MULTI-RING STRUCTURED OVERLAY NETWORK FOR

THE INTER-CLOUD COMPUTING ENVIRONMENT

Sumeth Lerthirunwong, Hitoshi Sato

Tokyo Institute of Technology, 2-12-1 Ookayama, Meguro-ku, Tokyo, Japan

Satoshi Matsuoka

Tokyo Institute of Technology, 2-12-1 Ookayama, Meguro-ku, Tokyo, Japan

National Institute of Informatics, 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo, Japan

Keywords: Cloud Computing, Inter-Cloud, Federated Cloud, Overlay Network, Structured Network, Churn,

Consistency of Global Information.

Abstract: Although the Inter-Cloud environment enables new possibilities for several data-intensive e-Sciences

applications, some challenging issues such as dynamic change of computing resources and management

complexity of large-scale overlay network remain. The structured peer-to-peer overlay network approach is

hereby adapted onto the Inter-Cloud environment to manage the consistency of the global information

among clouds and enhance the churn resistance. We propose a multi-ring structured overlay network for the

Inter-Cloud environment, in which nodes are organized into one sub-ring and all sub-rings are then

composed to form a large single ring structured network. The global information is managed within the sub-

ring so that the cost and the complexity for managing the global information can be reduced significantly.

We evaluate the efficiency of our proposed methodology by using overlay network simulator and compare

the results with other existing overlay networks. Moreover, several scenarios are further analyzed to show

the effectiveness in real-world cases. The results indicate that, with our approach, the management cost for

global information is reduced while the stability of the overlay network under churn can be maintained. The

average reachability of network under churn is more than 89.2% which is better than other structured

networks and by considering locality of node, the number of the distant messages also decreases by 58.4%

with respect to the overall messages.

1 INTRODUCTION

The Inter-Cloud environment, which is a federated

cloud environment where many clouds including

both public and private clouds are connected

together, is a possible platform for many e-Science

applications, since such environment can provide

much larger amounts of computational power and

storage resources. Such an environment brings

different cloud flavors and internal cloud resources

so that users are required to select computing

resources on demand in order to suite for particular

application workloads while considering budget,

urgency, and cost of computation, data movement

and consumption, etc. Therefore, computing

resources in this kind of environment may vary

dynamically, since users try to optimize the

computing resources according to the demands of

users and, as a result, the resources may join and

leave to the Inter-Cloud environment frequently. The

problem is that how can the environment maintain

the consistency of global information such as

availability of overall resources and running

applications in a highly dynamically changing cloud

environment, while keeping high availability of the

global information and minimizing resource

consumption. Furthermore, automatic resource

management is also a crucial research topic for these

large-scale and dynamic systems.

Our goal is to reduce cost and complexity for

managing the global information among clouds and

automate the system so that it can adjust itself

adaptively under churn situation when computing

nodes enter or leave to a federated cloud

Lerthirunwong S., Sato H. and Matsuoka S..

MULTI-RING STRUCTURED OVERLAY NETWORK FOR THE INTER-CLOUD COMPUTING ENVIRONMENT.

DOI: 10.5220/0003388400050014

In Proceedings of the 1st International Conference on Cloud Computing and Services Science (CLOSER-2011), pages 5-14

ISBN: 978-989-8425-52-2

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

environment at once. To that end, the structured

peer-to-peer overlay network approach is hereby

adapted onto the Inter-Cloud environment to manage

the consistency of the global information among

clouds and enhance the resistance to a large churn.

We propose a multi-ring structured overlay network

for the Inter-Cloud environment, in which nodes are

organized into one sub-ring and all sub-rings are

then composed to form a large single ring structured

network. The locality of nodes is considered when a

node decides to join into the sub-ring so that the

number of distant messages which are the messages

generated by the distant nodes can be decreased. The

global information is managed within the sub-ring so

that the cost and the complexity for managing the

global information can be reduced significantly.

Note that the cost for managing the global

information in our approach is a number of

messages used for requesting and updating data.

