 , ∀k ∈ [1, |Sj |]
In matrix C, let cii0 , i ∈ [1, |M |], i0 ∈ [1, |M |] denote the and k ∈ [1, |Sj |].

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

Since we focus on the network resource scheduling, Algorithm 2 Instance placement optimization for virtual
we assume that the assigned resource of each instance networks
is equal to required resource. We use fjkk0 to denote the 1: sort set M and matrix C in which c12 ≥ c13 ≥ ... ≥
total traffic between instance sjk and sjk0 . Thus, in our c21 ≥ c23 ≥ ... ≥ c|C|−1|C| ;
optimization problem, we assume function F (·) can be 2: ∀i ∈ [1, |M |], j ∈ [1, |V |], k ∈ [1, |Sj |], lijk ← 0;
simplified as 3: for j ← 1 to |V | do
P|Sj | d P|Sj | d
4: Sort Sj in which ( k0 ←1 cj1k0 + k←1 cjk1 ) ≥
a
P|Sj | d P|Sj | d
F (rj∗ , caj∗∗0 ) = F (caj∗∗0 ) ( k0 ←1 cj2k0 + cjk2 ) ≥ ... ≥
P|Sj | d P|Sjk←1
| d
P|Sj | PSj fjkk0 ( k0 ←1 cj|Sj |−1k0 k←1 cjk|Sj |−1 );
k=1 k0 =1 cd 0
(7)
jkk 5: for i ← 1 to |M | do
= P|S | PS
j j fjkk0
+
P|M | P|Sj | 6: for k ← 1 to Sj do
k=1 k0 =1 ca 0 i=1 k=1 lijk
jkk 7: if Ric ≥ rjkd
and Q0j > Qj then
8: place sjk to mi ;
a
where rj∗ d
= rj∗ , and k 6= k 0 . 9: lijk ← 1
Therefore, the SLA fulfillment of virtual network vj is 10: Ric ← Ric − rjk c
;
simplified as 11: end if
( 12: end for
100%, if F (caj∗∗0 ) > 1 13: end for
Qj = (8)
F (caj∗∗0 ), otherwise. 14: for k ← 1 to |Sj |−1 do
15: for k 0 ← j + 1 to |Sj | do
We consider the SLA fulfillment as a payoff function of 16: i ← arg place sjk ;
each virtual network. We consider the virtual networks 17: i0 ← arg place sjk0 ;
as game players since each virtual network behaves in 18: if i 6= i0 then
a selfish way to maximize its utility. Thus, we formulate 19: cii0 ← cii0 − cdjkk0 ;
the optimization problem as a non-cooperative game 20: ci0 i ← ci0 i − cdjk0 k ;
between individual virtual networks, given by 21: end if
22: end for
max
Sj Qj (Lj ), ∀j ∈ [1, |V |]. 23: end for
|V | |Sj | 24: sort set M and matrix C in which c12 ≥ c13 ≥
... ≥ c21 ≥ c23 ≥ ... ≥ c|C|−1|C| ;
XX
a
s.t., rjk · lijk ≤ Ric ,
j=1 k=1 (9) 25: end for
|V | |Sj | |Sj |
X XX
cajkk0 lijk li0 jk0 ≤ cii0 ,i 6= i0 , k 6= k 0 .
j=1 k=1 k0 =1 5.2 Solving Virtual Network Optimization Problem
We design an algorithm shown in Algorithm 2 to find
In a non-cooperative game, a Pareto efficient solution a Pareto efficient solution in the virtual network op-
is considered as an efficient manner. Thus, in the QoS- timization problem. The algorithm first sorts the set
awared combined optimization, we want to find a Pareto M and matrix C in descending order by the available
efficient solution to optimize the SLA fulfillment of each network bandwidth of each link. We also set the original
virtual network. placement matrix Lj , j ∈ [1, |V |] equal 0. Then, the
We first define a Pareto efficient assignment in the algorithm starts to traverse the ordered set |V | from v1 to
virtual network optimization problem. Let vector y = v|V | . In the traversal, the algorithm sorts the instances in
(L1 , ..., L|M | ) denote the outcome of the game regarding descending order by the sum of required network band-
the assignment of instance placement solutions of all the width. Then, the algorithm places the ordered instances
virtual networks, where Lj is the set of the placement into ordered physical servers iteratively, if the resource
for network vj . capacity of the server is enough. After placement, the
Definition 1: An assignment matrix y ∗ , is Pareto effi- algorithm decreases the available bandwidth between
ciency, if there exists no other assignment matrix y such the servers in which instances are placed. After placing
that Qj (y) ≥ Qj (y ∗ ) for all j ∈ |V | and Qj (y) > Qj (y ∗ ) all instances in one virtual network, set M and matrix
for some j ∈ |V |. C are sorted again with the descending order of the
The problem of virtual network optimization prob- available network bandwidth of each link.
lem: given a set of physical servers with a connected Lemma 1: The output vector y ∗ = (L1 , L2 , ..., L|V | ) of
network and a set of virtual networks with instances, Algorithm 2 is a Pareto efficient solution.
the virtual network optimization problem attempts to Proof: We first assume a different assignment vector
assign the network bandwidth to virtual networks and y 0 = (L01 , L02 , ..., L0|V | ) such that
place instances to the physical servers such that the SLA-
fulfillment of all virtual networks satisfies Definition 1. ∀j ∈ [1, |V |] : Qj (y 0 ) ≥ Qj (y ∗ ), and (10)

