DYNAMIC COALITION IN AGENT AWARE ADHOC VIRTUAL

P2P INTERCONNECT GRID COMPUTING SYSTEM –

VIGRID

Avinash Shankar, Chattrakul Sombattheera, Aneesh Krishna, Aditya Ghose, Philip Ogunbona

Decision Systems Lab [DSL]

School of IT and Computer Science,Univerity of Wollongong, NSW-2522, Australia

Keywords: Dynamic Coalition, Grid, P2P, Agents, A

viGRID.

Abstract: Artificial Intelligence and High performance Grid computing s

ystems are two different fields of

technologies that have been much researched upon and are as old as the development of personal computers

and its related technologies. Technologies such as Agent based systems and the Semantic Grid have evolved

with the use of Artificially Intelligent techniques such as the Turing system of intelligence measurement.

Similarly Peer to Peer computing and Supercomputing Grids have evolved from distributed and middleware

clustering systems such as Condor. In this paper a new architectural schematic is proposed where

technologies such as Agents, P2P computing and Adhoc systems are incorporated in a Grid Computing

framework for the optimal Job processing and delivery to the end user transparently. Applying Dynamic

Coalition techniques in Agent based Grid computing systems has been a meagrely researched Area. The

proposed system A

viGrid tends to deploy a service oriented schematic that enables users to search for,

negotiate using agents, do remote Job Processing and use resources without the need for a resource

discovery model in place that is commonly used in current day high performance Grid systems.

1 INTRODUCTION

Applying Agent based Concepts to Grid Computing

brings a new era of Computing systems that learn

and act according to the needs of the end user. Here

we give a brief introduction to the various

techniques and technologies that will be used in

designing the A

viGRID System.

“Grid computing can be defined as a

decent

ralized, pervasive, transparent and simplistic

schematic which is governed by no central body that

uses different types of heterogeneous computing

platforms / technologies and their resources over a

decentralized interconnected network such as the

Internet.”(Source:http://www.huihoo.com/grid/grid_

computing_info_centre.htm, Last visited 5/01/2005)

The driving force behind Grid and High

perform

ance computing research seems to be a

desire to harness and share idle computing resources

across organizations world over. Grid technology

has gotten a lot of attention from both academic and

commercial environments. The academic

community is seeking better ways to tackle High

performance problems, and the commercial

industries interest’s lies in more e-client usage of

commodity hardware to reduce operational costs (or)

replace expensive specialized computers. Good

examples of High-performance Cluster or Grid

Computing systems would be the TeraGrid project

(Source: www.teragrid.org), UK E-Science project

(Source:www.escience-grid.org.uk), etc, that

performs in the measure of Teraflops in terms of

processing speeds and Terabytes of storage space.

Such vast resources should be made available and be

used to the fullest extent possible. The A

viGrid

System tries to provide such a Framework for

effective utilization of such systems to bring service

oriented Grid computing to the desktop user

irrespective of the location and topology. Peer-to-

Peer computing is the concept of sharing resources

in ways of a give and take policy. P2P computing is

a subset of High performance Grids or Cluster

computing. The technology was an offspring of

many different technologies and has become famous

after the success of High performance computing. A

peer is a single entity (or) node that shares its

resources with other peers (or) nodes directly (or)

indirectly connected to it. Early peer-to-peer systems

were primarily made to enable users to share, often

illegal, files easily and in public. Early systems

either had scalability problems or were not pure

peer-to-peer systems. Now there are more serious

170

Shankar A., Sombattheera C., Krishna A., Ghose A. and Ogunbona P. (2005).

DYNAMIC COALITION IN AGENT AWARE ADHOC VIRTUAL P2P INTERCONNECT GRID COMPUTING SYSTEM – A3PVIGRID.

In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 170-175

DOI: 10.5220/0002530701700175

 SciTePress

research and product development efforts on peer-

to-peer technologies. Sun’s JXTA (Source:

www.jxta.org) effort is one important arena today.

In the open source community peer-to-peer systems

also seems to have matured and issues on scalability

have been addressed in some systems, e.g. (Condor /

Condor - G, www.cs.wisc.edu/condor/ and

Bearshare, Kazza, Gnutella, www.zeropaid.com).

Peer-to-peer systems also have a basic goal of

utilizing distributed resources and providing services

to the participating nodes. Today services have

primarily been file sharing systems, but Sun’s JXTA

framework shows peer-to-peer computing is not

limited to just that. Especially Sun Microsystems

JXTA and P2P-JXTA (Source: www.openp2p.com)

show that peer-to-peer computing actually has many

similarities with Grid and agent based systems.