To evaluate the effectiveness of our proposed

technique, we analyze the stability of our network

and the efficiency of our technique in managing the

global information across the Inter-Cloud. We

implemented the overlay network simulator to

analyze reachability of the network, consistency and

availability of the global information. The results are

compared with other existing overlay networks such

as ring structured network, tree structured network

and unstructured network. We also applied several

scenarios to show the effectiveness of our algorithm

in real-world cases, for example, the scenario when

a large group of nodes decide to leave network at

once when they are not in-use. The results indicate

that, with our approach, the management cost for

global information is reduced while the stability of

the overlay network under churn can be maintained.

The reachability of network under churn is 8.2%

better than ring structured network and 23.1% better

than tree structured network by average. The number

of distant messages also decreases by 58.4% with

respect to the overall messages when considering

node locality.

2 RELATED WORK

Overlay network is a logical network that is built on

top of another network such as physical network or

IP network. Overlay network has been widely used

in distributed system especially the peer-to-peer

network. Overlay network can be categorized into

either unstructured network or structured network.

The unstructured network has minimal constraints

on the network topology and content distribution, so

it is suitable in highly dynamic and often ad-hoc

peer-to-peer environments. Most of popular peer-to-

peer applications nowadays operate on unstructured

networks, e.g., Gnutella (http://www.gnutella.com/)

and Freenet (http://free netproject.org/). In these

networks, peers connect to other peers in an ad-hoc

fashion. The location of the resources is not

controlled by a single centralized entity in the

system and there are no guarantees of query

solution. CLON (Matos et al, 2009) is the example

of the unstructured overlay network for cloud

environment. The authors aim to approximate the

structure of the physical network, while ensuring the

connectivity properties desirable to ensure reliable

dissemination of the gossip-based protocol. The

algorithm is based on SCAMP (Ganesh et al, 2003)

but also considered the locality of joining node so

that the number of messages traversing the long-

distant links can be significantly reduced.

On the other hand, the structured network has a

strict control on an underlying network structure, a

content publication strategy and query routing.

While the structured network is suitable for fast, low

cost resource lookup mechanism, it has a high

management cost for maintaining the topology of the

network. The examples of these networks are Chord

(Stoica et al, 2001), CAN (Ratnasamy et al, 2001),

and Tapesty (Zhao et al, 2001). All these networks

utilize the distributed hash table (DHT) for fast and

accurate resource lookup but have different network

topologies. In the case of cloud environment, the

network tends to be more reliable than peer-to-peer

network because nodes in the cloud environment are

usually part of the large data-center which has rich

network utility, so churn is unlikely to occur

frequently. For that reason, the structured network is

more suitable for cloud environment

Other interesting overlay network that related to

our proposed technique is ChordSPSA (Merz et al,

2009). ChordSPSA is a Chord-enhanced self-

organizing super-peer overlay which is designed to

suit the communication requirements of a peer-to-

peer desktop grid system. This overlay network

combines favourable properties of Chord rings and

fully meshed super-peer networks. The result from

the simulation shows that the algorithm can reduce

average message hop count in half compared to pure

Chord. Unfortunately, this algorithm is not designed

for cloud environment, so applied this algorithm

onto cloud may reduce the efficiency of algorithm.

There are also several works which focus on the

federated cloud. Aneka-Federation (Ranjan et al,

2009) is a decentralized and distributed system that

combines enterprise Clouds, overlay networking,

CLOSER 2011 - International Conference on Cloud Computing and Services Science

and structured peer-to-peer techniques to create

scalable wide-area networking of compute nodes for

high-throughput computing. The system components

are self-organized based on a structured peer-to-peer

routing methodology and thus make the system more

flexible, efficient, and scalable. However, the

authors do not focus on overlay algorithm but rather

explain the detail of each component in the system

instead.