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

16 16 2 2 2
∃j ∈ [1, |V |] : Qj (y 0 ) > Qj (y ∗ ). (11) 15 15 3 3 3
14 14 1 1 1
13 13 3 3 3
Let Vd denote the virtual networks having different 12 12 1 1 1

Instance number

Instance number
11 11 2 2 2
placement between y 0 and y∗. For vj ∈ Vd , we first 10 10 1 1 1
9 9 3 3 3
consider a case as 8 8 44 4
7 72 2 2
 6 6 44 4
lijk = 1, 5 5 1 1 1
4 4 3 3 3


 0 3 3 4 4 4

 lijk = 0, 2 2 4 4 4
(12) 1 1 2 2 2
 li0 jk = 0, 1 2 3 4 5 6 7 8 910111213141516 1 2 3 4 5 6 7 8 910111213141516
Instance number Instance number


 li0 jk = 1

(a) Network traffic heat (b) Result
where the lijk = 1, instance sjk is placed. Since the 0
lijk = map
1, the s0jk is also placed.
Fig. 3. Example of Virtual network recognition algorithm
Then, we consider another case as

 lijk = 0,
layers tree topology network to connect all servers.

 l0 = 1,


ijk
(13) The bandwidth between the aggregation switch and

 l i0 jk = 1, instances is set to 10 Gbps and the bandwidth between

 li0 jk = 0

the aggregation switch and core switch is set to 40 Gbps,
which comes from a typical datacenter settings. The
where both instance sjk and s0jk are placed. maximum number of user virtual networks is set to 50,
Since we assume the algorithm can place all instances, and the number of instances per each virtual network is
∀k ∈ [1, |Sj |] : ∃i ∈ [1, |M |] : lijk = 1. Thus, the third case uniform distributed from 5 to 15.
is as ( In the simulations, we choose two types of workloads
lijk = 1,
0 (14) for measuring our method, one type is the big data
lijk =0 processing which has heavy traffic and large virtual
where ∀i0 ∈ [1, |M |] : li0 jk 6= 1. networks, and another is the web service which has
In this case, because an instance is not placed in lighter traffic and small virtual networks.
vj with solution y 0 , payoff Qj (y 0 ) is less than Qj (y ∗ ). For the traffic information from big data computing
Therefore, from three cases, Qj (y 0 ) ≤ Qj (y ∗ ), which is applications, we record the traffic datasets by a virtual
inconsistent with the assumed equation (10) and (11). cluster deployed in Google Compute Engine, which has
As a result, the output y ∗ is a Pareto efficient solution. maximum 30 instances. We install an Apache Hadoop
and a Spark system in the cluster and run several
example programs with a different number of instances
to generate traffic datasets. We choose Apache Hadoop
6 P ERFORMANCE E VALUATION 2.7.3 and Spark 2.0 as the big data applications. For the
In this section, we first introduce the experiment settings traffic information from web services, we use a traffic
and then discuss the experimental results of virtual trace dataset from a real-world IaaS environment [39].
network recognition and optimization. We define an integer number for measuring the re-
sources of each server, and the resource capacity of
6.1 Simulation Settings each physical server is 100. The required resource of an
instance in each virtual network is uniform distributed
We use a workstation computer as the simulation plat-
from 3 to 30 and the required bandwidth is set to 1 Gbps
form which is equipped a CoreTM i7 4770 (8 MB Cache,
which is a common value in an IaaS platform.
up to 3.90 GHz) CPU, 16 GByte RAM, and 2 Tbyte HDD.
For different number of instances in the virtual net-
We test each simulation 20 times and record the average
work, we set the traffic from the recorded datasets with
results.
the same size. We randomly place the instances of each
We first build a small cloud datacenter network on
virtual network to virtual servers and replay the traffic
mininet 2.3.0 which is a populate emulation platform
dataset in each link.
for SDN applications. We use a tree topology with 16
instances and 5 switches and set 4 leaf switches as 4
physic servers. The bandwidth between physical servers 6.2 Recognition Accuracy
and the core switch is 40 Gbps. We implement a traffic We first test the virtual network recognition in the small
monitor model in the Floodlight 1.2 controller then cloud datacenter network with the monitor module in
the controller deploys flow tables into each aggregation the Floodlight controller. As shown in 3, we use a
switch to monitor the end-to-end traffic information. network traffic heat map to show the settings and result.
For large-scale experiments, we implement all simu- The traffic between instances is shown as a heat map in