2 COALITIONS IN AGENT-BASED

GRIDS

A coalition, in the context of agent-based systems, is

usually defined to be a group of agents that come

together to solve a common task or achieve a

common objective. Coalition has its roots from

Game theory where players {agents} form groups

and plot a strategy for winning a Game. In general

with respect to Agent based Systems and Game

Theory, coalition formation occurs on the fly where

agents tend to form groups to achieve a common

goal such as Job processing or maximizing their

utility value. Here with respect to Grid computing

we en-route and define two new concepts called

Static coalition and dynamic coalition in agents

based grid systems. Two categories of coalitions are

of interest here: static and dynamic.

Static coalitions are typically formed on the basis

of more persistent common goals and tasks, and are

less likely to change from problem to problem.

Dynamic coalitions, on the other hand, are

groupings that are formed to address the needs of

a specific task or common objective. Once these

tasks are completed, or the common objectives

met, dynamic coalitions tend to disband, and re-

form in different ways. Here in A

viGRID we

use dynamic coalition formation techniques for

effective job processing and aggregation of

resources available. Our fundamental premise is

that coalition mechanisms add value in the

context of agent-based grids, for the following

reasons:

• Coalitions of peers can reduce the

communication overhead involved in executing

complex tasks and services which require the use

of multiple peers.

• Coalitions of peers can better enable better

matching between the requirements of tasks/services

and the infrastructure that is made available to

execute these. For instance, an appropriate coalition

formation mechanism can put together a collection

of peers with similar platforms and QoS

characteristics that are best suited for a given task.

• Coalition formation mechanisms can be used

to optimize complex trade-offs between the

objectives of maximizing the utility of the service

requester(s) and the service providers. For example a

service requestor could be maximizing its payoff for

the given task by being an intermediary service

provider that outsources its job to third party agents

thus maximizing his individual utility value.

• Coalition formation mechanisms can

economically increase system throughput as a

whole. After some negotiation among agents, tasks

will be allocated to appropriate coalitions who can

execute them with minimal costs and time. Thus

agents seem to be better off. A good example of this

would be the formation of coalition among agent in

a local Linux cluster where the maximum payoff is

achievable with minimal communication costs.

3 THE A

viGRID

ARCHITECTURE

The A

viGRID Architecture is primarily focused on

providing a Peer to Peer based Adhoc Multi Agent

Environment that enables users to remotely join the

viGRID system to search for new serial/parallel

programs and submit jobs for job processing. A very

good example would be P2P file-sharing systems

such as Bear share, Kazza, Gnutella etc that use

directory services to register the location of peers

along with the information of the list of files used

for sharing.

Figure 1: Peer to Peer A

viGrid Syste

DYNAMIC COALITION IN AGENT AWARE ADHOC VIRTUAL P2P INTERCONNECT GRID COMPUTING

SYSTEM – A3PVIGRID

171

3.1 Agent Based Peer Manager

[APM]

The Agent based Peer manager is an intelligent

agent that handles negotiating and registering of

services based on directory services. Although it

poses a threat as a centralized scheme, it’s primarily

used for discovery and communications between

agents and their respective peers based on a light

weight directory services model such as LDAP

(Source: Open LDAP directory service protocol -

Source: www.openldap.org). The primary role of the

APM is to register services of agents based on

commonality and self-interests factors and also help

in discovery and formation of Coalitions in agents

based on the commonality of their services rendered.

The APM also tends to act as a broker or

middleware for all agents associated with it. An

economic brokering system could be incorporated

based on the usage of the APM which renders as a

service to self-interested agents and agent based

systems in its locality.

3.2 Electing a RSD Leader

Let there be a set of agents {A, B...E} located in

nearby locations. We assume that these agents form

a Regional service domain [RSD] based on a set of

attributes that are adhered to by the agents and the

service requestor. That is all the agents available in a

common region are grouped by a leader called the

Regional leader. So each region can have a number

of RSD’s based on the how close agents are to each

other. Once the RSD is formed each agent votes for

a leader for representing the agents in a particular

region. Here the agents say {A, B, C, D, E} will bid

on the Job J1 along with A the elected leader by

forming new coalitions. The least loaded Coalition

having an optimal turn around time will get the job

based on the negotiated payoff value. R after

negotiations decides upon selecting the best

coalition for the job and offloads the job for job

processing. The primary use of the RSD leader is to

represent a set of closely-knit agents to minimize the

communication costs between the agents and the

APM.

