Reliable Virtual Data Center Embedding Across Multiple Data

Centers

Gang Sun

1,2

, Sitong Bu

, Vishal Anand

, Victor Chang

and Dan Liao

Key Lab of Optical Fiber Sensing and Communications (Ministry of Education), UESTC, Chengdu, China

Institute of Electronic and Information Engineering in Dongguan, UESTC, Chengdu, China

Department of Computer Science, The College at Brockport, State University of New York, New York, U.S.A.

Leeds Beckett University, Leeds, U. K.

Keywords: Virtual Data Center, Embedding, Reliability, Multi-domain.

Abstract: Cloud computing has become a cost-effective paradigm for deploying online service applications in large data

centers in recent years. Virtualization technology enables flexible and efficient management of physical

resources in cloud data centers and improves the resource utilization. A request for resources to a data center

can be abstracted as a virtual data center (VDC) request. Due to the use of a large number of resources and at

various locations, reliability is an important issue that should be addressed in large and multiple data centers.

However, most research focuses on the problem of reliable VDC embedding in a single data center. In this

paper, we study the problem of reliable VDC embedding across multiple data centers, such that the total

bandwidth consumption in the inter-data center backbone network is minimized, while satisfying the

reliability requirement of each VDC request. We model the problem by using mixed integer linear

programming (MILP) and propose a heuristic algorithm to address this NP-hard problem efficiently.

Simulation results show that the proposed algorithm performs better in terms of lowering physical resource

consumption and VDC request blocking ratio compared with existing solution.

1 INTRODUCTION

Cloud computing has become a promising paradigm

that enables users to take advantage of various

distributed resources (Armbrust et al., 2010). In a

cloud computing environment, the cloud provider(s)

owning the physical resource pool (i.e., physical data

centers) offer these resources to multiple

users/clients. A user’s demand for resources can be

abstracted as a virtual data center (VDC), also known

as a virtual infrastructure. A typical VDC consists of

virtual machines (VMs) connected through virtual

switches, routers and links with communication

bandwidth. The VDCs are able to provide better

isolation and utilization of network resources, thereby

improving the performance of services and

applications. The main challenge associated with

VDC management in cloud data centers is efficient

VDC embedding (or mapping), which aims at finding

a mapping of VMs and virtual links to physical

components (e.g., servers, switches and links).

Due to the use of a large number of resources and

at various locations in cloud data centers, providing

reliability of cloud services is an important issue that

needs attention. A service disruption can lead to

customer dissatisfaction and revenue loss. However,

restoring failed services is costly. Thus, many cloud

services are deployed in distributed data centers to

meet their high reliability requirements. Furthermore,

some services may require to be within the proximity

of end-users (e.g., Web servers) whereas others may

not have such location constraints and can be

embedded in any data center (e.g., MapReduce jobs)

(Amokrane et al., 2013). Therefore, reliable VDC

embedding across distributed infrastructures is

appealing for Service Providers (SP) as well as

Infrastructure Providers (InPs).

However, there is limited research on the problem

of reliable VDC embedding across multiple data

centers, with most research focusing on only VDC

embedding within a single data center. For example, C.

Guo et al. (2013) proposed a data center network

virtualization architecture, SecondNet, in which VDC

is used as the basic unit of resource allocation for

multiple tenants in the cloud. However, they do not

consider the VDC reliability. M. Zhani et al. (2013)

designed a migration- aware dynamic VDC embedding

framework for efficient VDC planning, which again

Sun, G., Bu, S., Anand, V., Chang, V. and Liao, D.

Reliable Virtual Data Center Embedding Across Multiple Data Centers.

DOI: 10.5220/0005842101950203

In Proceedings of the International Conference on Internet of Things and Big Data (IoTBD 2016), pages 195-203

ISBN: 978-989-758-183-0

195

without reliability considerations. Zhang et al. (2014)

proposed a framework, Venice, for reliable virtual data

center mapping. The authors prove that the reliable

VDC embedding problem is NP-hard problem.