3 BACKGROUND ON

INTER-CLOUD

3.1 Motivation behind Inter-Cloud

Although the cloud computing already has provide a

lot of capability for flexible, by demand computing

resources, it limited by some drawback which leads

to an emerging of the Inter-Cloud. First, one set of

cloud services is not going to be able to serve all the

needs of a customer and the geographical location of

cloud may limit ability of service to meet user QoS

because of high data transfer cost. Moreover, user

may want to use multiple clouds for different

applications to match business needs and allocate

different elements of an application to different

environments, whether internal or external and may

also want to move an application to meet

requirements at different stages in its life-cycle,

whether move between public clouds or move back

to the data-center.

To overcome such limitation of cloud, the Inter-

Cloud environment is developed to bring different

cloud flavours and internal cloud resources so that

users are required to select computing resources on

demand in order to suite for particular application

workloads while considering budget, urgency, and

cost of data movement and consumption, etc. Such

environment is a possible platform for many e-

Science applications. Nowadays, there are several

projects which already implemented the Inter-Cloud

such as CloudSwitch (http://www.cloudswitch.

com/), and Cisco/EMC Inter-Cloud project

(http://cloudven tures.sys-con.com/). Examples of

projects which built for the Inter-Cloud environment

are IETF’s Locator/ID Separation Protocol

(http://tool s.ietf.org/wg/lisp/), DMTF's Open

Virtualization Format (http:// www.dmtf.org/

standards/ovf/) and Google Code (http://code.

google.com/).

3.2 Challenges

Despite the potential of the Inter-Cloud environment

mentioned earlier, we also summarize some

challenging issues which it is still remained.

z How to manage the large-scale highly

dynamically changing environment which

may consists of millions nodes.

z How can the environment maintain the

consistency of global information such as

availability of overall resources and running

applications in this environment, while

keeping high availability of the global

information and minimizing resource

consumption.

Figure 1: Inter-Cloud Environment.

For example, in a data-intensive distributed software

framework like Hadoop (White, 2009), the job can

be effectively scheduled and allocated to the most

suitable node. As the result, the traffic cost to move

the job data can be reduced efficiently. In Hadoop

system, a JobTracker knows which node contains

the desired resources and a location of nearly

machines by using heartbeat mechanism to monitor

TaskTracker and using a distributed filesystem such

as HDFS (Borthakur, 2007) to share the job files. In

case of the Inter-Cloud environment where multiple,

different clouds are connected together, some clouds

may not provide the data of node in the distributed

filesystem and TaskTracker may be located in a

different cloud. Thus, due to lagging the sufficient

information, it is impossible for Job Tracker to

effectively schedule the jobs across the dynamically

changing clouds.

3.3 Requirement

To cope with the mentioned challenges, the system

should be able to meet these requirements.

MULTI-RING STRUCTURED OVERLAY NETWORK FOR THE INTER-CLOUD COMPUTING ENVIRONMENT

z The system should be automated, de-

centralized, load-balanced, self-organized and

should also have a low maintenance cost to be

able to deal with dynamic changing of

resources and for scalability

z Node locality should be concerned because of

high cost of distant messages in wide-area

network.

z The system should be able to maintain

consistency of the global information, so

sending or broadcasting messages should have

minimal cost

Moreover, there is the difference that needs to be

considered when designing the overlay for the Inter-

Cloud. The Inter-Cloud environments have more

reliable network connection than normal large-scale

peer-to-peer network. Churn in this environment

occurs when computing nodes has not been utilized

and leave the network. In contrast, in peer-to-peer

network, churn might occurs because network

failures or user shutdown the system. So the

structured overlay network approach is more

suitable for the Inter-Cloud environment.

4 MULTI-RING STRUCTURED

NETWORK

4.1 Overview

When considering all the requirements mentioned

earlier, the structured peer-to-peer overlay network

approach is hereby adapted onto the Inter-Cloud

environment to manage the consistency of the global

information among clouds and enhance the churn

resistance. Moreover, structured peer-to-peer

overlay network approach is known for its de-

centralized and self-organized manner which will

benefit the scalability of our network. We proposed

a new multi-ring structured overlay network.