4 COALITION FORMATION

METHODOLOGY

This architecture allows agents to dynamically form

coalitions in order to bid for, executing tasks. Firstly

the job description will be broadcast to all agents of

the RSD. Agents exchange information with respect

to their present ability to perform the tasks. The

information is composed of a set of attributes that

explains the current state of the agents. There is a

utility function that assigns a utility to the coalition.

The utility function takes the attributes of all the

coalition members and computes a utility that

indicates which coalitions are suitable for the tasks.

Agents then try to form coalitions based on this

indicative utility. Agents decide to join a coalition if

they are offered a payoff, which is not less than a

payoff threshold computed individually.

4.1 Ranking by Indicative Utility

The information is a set of attributes or properties

that affect the way jobs are to be processed by

agents and their respective peers. Attributes in

agent-based grids can be the load of machine, turn

around time of a process, latency, QoS factor,

bandwidth requirement, distance, etc. Each agent

maintains an attribute table, which contains all other

agents attributes it collects after the exchanging of

information. Based on the table shown in Figure 4

we can assume that each job is associated with an

attributed value with respect to the properties

associated by that attribute. Based on this we can

compute the utility value for each agent by. The

figure 4 indicates that Agents A, B and C have A1

…. A3 attributes which are needed for computing

the utility value of each potential coalition based on

the attributes of the job. Each agent can then

compute all the possible coalitions it may form.

These coalitions are ones that each agent would like

to propose to other agents. The agent uses each

coalition in this list as a proposal for forming

potential coalitions. It then computes the utility

value for each coalition. The utility value indicates

how well a coalition can perform a task. The utility

can be computed by applying a LESS or a MORE

value with respect to the properties of the attribute.

For example if we were computing the Load of

individual agents then its useful to have a less loaded

system than that of the jobs attribute value. So a

coalition formation will lead to an average of loads

to compute the average value of the potential

coalition and a LESS is specified for computing the

validity of the potential coalition formation between

say agents A and B. The agent then ranks coalitions

based on their utility values in a ranking table called

RTable in descending order. In the case that the

multiple coalitions yield the same utility value, then

smaller coalitions are preferred. Based on the

ranking table RTable the best coalition formation

strategy is deployed. The price for executing the job

successfully will be equally distributed among the

coalition members as their individual payoffs.

ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING

172

4.2 Coalition Formation Protocol and

Decision Making

Step 1: R is a remote node that needs to offload its

job say J1. R sends a Job Description message say

<ID, Job desc, T, A1, A2...An> to the APM for the

job J1. A set of attributes such as Latency, Time of

Completion, Payoff, etc is defined as A1, A2, A3

…An.

Step 2: The APM in turn sends a message to the

Regional service Domain [RSD] leader which in

turn advertises the job J1 to its local agents for

dynamic coalition formation.

Step 3: Let us take an agent A that is interested in

computing the job J1.

Step 4: To compute potentially good coalitions,

agent A sends messages to all other agents in the

RSD querying for the attributes as shown in Fig 4.

Step 5: Each agent uses an internal table to compute

the indicative utility and ranks the coalitions in

descending order in a ranking table RTable. Now

agents have the capability to decide which agents are

suitable for coalition formation.

Step 6: Each agent now computes individual payoff

for each coalition member in all coalitions by

distributing the price of the job equally.

Step 7: Each agent selects all the coalition members

from its top proposal based on its ranking table

RTable and then sends out a message <ID, coalition,

payoff vector> (ID is identity of the requestor while

payoff vector specifies the individual payoff for each

coalition member) to all agents in that proposal.

Step 8: Each receiving agent compares its individual

payoffs, Xi, in the proposed payoff vectors and

compares them to the payoff Xi* in its top individual

proposal. If

Xi >= Xi*

then agents reply with ACK, or else with a NO-ACK

signal.

Step 9: If the Head of the coalition receives ACK

signal from all the coalition members, it declares

that a coalition is formed and the task is bided for by

the new coalition that is formed.

Step 10: And if the requestor receives a NO-ACK, it

decreases its payoff Ui* in the top payoff vector by a

certain value, i.e. ∅ then increases the payoff Uj*

for all other coalition member by ∅ / (|S|-1), where

|S| is the size of the coalition. If the new Ui* is less

than the equivalent payoff in the second payoff

vector then delete the top proposal Repeat step 7.