lations with Python 2.7.12 and NetworkX 1.11. We set Fig. 3(a). The deeper color means heavier traffic while
the number of virtual servers as 100. We use a three lighter color means light traffic. We execute our virtual

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

Ratio of SLA fulfilled virtual networks

100% 80%

80%
Recognition accuracy

60%
60%
40%
40%
Original
20% 20% Average
Optimized (without placement)
Optimized
50 10 0%
4 90% 0% tions 10 20 30 40 50
Num 5 40 80 ca
ber 3 70% % a appli Number of virtual networks
of v 5 30 60 at
irtua
l ne 25 20 50% % big d Fig. 5. Ratio of SLA fulfilled virtual networks with different
40% rom
two
rks r a tio f number of virtual networks
fic
Traf
influence the recognition accuracy, which means our
Fig. 4. Accuracy of virtual network recognition
recognition method is suitable for heavy traffic scenarios.

network recognition algorithm, and the result is shown 6.3 QoS Performance
in Fig. 3(b). In the same heat map, we put a numerical The QoS performance is measured by the SLA fulfillment
digit on the traffic to mark which virtual network the ratio. In simulations, we test the ratio of the number
traffic belongs to and use 1 to 4 to mark 4 different virtual of SLA fulfilled virtual networks to the number of all
networks. The recognition algorithm works well and all virtual networks. We execute the recognition algorithm
virtual networks are recognized correctly. first and then run the optimization algorithm with the
Then, we test the recognition accuracy in large-scale recognized virtual networks. The ratio of big data traffic
simulations. We adjust the ratio of big data traffic to is set to 80%. We also compare the performance of
total traffic in simulations. Since users will deploy oth- our method to original settings and average bandwidth
er applications in cloud, different traffic records will assignment which averagely assigns bandwidth of each
influence the accuracy of the recognition algorithm. In link to instances. Meanwhile, for studying the advantage
the simulation, we set the ratio of big data computing of instance placement, we also test the performance
traffic to total traffic from 100% to 20%. We also set of our virtual network optimization solution without
the number of virtual networks from 20 to 50 and instance placement.
increase 5 instances in each step. Then, we execute the We first test the QoS performance with a different
recognition algorithm to find the virtual networks from number of virtual networks. We increase the number of
the traffic records. After recognition, we compare the virtual networks from 10 to 50 and increase 10 virtual
instances in recognized virtual networks and original networks in each step. As shown in Fig. 5, the SLA ful-
virtual networks. If the original virtual network and the filled ratio decreases with more virtual networks. When
recognized one have same instances, we consider that the number of virtual networks is set to 10, the SLA
recognized virtual network is correct. The accuracy is fulfilled ratio is more than 95% after optimization and
the ratio of the number of correct virtual networks to near to 80% by original settings. Without instance place-
the number of all virtual networks. ment, the performance of our optimization method is
From Fig. 4, our recognition algorithm recognizes near to the original settings. When the number of virtual
virtual networks with an acceptable accuracy. When the networks increases to 50, SLA fulfilled ratio with original
traffic ratio of big data computing is more than 80%, settings is less than 20%, while with our optimization
the accuracy is near to 70% with 30 virtual networks. method with instance placement, more than 80% virtual
The accuracy decreases with fewer instances and lower networks are SLA fulfilled with instance placement.
traffic ratio from big data computing. When the traffic Thus, instance placement increases the efficiency of the
ratio from big data computing decreases to 40%, the virtual network optimization obviously with more vir-
recognition accuracy decreases to less than 80% with tual networks. Average bandwidth assignment performs
50 virtual networks or 40% with 20 virtual networks. better than original settings, when the number of virtual
Thus, both the traffic type and the number of instances networks increases to 20.