However, this study is limited to a single data center.

There are also some studies on virtual data center

or virtual network (VN) embedding across multiple

domains. The work in (Amokrane et al., 2013) studied

the problem of VDC embedding across distributed

infrastructures. The authors in (Yu et al., 2014)

presented a framework of cost efficient VN

embedding across multiple domains, called MD-

VNM. Di et al. (2014) proposed an algorithm for

reliable virtual network embedding across multiple

data centers. However, there are differences between

VDC embedding and VN embedding in some aspects.

For example, a VDC is a collection of VMs, switches

and routers that are interconnected through virtual

links. Each virtual link is characterized by its

bandwidth capacity and its propagation delay. VDCs

are able to provide better isolation of network

resources, and thereby improve the performance of

service applications. A VN is a combination of active

and passive network elements (network nodes and

network links) on top of a Substrate Network (SN)

(Fischer et al., 2013).In addition, a VN node generally

cannot be embedded on a physical node that hosts

another VN node of the same VN, whereas each

physical server can host multiple VMs from the same

VDC in the VDC embedding problem. Therefore,

most existing VDC embedding approaches cannot be

directly applied to solve the problem of reliable VDC

embedding across distributed infrastructures.

In this paper, we study the problem of reliable

VDC embedding (RVDCE) across multiple data

centers. In our work, we minimize the total backbone

bandwidth consumption, while satisfying the

reliability requirement of VDC request. We formulate

the RVDCE problem as a mixed integer linear

programming (MILP) problem and propose a cost

efficient algorithm for solving this problem.

Simulation results show that the proposed algorithm

performs better than the existing approach.

2 PROBLEM STATEMENT

In this section, we give the descriptions on the

problem of reliable VDC embedding across

distributed infrastructures.

2.1 VDC Request

A VDC request consists of multiple VMs and virtual

links that connect these VMs. Figure 1 shows an

example of a VDC request with 5 VMs,

interconnected by virtual links. The i-th VDC request

is represented by G

= (V

, E

), where V

denotes the set

of virtual machines, E

is the set of virtual links. The

VMs with location constraints can only be embedded

onto the specified data centers, whereas the VMs

without location constraints can be embedded onto

any data center in the substrate infrastructure. We use

R to denote the types of resources (e.g., CPU or

memory) offered by each physical node. We use







to denote the requirement of resource of virtual

node v, b

to present the amount of bandwidth

required by virtual link e. Similarly, we define s

and

as the variables that indicate whether virtual node

v is the source or destination of link e.

We b

App

App Database

Database

Figure 1: Example of a VDC request.

2.2 Distributed Infrastructure

The distributed infrastructure consists of the

backbone network and data centers managed by

infrastructure providers (InP).Thus, the InP knows

the information of all the distributed data centers.

Typically, the cost of per unit bandwidth in the

backbone network is more expensive than the cost of

per unit bandwidth within data center (Greenberg et

al., 2008).

Figure 2: The Fat-Tree topology.

We model the physical infrastructure as a undirected

graph G=(



∪



,



∪



), where 



denotes the

set of physical nodes in the data centers, 



denotes

IoTBD 2016 - International Conference on Internet of Things and Big Data

196

the set of nodes in backbone network, 



denotes the

set of physical links of data centers, and 



denotes

the set of physical links of the backbone and physical

links that connect the backbone and data centers. Let





=(





,





) represent the physical data center k,

where





denotes the set of physical nodes and 





denotes the set of physical links in the data center. We

use 







to indicate the capacity of resource on the

physical node ̅ , and 



to indicate bandwidth

capacity of physical link ̅. Let 







,







be indicator

variables that denote whether ̅ is the source or

destination of physical link ̅ .Without loss the

generality, in this work, we assume that each data

center has a Fat-Tree (Leiserson, 1985) topology as

shown in Figure 2.