Basically, our network is a composition of ring

structured network as can be recognized as the group

of ring network, called sub-ring, connected together

to form a large single ring structured network as

depicted in Figure 2. The blue circle represents a

peer and the pink circle represents a super-peer. The

function of peer and super-peer will be described in

later section. Note that the reason for adopting the

ring geometry is because of its simplicity, flexibility

and resilience, especially when compared to other

structured networks likes Pastry, CAN, Tapestry

(Gunmadi et al, 2004). The main features of our

network are:

z The load of the global information

management is divided among sub-ring and

managed within the sub-ring which leads to

less management cost and complexity

z User can choose area size when gathering the

resource information, and thus eliminate

dispensable messages

z By considering node locality when new node

joins the system, the number of the distant

messages can be decreased

z The ring structured overlay requires relatively

low cost to maintain the topology (Gunmadi et

al, 2004) which make it more scalable and

more resistance to churn

Figure 2: The multi-ring structured network.

To make our approach easier to understand, let take

a look back to the Hadoop example mentioned in

previous section. In case of our approach, super-peer

can function as a Job Tracker and keep track of all

the TaskTracker in the sub-ring. Super-peer can also

contact and ask information from JobTracker in

other sub-rings to schedule most suitable resources

for specific jobs. In this way, not only the global

information, such as resources location, can be

managed efficiently across the clouds, the work load

of the JobTracker can also be lighten and shared

between sub-rings. Note that our sub-ring does not

necessary represent one cloud, the sub-ring rather

represents group of nodes that is located within the

same geographical area. One cloud can be a subset

of a sub-ring or can be divided into multiple sub-

rings depending on the size of cloud.

4.2 Building the Overlay

The overlay network starts from a single node, other

nodes will join the system later to form the overlay

network. When a node wants to join the system, it

has to contact one of member of the system. That

CLOSER 2011 - International Conference on Cloud Computing and Services Science

member forwards the request to a super-peer which

locates within the same sub-ring. Sequentially, the

super-peer forwards this request to another super-

peer in a sub-ring which a geographically location is

closest to the joining node. The details of node’s

locality will be discussed later. Finally, the node

joins in that sub-ring. The sub-ring will be divided

into two sub-rings if member of that sub-ring

exceeds a threshold. The node which already

established itself in the sub-ring will get a local view

from peer in the same sub-ring. A local view

consists of a joining point and super-peers location.

It also has to deal with a replica which will be

explained later.

If node wants to leave the system, it has to

inform super-peers and adjacent peers in sub-ring

before leaving the system. The node is given two

choices of leaving pattern which are:

z If node temporary leaves the system,

temporary link between leaving node’s

neighbours is created to bypass the node. In

this case, links and replicas are kept in

unavailable state so that when the node rejoins

the system, the node can join at the point

where it leaves and get its replicas back from

its neighbours.

z If node permanently leaves the system, links

and replications will be erased. Sub-ring is

merged if number of node in sub-ring is less

than threshold. Node has to do the join

process all over again if it wants to rejoin the

system. When considering all the

requirements mentioned earlier, the structured

peer-to-peer overlay network approach is

hereby adapted onto the Inter-Cloud

environment to manage the consistency of the

global information among clouds and enhance

the churn resistance. Moreover, structured

peer-to-peer

Node in the Inter-Cloud environment may locate in

different cloud which may be as far as different

continent. The cost for sending message between

these two nodes is very high considering delay. The

communication between these two nodes is likely to

occur frequently if these nodes are neighbours or are

within the same sub-ring. In our approach, we

consider node’s locality before let the node joins

into the sub-ring and thus, the distant message can

be reduced. There are many approaches to compare

the locality of node (Matos et al, 2009), (Balbosa et

al, 2004). In case of our network, we use message’s

delay time between nodes to measure their locality.

If the delay is less than threshold, we assume that

these two nodes located within the same locality.

4.3 Choosing Super-peers

We adopt the super-peer approach in our overlay

network to manage the global information. In our

multi-ring network, each sub-ring consists of

multiple peers and super-peers. Normal peer can

only contact with peers within same sub-ring,

whereas super-peer can contact with super-peers in

adjacent sub-rings.