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

Ratio of SLA fulfilled virtual networks

Ratio of SLA fulfilled virtual networks

80% 80%

60% 60%

40% 40%

20% 20%
Original Original
0% Average 0% Average
Optimized (without placement) Optimized (without placement)
Optimized Optimized
[5, 15] [10, 20] [15, 25] [20, 30] [25, 35] 20 30 40 50 60 70 80
Number of instances in each virtual network Bandwidth of core switches (Gbps)
Fig. 6. Ratio of SLA fulfulled virtual networks with different Fig. 7. Ratio of SLA fulfilled virtual networks with different
number of instances in each virtual network bandwidth betwee aggreation and core switches

We test the QoS performance with different number virtual networks increases to 30% with original instance
of instances in each virtual networks. We set the num- placement. The fulfilled ratio with our optimization
ber of virtual networks as 50, increase the number of without instance placement is increased to 80% because
instances in each virtual network from [5, 15] to [25, of the enough bandwidth of core switches. Therefore,
35], and increase 10 instances in each step. As shown the performance is not improved obviously with enough
in Fig.6, the SLA fulfilled ratio is decreased with more bandwidth of core switches.
instances per each virtual network. When the number of As a result, our virtual network recognition and op-
instances in each virtual network is uniform distributed timization method performs better than the average
from 5 to 15, the SLA fulfilled ratio is more than 30% assignment and the original settings. The performance
with original settings and more than 80% with our of our solution is slightly influenced by different vir-
optimization. Our optimization method without instance tual network parameters. In all simulation results, our
placement performs similarly to the one with placement, optimization method with instance placement achieves
while average assignment performs better than original a SLA fulfilled ratio of more than 80%. Meanwhile,
settings. When we uniform distribute the number of the performance with instance placement is related to
instances per each virtual network from 25 to 35, the the available network resource. Instance placement im-
SLA fulfilled ratio is less than 20% with original instance proves the number of SLA fulfilled virtual networks
placement and more than 80% after our optimization. with the scarce network resource. When the cloud envi-
The ratio of SLA fulfilled virtual networks is near to ronment provides enough network resource, bandwidth
50 % by our optimization without instance placement, adjustment without instance placement can also improve
while the performance of average assignment is near to the SLA fulfilled ratio.
the original settings. Thus, instance placement improves
efficiency of the optimization algorithm with more in-
stances in virtual networks. 7 C ONCLUSION AND F UTURE W ORK
We also test the influence on the QoS performance In this paper, we first investigate the virtual network
with different bandwidth of core switches. The number recognition problem in a general-purpose cloud envi-
of instances per each virtual network is uniform dis- ronment. Since the user virtual network topologies is
tributed from 5 to 35. The bandwidth of core switches not acknowledged, it is hard to implement an efficient
increases from 20 Gbps to 80 Gbps and increases 10Gbps optimization for cloud computing. Thus, we propose
in each step. As shown in Fig. 7, SLA fulfilled ratio an optimization method that first recognizes virtual
increases with higher bandwidth of core switches. When networks then optimizes the performance of the rec-
the bandwidth of core switches is set to 20 Gbps, SLA ognized virtual networks. We design a virtual network
fulfilled ratio is less than 10% with original settings recognition algorithm by formulating the recognition
and near to 80% with our optimization. The fulfilled problem as a classic community detection problem. The
ratio with our optimization method without instance optimization algorithm focuses on the QoS performance
placement is near to 40%. Average assignment performs of each virtual network. We adopt the SLA fulfillment
similarly with original settings. After the bandwidth of as the measurement of QoS and formulate the optimiza-
core switches increases to 80Gbps, the ratio of fulfilled tion problem as a non-cooperative game with a Pareto