2.3 Reliable VDC Embedding

We define the reliability of a VDC as the probability

that the service is still available even after the failure

of physical servers. Similar to (Zhang et al., 2014),

we use replication groups to guarantee the reliability

requirements of data centers. The basic idea is that if

one VM in the replication group fails, the other VM

in the same replication group can run as a backup. In

other words, a replication group is reliable as long as

one of the VMs in the group is available. The VMs in

different replication groups have different

functionalities, and the set of all replication groups

form the complete service. Therefore, when any

group fails, the entire service is unavailable, since the

rest of groups cannot form a complete service. In this

work, the key objective is to guarantee the reliability

of the whole VDC, while satisfying its resource

requirements.

The reliability rl of a VDC which has been

embedded can be calculated as follows:

(1 ) ,

iRC

vF vNF

rfr NlvPr

∈

∈∈

=−∀∈



∏∏

, (1)

where PN denotes the set of physical nodes which

host the VMs in the VDC; RC denotes the set of cases

which can guarantee the availability of the VDC; 





denotes the failure probability of the physical node ̅;

F is the set of the failed physical nodes in all data

centers; and NF is the set of the available physical

nodes in data centers.

The problem of VDC embedding across multiple

domains (i.e., multiple data centers) means that VMs

in a VDC may be embedded in multiple data centers.

Since VMs have different location constraints, they

may not be embedded in the same data center. Figure

3 shows an example of VDC embedding across

multiple domains.

Backbone

network

Virtual data

center

DC1

DC2

DC4

DC3

Figure 3: Example of VDC cross-domain embedding.

3 MILP MODEL

In this section we describe the constraints and

objective of our researched problem and then

construct a MILP model.

3.1 The Constraints

In order to ensure that the VDC embedding does not

violate the physical resource capacity limits, the

following capacity constraints must be met.

, ,

iir r

vv v

cc vVrR

∈

≤∀∈∈



, (2)

ee e

bb eE

∈

≤∀∈



, (3)

Where 







is a variable denoting whether the virtual

node v(i.e., VM v)of VDCi is embedded on the

physical node ̅ ; 





denotes the amount of

bandwidth resources provisioned by physical

link̅for virtual link e of VDC i. In addition, the flow

conservation constraints below must also be satisfied.

,,e,

ve ee ve ee

eE eE

ii i i i

ve e ve e

vv vv

vV vV

sb db

sb xdb i I E v V

∈∈

−

=− ∀∈∈∈





(4)

We define 







as a variable that indicates whether the

VM v can be embedded on physical node . Then the

virtual node placement constraints can be formulated

as follows:



,, ,

xiIvVvV≤∀∈ ∈ ∈

, (5)

Reliable Virtual Data Center Embedding Across Multiple Data Centers

197

1, ,

iIvV

∈

=∀∈∈



. (6)

We define 





as a variable to indicate whether the

physical node ̅ is active. If a physical node hosts at

least one VM, then this node must be active,

otherwise inactive. It means that the following

constraints should be satisfied:

, , ,

vvv

yx iIvVvV≥∀∈∈∈

, (7)

, , , ,

veeve

bs iIvVeEeE

≥∀∈∈∈∈

, (8)

, , , ,

veeve

bd iIvVeEeE

≥∀∈∈∈∈

.(9)

Furthermore, the following geographical location

constraints must be satisfied:

1 if can be assigned to DC

0otherwise







, (10)

1 if is assigned to DC

0otherwise







, (11)

ii i

kv kv

wz vV≤∀∈

, (12)

1, : G

∈

=∀



. (13)

3.2 The Objective

We use 



to represent the reliability of i-th VDC. If

the InP fails to meet the reliability requirement 



there will be a penalty:

(1 )

unavail

∈

=−



, (14)

where π

denotes the penalty from failing to meet the

reliability of the i-th VDC.