Super-peer manages all shared-information

within the sub-ring by providing or gathering data

when needed. More details on how super-peer

managing the global information will be discuss

later in next section. Super-peer also functions as a

hub for peers to connect to adjacent sub-ring.

Number of super-peer in sub-ring is a portion of

number of all peers in sub-ring (Mers et al, 2008)

In our algorithm, super-peer is randomly chosen

and so it can be any node in the system, a computing

node or a front-end node. But in the nature of the

Inter-Cloud, super-peer’s duty should be assigned to

front-end node of the cloud because super-peer has

to function as a hub to connect the sub-ring with

other sub-rings. Before leaving the system, super-

peer has to promote one peer to act as super-peer

and send all the sub-ring information to that

promoted peer.

4.4 Managing Global Information

In the Inter-Cloud environment where there is a

wide variety of accessible computing resources and

services for user, the available of global information

is very essential for the system to efficiently utilize

all the resources. Maintaining the consistency of

global information such as availability of overall

resources and running applications in a highly

dynamically changing cloud environment, while

keeping high availability of the global information

and minimizing resource consumption is very

challenging task.

The simplest way to make global information

available to all nodes in the system is to have every

node maintain information of all the system but the

cost for managing the information will be

unacceptable. In the Inter-Cloud scenario, nodes

tend to be aggregated in sub-nets inside a data-center

or in multiple data-centers, which are connected by

costlier, long-distant links. We take advantage from

this nature of the Inter-Cloud and divide our systems

into multiple sub-rings. Each sub-ring represents a

group of nodes which aggregated in sub-nets inside

a data-center or within the same local area. In our

approach, the cost of managing the information is

MULTI-RING STRUCTURED OVERLAY NETWORK FOR THE INTER-CLOUD COMPUTING ENVIRONMENT

distributed to each sub-ring and the information is

managed within the sub-ring. As the result of this,

the cost and the complexity for managing the global

information can be reduced significantly.

Furthermore, to enhance the availability of the

global information, we apply the replica strategy to

our overlay network. Each peer has to make a

replica of its resource information and send the

replica to adjacent peers and also every super-peer in

the sub-ring. Replica is proactively updated from

replica’s owner for data consistency which means

that each peer has to inform adjacent peers and

super-peers if there is any update in the resource

information.

Whenever a peer needs the information, the peer

sends request to super-peer within the same sub-

ring. If aggregated global information is needed,

super-peer will broadcast the request to super-peers

located in another sub-rings. Generally user likely to

requests the global information to obtain resource

information of overall system. But in some

scenarios, a job may limit by geographically location

of the data and thus the resource locate faraway may

not be able to meet user’s QoS because of costlier,

long-distant links. For example, user in Japan may

want to use only resources which located within

Japan or Southeast Asia because of a high data

transfer cost. In this case, user can choose to get the

information within the sub-ring or adjacent sub-rings

only since our sub-ring already represents the

approximation of the geographical location of node.

4.5 Dealing with Fault Tolerant

Our overlay network adopts heartbeat mechanism to

detect node failure. Each node in the sub-ring sends

heartbeat message to its neighbour periodically to

monitor the existence of its neighbour. Recovery

process is initiated when any failure is detected.

First, a failed node is treated as a temporary leaved

node and a time out counter is set. If the time out is

reached, the failed node will be treated as a

permanent leaved node and all of its temporary data

will be deleted from the system. This algorithm

provides a possibility for a node to re-join the

system smoothly with less cost if the node fails

unexpectedly.

5 EVAULATION

To evaluate the effectiveness of our proposed

technique, we evaluate the reachability of the

network, the consistency and the availability of the

global information. We implemented the overlay

network simulator using C# which the details is

listed in Table 1. Note that in our simulation, a

number of super-peers per sub-ring are fixed. We

compared the results with other existing overlay

networks such as ring structured network, tree

structured network and unstructured network. The

reason that we also compare the results with the

unstructured network is to analyze the pros and cons

of our decision to choose the structured overlay. In

the first part, we investigate the stability of network

by observing a percentage of reachable nodes when

the network facing churn. Next we investigate the

consistency and the availability of the global

information when the churn rate is varied and the

information is updated frequently.