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

10

efficient solution. Finally, we evaluate our solution and [16] X. Meng, V. Pappas, and L. Zhang, “Improving the scalabili-
compare the performance to the original placement. In ty of data center networks with traffic-aware virtual machine
placement,” in Proceedings of the 29th Conference on Information
the future, we first plan to adopt pattern recognition Communications, ser. INFOCOM’10. Piscataway, NJ, USA: IEEE
to detect different applications from the network traffic Press, 2010, pp. 1154–1162.
records, since different applications have different re- [17] D. Jayasinghe, C. Pu, T. Eilam, M. Steinder, I. Whally, and
E. Snible, “Improving performance and availability of services
quirements on cloud resources. hosted on iaas clouds with structural constraint-aware virtual
machine placement,” in 2011 IEEE International Conference on
Services Computing, July 2011, pp. 72–79.
ACKNOWLEDGMENTS [18] V. Shrivastava, P. Zerfos, K. w. Lee, H. Jamjoom, Y. H. Liu, and
This work is supported by JSPS KAKENHI Grant Num- S. Banerjee, “Application-aware virtual machine migration in data
centers,” in 2011 Proceedings IEEE INFOCOM, April 2011, pp. 66–
ber JP16K00117, JP15K15976, JP17K12669, KDDI Foun- 70.
dation, and Research Fund for Postdoctoral Program of [19] X. Wen, K. Chen, Y. Chen, Y. Liu, Y. Xia, and C. Hu, “Virtualknot-
Muroran Institute of Technology. Mianxiong Dong is the ter: Online virtual machine shuffling for congestion resolving in
virtualized datacenter,” in 2012 IEEE 32nd International Conference
corresponding author. on Distributed Computing Systems, June 2012, pp. 12–21.
[20] O. Biran, A. Corradi, M. Fanelli, L. Foschini, A. Nus, D. Raz,
R EFERENCES and E. Silvera, “A stable network-aware vm placement for cloud
systems,” in 2012 12th IEEE/ACM International Symposium on
[1] B. He, M. Lu, K. Yang, R. Fang, N. K. Govindaraju, Q. Luo, Cluster, Cloud and Grid Computing (ccgrid 2012), May 2012, pp.
and P. V. Sander, “Relational query coprocessing on graphics 498–506.
processors,” ACM Trans. Database Syst., vol. 34, no. 4, pp. 21:1– [21] M. Alicherry and T. V. Lakshman, “Network aware resource allo-
21:39, Dec. 2009. cation in distributed clouds,” in 2012 Proceedings IEEE INFOCOM,
[2] L. Hu, X. Che, and Z. Xie, “Gpgpu cloud: A paradigm for general March 2012, pp. 963–971.
purpose computing,” Tsinghua Science and Technology, vol. 18, [22] L. Hu, K. D. Ryu, D. D. Silva, and K. Schwan, “v-bundle: Flexible
no. 1, pp. 22–23, Feb 2013. group resource offerings in clouds,” in 2012 IEEE 32nd Interna-
[3] A. Beloglazov and R. Buyya, “Energy efficient resource man- tional Conference on Distributed Computing Systems, June 2012, pp.
agement in virtualized cloud data centers,” in Proceedings of the 406–415.
2010 10th IEEE/ACM International Conference on Cluster, Cloud and [23] T. Benson, A. Akella, A. Shaikh, and S. Sahu, “Cloudnaas: A cloud
Grid Computing, ser. CCGRID ’10. Washington, DC, USA: IEEE networking platform for enterprise applications,” in Proceedings of
Computer Society, 2010, pp. 826–831. the 2Nd ACM Symposium on Cloud Computing, ser. SOCC ’11. New
[4] W. Fang, X. Liang, S. Li, L. Chiaraviglio, and N. Xiong, “Vm- York, NY, USA: ACM, 2011, pp. 8:1–8:13.
planner: Optimizing virtual machine placement and traffic flow [24] F. Francois and E. Gelenbe, “Towards a cognitive routing engine
routing to reduce network power costs in cloud data centers,” for software defined networks,” in 2016 IEEE International Confer-
Computer Networks, vol. 57, no. 1, pp. 179 – 196, 2013. ence on Communications (ICC), May 2016, pp. 1–6.
[5] R. Mijumbi, J. Serrat, J. L. Gorricho, N. Bouten, F. D. Turck, and [25] F. Francois and E. Gelenbe, “Optimizing secure sdn-enabled
S. Davy, “Design and evaluation of algorithms for mapping and inter-data centre overlay networks through cognitive routing,” in
scheduling of virtual network functions,” in Network Softwariza- 2016 IEEE 24th International Symposium on Modeling, Analysis and
tion (NetSoft), 2015 1st IEEE Conference on, April 2015, pp. 1–9. Simulation of Computer and Telecommunication Systems (MASCOTS),