Recovery costs come from restarting VMs and

reconfiguring the network equipment. We thus define

the failure recovery costs of node ̅ and link ̅ as

shown in Equation (15) and (16), respectively. Let 





and 



represent the energy cost of an active node ̅

and link ̅, respectively.

restore i i

v v vv ee ve

vV eE

PxbvV

ρλμλ

∈

∈∈

=+ + ∀∈

 

,(15)

, \

restore i

ee ee

beEBE

ρλ

∈

=+ ∀∈



, (16)

where 



and 



are the recovery costs of virtual

node v and virtual link e, and 







{







,







} denotes that whether node ̅ is the

source or destination of link ̅ . The cost for

embedding e onto backbone network is:

backbone i

eee

PbeBE

∈

=∀∈



, (17)

where 



denotes the provisioning cost of virtual

link e in backbone network.

We use 





and 



to denote the failure probability

of node ̅ and link ̅, respectively. Then the total cost

of unreliable VDC is as follows:

unavail restore

restore backbone

ee e

PP FP

FP P

∈



. (18)

Therefore, the objective function to minimize the

total cost defined as follows:

inimize P

. (19)

4 ALGORITHM DESIGN

In this section, we propose an efficient algorithm for

solving the problem of reliable VDC embedding

(RVDCE) across multiple domains. The RVDCE

problem consists of two key issues: i) how to ensure

the reliability of the VDC request; ii) how to reduce

the backbone bandwidth consumption.

Multiple VMs may be embedded on the same

physical node, and the reliability of embedding all

VMs of a VDC on the same physical node is higher

than that of embedding the VMs on different nodes.

The reason is as follows: according to the definition

of the reliability requirement of a VDC, reliable

service implies that all replication groups are reliable.

All replication groups form the complete service, and

each plays a unique role as explained in the third part

of Section 2.Therefore, when any group fails, the

entire service is unavailable. Hence, if we embed all

VMs belonging to different replication groups on the

same physical node, the service reliability is equal to

the reliability of that physical node. On the other

hand, if we embed all VMs on different physical

nodes, the service reliability is equal to the product of

the reliability of those physical nodes. In addition, in

order to improve the service reliability, we need to

embed VMs in the same replication group on to

different physical node so that these VMs can backup

IoTBD 2016 - International Conference on Internet of Things and Big Data

198

each other. However, physical nodes with limited

resources and VMs with location constraints may not

allow all VMs in a VDC to be embedded onto the

same physical node, or even the same data center.

Moreover, reducing the backbone bandwidth

consumption is the other issue we have to address in

this work. For addressing this issue, we partition a

VDC into several partitions before mapping it. The

reasons are as follows: i) if we place VMs one by one,

it will cause a large amount of backbone bandwidth

consumption, because the backbone bandwidth

consumption is not considered in the VM mapping

process; ii) the network resource in inter-data centers

is more expensive than that of intra-data centers

networks. We put the VMs that have a large amount

of communication bandwidth requirement between

each other into the same partition, and the VMs in the

same partition will be mapped into the same data

center. Thus, the backbone bandwidth consumption

can be reduced. On the other hand, the way of

mapping VMs one by one mentioned above will

cause much higher reliability than required. It is

unnecessary that the data centers provide much higher

reliability than that required by VDC request, so it is

necessary to just meet the required reliability for

reducing the resource consumption.

In addition, if we map virtual components with

low reliability requirement on the servers with high

reliability, the physical server resources will be

wasted and may result in increasing in the blocking

ratio of the VDC requests in the long term. Therefore,

in order to use physical server resource with different

reliability rationally, we grade the servers in data

center according to their reliability.

The RVDCE algorithm consists of three main

steps: i) group the physical servers; ii) partition the

VDC; iii) embed the VDC partitions. The detailed

RVDCE algorithm as shown in Algorithm 1.

Algorithm 1: Reliable VDC Embedding.

Input:1. Size of partitions:

;

2. The VDC request: (V ,E )

iii

G = ;

3. Substrate data center:

(,)GVBVEBE= 

Output: The VDC embedding result.