To analyze data manage cost, we evaluate the

cost of updating the information and requesting the

information. Finally, we investigate the percentage

of distant message which node received during a

period of time to see how our technique can utilize

the locality of nodes.

Table 1: Simulation Details.

Type of Overlay Multi-ring, Ring

Tree, Unstructured

System Size 1000 node

Churn Rate 10-80% per hour

Sub-ring Size 100, 200, 400 node

Number of super-peer

per sub-ring

1,3,5,10 node per sub-ring

5.1 Stability of Network

The stability of overlay network is very important

especially the large scale dynamically changing

network likes the Inter-Cloud environment. The

overlay should be able to repair itself when facing

churn to maintain the topology and the reachability

of the network. We investigate the stability of

network by observing a percentage of reachable

nodes when the network facing churn. In this

simulation, we set churn rate at 10% per hour which

means only half a day, all the nodes will be almost

replaced with new nodes. Although in the real cloud

environment, normally the churn rate may not be as

high as our simulation, we want to simulate the

worst case scenario to evaluate effectiveness of our

approach.

Figure 3 shows the result from our simulation.

We compared our multi-ring structured network

with ring structured network, tree structured network

and unstructured network. The graph indicated that

our network can maintain the reachability as good as

the unstructured network but far better than ring and

CLOSER 2011 - International Conference on Cloud Computing and Services Science

tree structured network. The reachability is 8.2%

better than ring structured network and 23.1% better

than tree structured network by average. Note that

the unstructured network does not have any

constraint on network topology so it has more

resistance to churn and better reachablity than those

structured networks. However, when considering a

cost for broadcasting messages which is essential for

managing the global information, the unstructured

network is not suitable approach since the messages

will be flooded over the network.

Figure 3: Reachability of Network.

Figure 4: Availability of replicas.

5.2 Availability of Global Information

In this section, we investigate the effective of our

approach on the availability of the global

information. According to our approach mentioned

earlier in Section 4, the information of each node

such as available resources or running tasks is

replicated to neighbour nodes and to super-peers.

Therefore, the number of super-peer is a key factor

of the availability of global information since the

number of neighbour node is fixed at two.

In our first experiment, we set a number of peers

to 1000, a number of super -peer per sub-ring to 3

and also set the same total number of replica to other

overlay networks for the fair comparison. Note that

for other overlay networks, we randomly choose the

nodes which are responsible for keeping the replicas.

The result from Figure 4 is matched with the result

from the previous section which indicates that our

technique can maintain the availability of replicas

better than other structured network.

Next, we vary the number of super-peer in our

approach to further analyze the effect of replicas on

both the availability and consistency of the data

under churn.

Figure 5: Effect of super-peer on availability of replicas.

Figure 6: Effect of super-peer on consistency of replicas.

Figure 5 shows that as number of super peer

increases, the availability of global resources can be

properly maintained at higher churn rate. However,

as in Figure 6, the consistency of global resources is

reduced when number of super peer increases since

it requires higher cost to keep the information

updated. The availability of global information is

likely to be a trade-off with a consistency.

MULTI-RING STRUCTURED OVERLAY NETWORK FOR THE INTER-CLOUD COMPUTING ENVIRONMENT

5.3 Data Management Cost

Data management cost is a cost for requesting or

updating the information. In case of requesting the

data, the request message is forwarded to super-peer

and then to another sub-rings. We measure hop-

count that is needed for the node to get the request

information, while in case of updating the data, node

needs to send the update of its information to entire

information replica. We also measure hop-count

which is needed for the all replica to be updated. We

set the number of super-peers per sub-ring to 3 and

the sub-ring size to 100. Note that for other overlay

networks, we randomly choose the nodes which are

responsible for keeping the replicas as done in the

previous experiment.