[6] T. Truong Huu, G. Koslovski, F. Anhalt, J. Montagnat, and P. Vicat- Sept 2016, pp. 283–288.
Blanc Primet, “Joint elastic cloud and virtual network framework [26] A. Tavakoli, M. Casado, T. Koponen, and S. Shenker, “Applying
for application performance-cost optimization,” Journal of Grid nox to the datacenter,” in Proc. of workshop on Hot Topics in
Computing, vol. 9, no. 1, pp. 27–47, 2011. Networks (HotNets-VIII), 2009.
[7] M. Xu, L. Cui, H. Wang, and Y. Bi, “A multiple qos constrained [27] B. Pfaff, J. Pettit, K. Amidon, M. Casado, T. Koponen, and
scheduling strategy of multiple workflows for cloud computing,” S. Shenker, “Extending networking into the virtualization layer,”
in 2009 IEEE International Symposium on Parallel and Distributed in HotNets, 2009.
Processing with Applications, Aug 2009, pp. 629–634. [28] B. P. M. C. Justin Pettit, Jesse Gross, “Virtual switching in an era
[8] R. L. Krutz and R. D. Vines, Cloud security: A comprehensive guide of advanced edges,” in 2nd Workshop on Data Center-Converged and
to secure cloud computing. Wiley Publishing, 2010. Virtual Ethernet Switching, 2010, pp. 1–7.
[9] R. Jain and S. Paul, “Network virtualization and software defined [29] M. Moshref, M. Yu, A. Sharma, and R. Govindan, “Scalable rule
networking for cloud computing: a survey,” IEEE Communications management for data centers,” in Proceedings of the 10th USENIX
Magazine, vol. 51, no. 11, pp. 24–31, November 2013. Conference on Networked Systems Design and Implementation, ser.
[10] P. Wieder, J. M. Butler, W. Theilmann, and R. Yahyapour, Service nsdi’13. Berkeley, CA, USA: USENIX Association, 2013, pp. 157–
level agreements for cloud computing. Springer Science & Business 170.
Media, 2011. [30] Q. Zhang, L. Cheng, and R. Boutaba, “Cloud computing: state-
[11] L. Wang and E. Gelenbe, “Adaptive dispatching of tasks in the of-the-art and research challenges,” Journal of Internet Services and
cloud,” IEEE Transactions on Cloud Computing, vol. PP, no. 99, pp. Applications, vol. 1, no. 1, pp. 7–18, 2010.
1–1, 2015. [31] Q. Li, J. Huai, J. Li, T. Wo, and M. Wen, “Hypermip: Hypervisor
[12] L. Wang, O. Brun, and E. Gelenbe, “Adaptive workload distri- controlled mobile ip for virtual machine live migration across
bution for local and remote clouds,” in 2016 IEEE International networks,” in 2008 11th IEEE High Assurance Systems Engineering
Conference on Systems, Man, and Cybernetics (SMC), Oct 2016, pp. Symposium, Dec 2008, pp. 80–88.
003 984–003 988. [32] P. Raad, G. Colombo, D. P. Chi, S. Secci, A. Cianfrani, P. Gallard,
[13] E. Gelenbe and L. Wang, “Tap: A task allocation platform for the and G. Pujolle, “Achieving sub-second downtimes in internet-
eu fp7 panacea project,” in Advances in Service-Oriented and Cloud wide virtual machine live migrations in lisp networks,” in 2013
Computing: Workshops of ESOCC 2015, Taormina, Italy, September IFIP/IEEE International Symposium on Integrated Network Manage-
15–17, 2015, Revised Selected Papers, vol. 567, 2016, p. 425. ment (IM 2013), May 2013, pp. 286–293.
[14] M. F. Bari, R. Boutaba, R. Esteves, L. Z. Granville, M. Podlesny, [33] M. Coudron, S. Secci, G. Maier, G. Pujolle, and A. Pattavina,
M. G. Rabbani, Q. Zhang, and M. F. Zhani, “Data center network “Boosting cloud communications through a crosslayer multipath
virtualization: A survey,” IEEE Communications Surveys Tutorials, protocol architecture,” in 2013 IEEE SDN for Future Networks and
vol. 15, no. 2, pp. 909–928, Second 2013. Services (SDN4FNS), Nov 2013, pp. 1–8.
[15] Y. Chen, R. Griffith, J. Liu, R. H. Katz, and A. D. Joseph, “Under- [34] The OpenDaylight Project, Inc., “OpenDaylight - Technical
standing tcp incast throughput collapse in datacenter networks,” Overview,” 2013. [Online]. Available: [Link]
in Proceedings of the 1st ACM Workshop on Research on Enterprise org/project/technical-overview
Networking, ser. WREN ’09. New York, NY, USA: ACM, 2009, [35] P. S. Pisa, N. C. Fernandes, H. E. T. Carvalho, M. D. D. Moreira,
pp. 73–82. M. E. M. Campista, L. H. M. K. Costa, and O. C. M. B. Duarte,