1:Sort servers in data centers in

ascending order of their reliability;

2:Divide the servers in data centers

into

levels, where a lower level has

a lower reliability;

3: The initial partition size denoted as

, let

;

4: letisSuccessful ← false;

5:while

0K >

6: Call Procedure1 to partition the

VDC request;

7: for all

lL∈

←the set of servers in all data

centers whose levels are lower

than or equal to

;

9: Call Procedure 2 to embed

partitions;

10: if all partitions are embedded

successfully

11: isSuccessful ← true;

12: return VDC embedding result;

13: end if

14: end for

15:

K=−

16:end while

17:if (isSuccessful == false)

18: return VDC embedding is failed

19:end if

4.1 Group the Physical Servers

We sort the servers in descending order of their

reliability, and then group the servers into groups with

different reliability levels. That is, higher level means

higher reliability. In the embedding process, we chose

level l servers to host a VDC, meaning that we can

use the servers whose levels are less than or equal to

l for hosting the VDC. If the reliability does not meet

VDC requirement after completing the VDC

embedding, it is necessary to increase the reliability

level l and re-embed the VDC. If the reliability is

lower than the required for any available physical

servers, we have to change the partition size and re-

embed the VDC.

4.2 Partition the VDC

Before embedding a VDC, we first partition the VDC

into several sub-VDCs, with the aim of minimizing

the bandwidth demands between partitions, thereby

reducing the bandwidth consumption in the backbone

network.

Assume the number of VMs in a VDC is N, the

partition size is K(i.e., the number of VMs in each

partition cannot exceed K), where the partition size K

is adjustable. The initial value of K is equal to N, then

the K is gradually reduced in the process of adjusting

the size. If a VDC is successfully embedded while K=

N, we do not need to adjust the size K, since the

backbone bandwidth consumption is minimized.

In the VDC partition process, we first make each

node as a partition, and calculate the total amount of

bandwidth demands between the original partitions.

Then the algorithm traversals all nodes ∈



, and

Reliable Virtual Data Center Embedding Across Multiple Data Centers

199

finds partition P that allows us move v from its

original partition to P and satisfy the follow

conditions: i) reduce the amount of bandwidth in

backbone network; ii) the VMs in P have same

location constraints; iii) the number of VMs in P is

smaller than k. If a partition P that meets the above

conditions is found, we will move VM v to P. As long

as there are nodes moving between different

partitions, algorithm keeps traversing the VMs until

there is no node need to be moved. If the current

bandwidth demands in inter-data center is less than

the bandwidth demands between original partitions,

algorithm will regenerate a new graph where a

partition of 



is as a node in this new graph.

Procedure 1: VDC Partition.

1: Let flag← true;

2: while(flag)

3: Denote each node of

as apartition;

4: Record the initial bandwidths

between partitions;

5: while nodes need to be moved between

partitions do

6: for each

vV∈

, do

7: Find a partition

and move

, such that: a) the

number of VMs in

does not

exceed

; b)bandwidths

consumptionis minimized; c)

all of the nodes in

with

same location constraint.

8: end for

9: end while

10: if current bandwidth demands <

initial bandwidth demands

11: Change

as the graph of

partitions;

12: else

13: flag ←false

14: end if

15:end while

Figure 4 shows an example of partitioning a VDC

request. We assume that two data centers included in

the substrate network, denoted as DC1 and DC2. The

VDC request m is shown in Figure 4(a). The location

constraints of the four VMs are as follows: i)VM a

must be embedded in DC1;ii) VM b and VM c must

be embedded in DC2; iii)VM d can be embedded in

DC1 orDC2. Therefore, VM a cannot be in the same

partition with VM b and VM c. Figure 4(b) shows the

partitioning of VDC m when the partition size is 1.

When the partition size is 2, the feasible partitions of

m are shown in Figure 4(b) and Figure 4(c). When the

partition size is 3, the feasible partitions of m are

shown in Figure 4(b), Figure4(c) and Figure 4(d).

a d

(a) (b)

a d

2 4

3 5

Figure 4: Example of partitioning a VDC.