Figure 7: Request cost.

Figure 8: Update cost.

Figure 7 and Figure 8 shows that our sub-ring

approach can reduce number of message’s hop-

count significantly in both request and update case

compared to other overlay network, especially the

unstructured network which cost is very high

because the messages are flooded over the entire

network. These results also show that by dividing

the network into multiple sub-rings, the entire global

management can also be distributed into each sub-

ring effectively.

5.4 Effect of Locality

This section, we evaluate the effect of using node

locality by measuring the percentage of distant

message which node received during a period of

time to see how efficient our technique can utilize

the locality of nodes. The distant message is defined

as a message which a delay time is more than our

threshold. If the percentage of distant messages

decreases, the average message delay also decreases.

We set a number of super-peers per sub-ring to 3, a

number of peers to 1000 and a sub-ring size to 100.

Figure 9: Percentage of Distant messages.

The result from Figure 9 shows that our approach

can reduced the percentage of distant message by

47.2% compared to the ring structured network and

by 58.4% compared to the multi-ring structured

network without considering node locality. This

result indicates that both node locality and sub-ring

approach contribute to the reduction of a number of

distant messages.

5.5 Summary

The results from our simulation experiments

indicated that, with our approach, the management

cost for global information is reduced while the

stability of the overlay network under different

churns can be maintained significantly compared to

other structured networks. In case of the

unstructured network, although the stability of

network and availability of replica is better than our

approach, the managing cost for the global

information is unacceptable since the messages are

flooded over the entire network.

CLOSER 2011 - International Conference on Cloud Computing and Services Science

6 SCENARIO STUDY

In this section, we simulate the scenario which could

happen in the real Inter-Cloud environment

situation. Since the Inter-Cloud environment

provides lots of computing resources on demand.

User tends to optimize the computing resources

according to their demands; as a result, the resources

may join and leave to the Inter-Cloud environment

frequently. However, the scenario is not as same as a

normal churn in the peer-to-peer network where

resources randomly and uniformly leave the

network. In case of the Inter-Cloud network, a group

of resources in the same area such as within the

same data-center tends to leaves the network

together when they are not in-use. The churn in this

scenario is still large as in the peer-to-peer but not

uniform. We call this type of churn, an area churn.

The simulation is set as a group of adjacent

nodes periodically leave of system together. We also

vary the group size to further investigate its effect

when the group size exceeds the sub-ring size. We

evaluate the reachability of the network and the

availability of the information to show a capability

of our technique facing these scenarios. We set a

number of peers to 1000, a number of super-peers

per sub-ring to 3, the sub-ring size to 100. Note that

for other overlay networks, we randomly choose the

nodes which are responsible for keeping the replicas

as done in the previous experiment.

Figure 10: Reachability of network under area churn.

The result from Figure 10 shows that our

approach can effectively maintain reachability of the

network compared to other structured overlay

networks. While the result from Figure 11 shows

that our approach can also maintain the availability

of the replicas even when facing area churn situation.

More than 79% of replicas are available when area

churn size is 160 nodes, which is larger than the size

of the sub-ring.

Figure 11: Availability of replica under area churn.

7 CONCLUSIONS

In this paper, we presented the multi-ring structured

overlay network for the Inter-Cloud computing

environment. The structured peer-to-peer overlay

network approach is hereby adapted onto the Inter-

Cloud environment to manage the consistency of the

global information among clouds and enhance the

churn resistance. We evaluated the efficiency of our

proposed methodology by using overlay network

simulator with several scenarios are further analyzed

to show the effectiveness in real-world cases. The

results indicated that, with our approach, the

management cost for global information is reduced

while the stability of the overlay network under

different churns can be maintained. The average

reachability of network under churn is more than

89.2% which is better than other structured networks

and by considering locality of node, the number of

the distant messages also decreases by 58.4% with

respect to the overall messages.