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2871118, IEEE
Transactions on Cloud Computing

11

OpenFlow and Xen-Based Virtual Network Migration. Berlin, Hei- Mianxiong Dong received B.S., M.S. and Ph.D.
delberg: Springer Berlin Heidelberg, 2010, pp. 170–181. in Computer Science and Engineering from The
[36] V. Mann, A. Vishnoi, K. Kannan, and S. Kalyanaraman, “Cross- University of Aizu, Japan. He is currently an
roads: Seamless vm mobility across data centers through software Associate Professor in the Department of Infor-
defined networking,” in 2012 IEEE Network Operations and Man- mation and Electronic Engineering at the Muro-
agement Symposium, April 2012, pp. 88–96. ran Institute of Technology, Japan. He was a
[37] S. Fortunato, “Community detection in graphs,” Physics Reports, JSPS Research Fellow with School of Com-
vol. 486, no. 3C5, pp. 75 – 174, 2010. puter Science and Engineering, The University
[38] J. Leskovec, K. J. Lang, and M. Mahoney, “Empirical comparison of Aizu, Japan and was a visiting scholar with
of algorithms for network community detection,” in Proceedings BBCR group at University of Waterloo, Canada
of the 19th International Conference on World Wide Web, ser. WWW supported by JSPS Excellent Young Researcher
’10. New York, NY, USA: ACM, 2010, pp. 631–640. Overseas Visit Program from April 2010 to August 2011. Dr. Dong was
[39] T. Benson, A. Akella, and D. A. Maltz, “Network traffic character- selected as a Foreigner Research Fellow (a total of 3 recipients all over
istics of data centers in the wild,” in Proceedings of the 10th ACM Japan) by NEC C&C Foundation in 2011. His research interests include
SIGCOMM Conference on Internet Measurement, ser. IMC ’10. New Wireless Networks, Cloud Computing, and Cyber-physical Systems.
York, NY, USA: ACM, 2010, pp. 267–280. He has received best paper awards from IEEE HPCC 2008, IEEE
ICESS 2008, ICA3PP 2014, GPC 2015, IEEE DASC 2015, IEEE VTC
2016-Fall, FCST 2017, 2017 IET Communications Premium Award and
IEEE ComSoc CSIM Best Conference Paper Award 2018. Dr. Dong
serves as an Editor for IEEE Transactions on Green Communications
and Networking (TGCN), IEEE Communications Surveys and Tutorials,
IEEE Network, IEEE Wireless Communications Letters, IEEE Cloud
Computing, IEEE Access, as well as a leading guest editor for ACM
He Li received the B.S., M.S. degrees in Com-
Transactions on Multimedia Computing, Communications and Applica-
puter Science and Engineering from Huazhong
tions (TOMM), IEEE Transactions on Emerging Topics in Computing
University of Science and Technology in 2007
(TETC), IEEE Transactions on Computational Social Systems (TCSS).
and 2009, respectively, and the Ph.D. degree
He has been serving as the Vice Chair of IEEE Communications Soci-
in Computer Science and Engineering from The
ety Asia/Pacific Region Information Services Committee and Meetings