4.3 Embed the Partitions

We guarantee the reliability of VDC by using

replication groups. If one VM in the replication group

fails, the other VM in the same replication group can

run as a backup. Therefore, a replication group is

reliable as long as at least one of the VMs in the group

is reliable. We choose one node in a replication group

as the working node, the other nodes in that

replication group can be used for backup. We take

different approaches for embedding the working node

and backup nodes.

4.3.1 Embedding of Working Nodes

Since the VMs in partition have location constraints,

each partition has a corresponding location constraint.

IoTBD 2016 - International Conference on Internet of Things and Big Data

200

Partitions can only be embedded on to data centers

that satisfy the location constraints.

We choose the first partition randomly and find a

data center that can provide the highest reliability for

the chosen partition. Then we embed the VMs in the

partition according to the following two steps: 1)

embedding the VMs without backups;2) embedding

the VMs that with backups. If any VM in the

replication group has been embedded, the VMs

belong to this replication group in the partition should

be skipped. In other words, only one VM of each

replication group needs to be embedded. We

preferentially embed it on the servers that have hosted

other VMs belong to the same VDC. If the resources

of the physical server are not enough, we need to

embed the VM on a new available server which with

the highest reliability.

4.3.2 Embedding of Backup Nodes

We calculate the reliability of the VDC by using the

reliability computing algorithm (Zhang et al., 2014),

after embedding the working node. If the reliability is

higher than required, we embed the backups into the

physical servers with lower reliability. If the reliability

is lower than required, we embed the backups into the

physical servers with higher reliability.

Procedure 2: Partition embedding.

1:Randomly chose a partition

from

partition set

artitions

;

2:Choose a data center

with the

highest reliability and enough

resources from data center set

;

artitionToDC

∪<

4:M: the set of nodes in VDC that have

been embedded, let

;

5:while

artitions

is not empty do

6: for all

7: if none of the virtual node in

the replication group that

belongs to has been embedded

8: if v can be embedded on the

server

S∈

hosting the

nodes in M

9: Embed

on server

;

10:

{};

Mv←∪

11: else

12: Chose the server has

highest reliability that

belongs to

and

to host

;

13: end if

14: end if

15: end for

16:

\{g};Partitions Partitions←

17:

← the partition in

artitions

which

has the largest amount of bandwidth

demand to communicate with the

embedded partitions;

18: Choose a

that can provide highest

reliability and enough resources to

g from

;

19:

artitionToDC

∪<

20:end while

21:Compute the current reliability

VDC request by using the reliability

computing algorithm proposed in

(Zhang et al., 2014);

22:for all <

> in

artitionToDC

23: if

24: for each node

25: Embed

on the server

with enough resources

and the lowest

reliability in

;

26: end for

27: else

28: for remaining nodes

29: if v can be embedded on

the server that node

in g has been embedded

on and the replication

groups of

and

are

different

30: Embed

;

31: else

32:

Embed

on the server

has highest

reliability;

33: end if

34: end for

35: end if

36:end for

We note that all VMs belonging to the same partition

must be embedded in the same data center.

Accordingly, when embedding the backup VMs, we

need to choose the data center that has the embedded

VMs belonging to the same partition. In addition, it is

important to note that the VMs belonging to the same

replication group cannot be embedded on to the same

server. After embedding the backup nodes, we

calculate the reliability and check whether the

reliability meets the VDC requirement.

Reliable Virtual Data Center Embedding Across Multiple Data Centers

201

5 SIMULATIONS AND

ANALYSIS

5.1 Simulation Environment

In our simulation, each data center has a Fat-Tree

(Leiserson, 1985) topology composed of 128 physical

servers connected through 80 switches. Each physical

server has 1000units CPU resource. The capacity of

physical link in data center is 100 units. We generate

the VDC requests by using GT-ITM tool (GT-ITM)

and the parameter settings are similar to (Zhang et al.,

2014). The number of VMs in each VDC request is

randomly set between 3 and 15.The CPU resource

demand of each VM is generated randomly between

8 to 16 units. The bandwidth requirement of each

virtual link is set randomly between 1 and 3 units.