8 FUTURE WORK

We plan to do more detailed simulations, with more

comparisons and scenarios. In the next step, we

intend to implement our technique onto real cloud

environments such as Windows Azure and Amazon

EC2. We’re also interested in exploring the

possibility to collaborate our system with the super-

computer system. The different nature of the cloud

environment and the super-computer will bring more

challenges to our approach.

MULTI-RING STRUCTURED OVERLAY NETWORK FOR THE INTER-CLOUD COMPUTING ENVIRONMENT

ACKNOWLEDGEMENTS

This research is supported in part by the MEXT

Grant-in-Aid for Scientific Research on Priority

Areas 18049028.

REFERENCES

Abhilash Gummadi and Jong P. Yoon,” Modeling Group

Trust for Peer-to-Peer Access Control”, in Proceeding

DEXA '04, 2004

Abhishek Chandra and Jon Weissman, “Nebulas: using

distributed voluntary resources to build clouds.”, in

Proceedings of the 2009 conference on Hot topics in

Cloud Computing, pp 2-2, 009

Ayalvadi J. Ganesh, Anne-Marie Kermarrec, and Laurent

Massoulie, “Peer-to-Peer Membership Management

for Gossip-Based Protocols”, IEEE Transactions on

Computers, vol 52, no. 2, 2003

Ben Y. Zhao, John Kubiatowicz, and Anthony D. Joseph,

“Tapestry: An Infrastructure for Fault-tolerant Wide-

area Location and Routing”, Report No. UCB/CSD-

01-1141, April 2001

Dhruba Borthakur, “The Hadoop Distributed File System:

Architecture and Design”, the Apache Software

Foundation, 2007

Ion Stoica, Robert Morris, David Karger, Frans Ka~shoek,

and Hari Balakrishna, “Chord: A scalable peer-to-peer

lookup service for internet applications”, In

Proceedings of SIGCOMM, pp. 149-160, 2001

Kazuyuki Shudo, “Churn Tolerance Improvement

Techniques in an Algorithm-Neutral DHT”, in

Proceeding of AIMS 2009, LNCS 5637, pp. 42–55,

2009

M´ark Jelasity and Anne-Marie Kermarrec, “Ordered

Slicing of Very Large-Scale Overlay Networks”, in

Proceeding of IEEE P2P, pp. 117–124, 2006

Marcelo Werneck Barbosa, Melissa Morgado Costa,

Jussara M. Almeida and Virgflio A. F. Almeida,

“Using Locality of Reference to Improve Performance

of Peer-to-Peer Applications”, in Proceeding of

WOSP'04, 2004

Miguel Matos, António Sousa, José Pereira and Rui

Oliveira, “CLON: Overlay Network for Clouds”, in

Proceeding of WDDDM ’09, March 31, 2009

Peter Merz, Steffen Wolf, Dennis Schwerdel, and Matthias

Priebe, “A Self-Organizing Super-Peer Overlay with a

Chord Core for Desktop Grids”, in Proceeding of

IWSOS 2008, LNCS 5343, pp. 23–34, 2008

Rajiv Ranjan, Rajkumar Buyya, “Decentralized Overlay

for Federation of Enterprise Clouds”, Handbook of

Research on Scalable Computing Technologies, K.C.

Li, C. H. Hsu, L. T. Yang, J. Dongarra, and H. Zima,

ISBN: 978-1-60566-661-7, IGI Global, Hershey, PA,

USA, July 2009

Rajkumar Buyya1, Rajiv Ranjan, and Rodrigo N.

Calheiros, “InterCloud: Utility-Oriented Federation of

Cloud Computing Environments for Scaling of

Application Services”, in Proceeding of ICA3PP 2010,

Part I, LNCS 6081, pp. 13–31, 2010

Sylvia Ratnasamy, Paul Francis, Mark Handley, Richard

Karp and Scott Shenker, “A Scalable Content-

Addressable Network”, Proceedings of ACM

SIGCOMM, pp. 161-170, 2001

Tom White, “Hadoop: The Definitive Guide”, O’REILLY,

Yahoo Press, 2009

CLOSER 2011 - International Conference on Cloud Computing and Services Science