University of Aizu in 2015. He is currently a
and Conference Committee, Leading Symposium Chair of IEEE ICC
Postdoctoral Fellow with Department of Informa-
2019, Student Travel Grants Chair of IEEE GLOBECOM 2019, and
tion and Electronic Engineering, Muroran Insti-
Symposium Chair of IEEE GLOBECOM 2016, 2017. He is the recipient
tute of Technology, Japan. His research interests
of IEEE TCSC Early Career Award 2016, IEEE SCSTC Outstanding
include cloud computing and software defined
Young Researcher Award 2017, The 12th IEEE ComSoc Asia-Pacific
networking. He has received the best paper
Young Researcher Award 2017 and Funai Research Award 2018. He
award from IEEE VTC2016-Fall. Dr. Li serves as an Associate Editor
is currently the Member of Board of Governors and Chair of Student
for Human-centric Computing and Information Sciences (HCIS), as well
Fellowship Committee of IEEE Vehicular Technology Society.
as a Guest Associate Editor for IEICE Transactions on Information
and Systems. He is the recipient of IEEE TCSC Outstanding Ph.D.
Dissertation Award 2016.

Kaoru Ota was born in Aizu-Wakamatsu, Japan.

She received M.S. degree in Computer Science
from Oklahoma State University, USA in 2008,
B.S. and Ph.D. degrees in Computer Science
and Engineering from The University of Aizu,
Japan in 2006, 2012, respectively. She is cur-
rently an Assistant Professor with Department
of Information and Electronic Engineering, Muro-
ran Institute of Technology, Japan. From March
2010 to March 2011, she was a visiting scholar
at University of Waterloo, Canada. Also she was
a Japan Society of the Promotion of Science (JSPS) research fellow
with Kato-Nishiyama Lab at Graduate School of Information Sciences at
Tohoku University, Japan from April 2012 to April 2013. Her research
interests include Wireless Networks, Cloud Computing, and Cyber-
physical Systems. Dr. Ota has received best paper awards from ICA3PP
2014, GPC 2015, IEEE DASC 2015, IEEE VTC 2016-Fall, FCST 2017,
2017 IET Communications Premium Award and IEEE ComSoc CSIM
Best Conference Paper Award 2018. She is an editor of IEEE Trans-
actions on Vehicular Technology (TVT), IEEE Communications Letters,
Peer-to-Peer Networking and Applications (Springer), Ad Hoc & Sensor
Wireless Networks, International Journal of Embedded Systems (Inder-
science) and Smart Technologies for Emergency Response & Disaster
Management (IGI Global), as well as a guest editor of ACM Transactions
on Multimedia Computing, Communications and Applications (leading),
IEEE Internet of Things Journal, IEEE Communications Magazine, IEEE
Network, IEEE Wireless Communications, IEEE Access, IEICE Trans-
actions on Information and Systems, and Ad Hoc & Sensor Wireless
Networks (Old City Publishing). She is the recipient of IEEE TCSC Early
Career Award 2017.

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.