For evaluating the effectiveness and correctness

of our proposed algorithm, we have implemented

three algorithms for comparison purposes. The

algorithms compared in our simulation experiments

are shown in Table 1.

Table 1: Algorithms compared in our simulations.

Algorithms Descriptions

RVDCE

The algorithm proposed in this work for

provisioning reliable VDC across

multiple data centers.

RVDCE_R

The algorithm proposed in this work for

provisioning reliable VDC across

multiple data centers in which the VMs

are embedded on the available servers

with the highest reliability.

RVNE

Algorithm proposed in (Yu et al., 2014)

for reliable virtual network embedding

across multiple domains.

5.2 Simulation Experiment Results

Backbone bandwidth consumption: Figure 5 shows

the performance of average backbone bandwidth

consumption for provisioning VDC requests against

the number of VDC requests. The K in Fig. 5

represents the size of the partition. It can be seen that

the RVDCE algorithm leads to a lower average

backbone bandwidth consumption compared to that

of RVNE and RVDCE_R. This is because RVDCE

divides a VDC request into multiple partitions and

consumes the bandwidth of backbone network as few

as possible while mapping these partitions.

Furthermore, average backbone bandwidth

consumption of RVDCE decreases with the growth of

the value of K. Since larger partition size leads to a

lower bandwidth consumption between partitions.

Blocking ratio of VDC request: Figure 6 shows the

performance of blocking ratio under various reliability

requirements. In this set of simulations, the reliability

requirement rr varies from 0.92 to 0.98 with a step of

0.02.It is clear that when the number of VDCs is small,

the blocking ratios of the three compared algorithms

are very low. For example, in Figure 6(a), when the

number of VDC requests is smaller than 20, the

blocking ratios of these algorithm is zero. The blocking

ratio increases with the growth of the number of VDC

requests. The blocking ratio of RVDCE is lower than

that of the other two algorithms under different

reliability requirements. It is due to fact that RVDCE

algorithm embeds VMs onto servers with as lower

reliability as possible while satisfying the reliability

requirements of VDCs, for avoiding over-provisioning

and thus can admit more VDC request.

Cumulative CPU resource consumption: Figure 7

presents the results of total CPU resource

consumption for various number of VDC requests.

The cumulative CPU resource consumption increases

with the growth of the number of VDC requests. It

also can be seen from Figure 7 that the CPU resource

consumption of RVDCE is lower than that of RVNE.

This is because our RVDCE algorithm does not use

redundant resources as the backups for satisfying the

VDC reliability. Furthermore, our proposed

algorithm selects the servers to host the VMs

according to the dependencies between VMs, thus

avoiding over-provisioning and results in a lower

resource consumption.

0 100 200 300 400 500 600 700 800 900 1000

Backbone Bindwidth Consumption

Number of VDC Requests

RVDCE, K=1 RVNE, K=1

RVDCE_R, K=1

0 100 200 300 400 500 600 700 800 900 1000

Backbone Bindwidth Consumption

Number of VDC Requests

RVDCE, K=5 RVNE, K=5

RVDCE_R, K=5

0 100 200 300 400 500 600 700 800 900 1000

Backbone Bindwidth Consumption

Number of VDC Re

uests

RVDCE, K=11 RVNE, K=11

RVDCE_R, K=11

Figure 5: Average backbone bandwidth consumption.

IoTBD 2016 - International Conference on Internet of Things and Big Data

202

500 1000 1500 2000 2500 3000 3500 4000

-0.1

0.0

0.1

0.2

0.3

0.4

0.5

0.6

Blocking Ratio

Number of VDC Requests

RVDCE, rr=0.92

RVNE, rr=0.92

RVDCE_R, rr=0.92

500 1000 1500 2000 2500 3000 3500 4000

-0.1

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Blocking Ratio

Number of VDC Requests

RVDCE, rr=0.96

RVNE, rr=0.96

RVDCE_R, rr=0.96

0 500 1000 1500 2000 2500 3000 3500 400

0.0

0.2

0.4

0.6

0.8

Blocking Ratio

Number of VDC Requests

RVDCE, rr=0.98

RVNE, rr=0.98

RVDCE_R, rr=0.98

Figure 6: Blocking ratios under different reliability requirements.

0 1020304050

2000

4000

6000

8000

10000

CPU Resource Consumption

Number of the VDCs

RVDCE

RVNE

Figure 7: Total CPU resource consumption.

6 CONCLUSION

In this paper, we have studied the problem of reliable

VDC embedding across multiple infrastructures and

proposed a heuristic algorithm for solving this

problem. The aim of our research is to minimize the

total bandwidth consumption in backbone network

for provisioning a VDC request, while satisfying its

reliability requirement. Our algorithm makes a trade-

off between backbone bandwidth consumption and

reliability. Simulation results show that the proposed

algorithm significantly reduced the resource

consumption and blocking ratio than the existing

approach.

ACKNOWLEDGEMENTS

This research was partially supported by the National

Grand Fundamental Research 973 Program of China

under Grant (2013CB329103), Natural Science

Foundation of China grant (61271171, 61571098),

China Postdoctoral Science Foundation

(2015M570778), Fundamental Research Funds for

the Central Universities (ZYGX2013J002),

Guangdong Science and Technology Project

(2012B090400031, 2012B090500003,

2012B091000163), and National Development and

Reform Commission Project.

REFERENCES

Armbrust, M., Fox, A., Griffith, R., et al, 2010. A view of

cloud computing. Communications of the ACM.

Bari, M., Boutaba, R., Esteves, R., Granville, L., Podlesny,

M., Rabbani, M., Zhang, Q., and Zhani, M., 2013. Data

center network virtualization: A survey. IEEE

Communications Surveys and Tutorials.

Amokrane, A., Zhani, M., Langar, R., Boutaba, R., Pujolle,

G., 2013. Greenhead: Virtual Data Center Embedding

Across Distributed Infrastructures. IEEE Transactions

on Cloud Computing.

Zhang, Q., Zhani, M., Jabri, M., Boutaba, R., 2014. Venice:

Reliable Virtual Data Center Embedding in Clouds.

IEEE INFOCOM.

Guo, C., Lu, G., Wang, H., Yang, S., Kong, C., Sun, P., Wu,

W., and Zhang, Y., 2010. SecondNet: a data center

network virtualization architecture with bandwidth

guarantees. ACM CoNEXT.

Zhani, M. F., Zhang, Q., Simon, G., and Boutaba, R., 2013.

VDC planner: Dynamic migration-aware virtual data

center embedding for clouds. IFIP/IEEE IM.

Yu, H., Wen, T., Di, H., Anand, V., Li, L., 2014. Cost

efficient virtual network embedding across multiple

domains with joint intra-domain and inter-domain

embedding. Optical Switching and Networking.

Di, H., Anand, V., Yu, H., Li, L., Lianand D., Sun, G., 2014.

Design of Reliable Virtual Infrastructure Using Local

Protection. IEEE International Conference on

Computing, Networking and Communications.

Fischer, A., Botero, J., Beck, M., Meer, H., Hesselbach, X.,

2013. Virtual Network Embedding: A Survey. IEEE

Communications Surveys &Tutorials.

Greenberg, A., Hamilton, J., Maltz, D., and Patel, P., 2008.

The cost of a cloud: research problems in data center

networks. ACM SIGCOMM Computer Communication

Review.

Leiserson, C., 1985. Fat-Trees: Universal Networks for

Hardware Efficient Supercomputing. IEEE

Transactions on Computers.

GT-ITM. http://www.cc.gatech.edu/projects/gtitm/

Reliable Virtual Data Center Embedding Across Multiple Data Centers

203