Step 11: The newly formed Coalitions then bid for

the job J1 by negotiating with the originator R to

obtain the job J1. The negotiation can involve

comparison of attributes such as payoff value, goals

of agents, etc, to negotiate for J1.

Step 12: R then compares and contrasts the attributes

of the coalitions and then offloads to the best

negotiated coalition formation available for its job

namely J1.

Step 13: As soon the job J1 is allocated to a specific

coalition, all the other coalitions are dispersed to

form autonomous agents once more.

Step 14: The selected coalition then performs the

Job J1 and on completion sends a finished FACK

status to the originator R of job J1; it then receives

its payoff for the job and disperses the coalition to

join the host of autonomous agents in the RSD to

become autonomous once more.

Step 15: Continue back to Step 1 for originator of

new jobs and coalitions respectively.

Figure 2: Location Based Dynamic Coalition

DYNAMIC COALITION IN AGENT AWARE ADHOC VIRTUAL P2P INTERCONNECT GRID COMPUTING

SYSTEM – A3PVIGRID

173

5 EXAMPLE OF APPLYING

DYNAMIC COALITION TO THE

viGrid SYSTEM.

As shown in Fig 3, there is a remote node R with IP

address 203.54.6.35 and has a Job processing

requirement called J1. The APM stores information

about the locations of the regional agents and there

services. An example of an attribute can be the

turnaround time, Latency, Payoff, etc. When node R

wants to execute the job J1, it authenticates itself

with the APM by sending its IP address say

203.54.6.35; sends an advertisement about a remote

Job J1 that has a requirement for a service called

padder.exe to the APM. Based on the location of R

and the services requested, the APM passes the

request to the respective regional service domains

[RSD] leader A. Attributes are properties that are

adhered to by the agents based on self-interests. So

each agent is specifically interested in job processing

based on a set of special value attributes that plays

an important role in the way jobs are processed.

Attributes can be anything from resource

requirements, latency to trust issues with respect to

agents and the originator of the job. As shown in

Figure 4 we can see that A1…..A3 are the attributes

values of CPU %, Load % and Storage. Now each

agent based on the properties of its environment will

agree or are bound to satisfy certain attribute values.

As shown in the figure, let A, B and C be three

agents having different attribute values. Whenever

an agent satisfies an attribute value, the agent puts

its calculated value to compute utility value of the

potential coalition based on the available attributes

of its immediate environment. Attributes based on

the job description tend to form the basis for agents

to form coalitions. Our example denotes attribute

values to be properties which affect the way jobs are

being processed by agents and their respective peers.

Attributes with respect to agent based Grids can be

the Load of a machine {A1}, turnaround time of a

process {A2}, Latency {A3}, QOS factors {A4},

Bandwidth requirement {A5}, distance {A6}, etc.

As shown in Figure 4 let job J1 that originated from

remote node R have the following three attributes

that needs to be satisfied in order to do job

processing. Now we can see that A1…..A3 are the

attributes having CPU %, Load % and Storage

computed approximately for processing job J1. The

attributes become a requirement here for job

processing. Each agent say A, B, C compute their

individual values and the appropriate potential

coalitions by comparing which potential coalitions

formations tries to satisfy the attributes of the Job J1

to the fullest. As we can A, B and C first check to

see if they satisfy the requirements of the Job J1

independently before proposals are sent for

coalition formation. Then a set of possible coalitions

ure 3: A

namic Coalition in A3

viGrid Schematic

ure 4: Com

utin

Potential Coalitions

ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING

174

proposals are formulated and computed based on the

attributes available. Let us take one of the attributes

say Load % to examine how the computation works.

For every value computed we see if a LESS or

MORE value is required by that attribute. That is

Job J1 here needs the load of a system to be less than

or equal to 50% then we compute every possible

coalitions and do an average of the loads for all the

coalitions to see if they satisfy the load value say <=

50%. As shown all the possible coalitions are

calculated and compared to see which of the

coalitions are <= 50% loaded. Similarly we need

storage space of 6 megabytes, and that becomes a

MORE as we need to have a system which can offer

6 or more number of megabytes for storage. For

attributes such as load or CPU we have to calculate

the average by dividing the sum with the number of

agents forming the coalition. After calculating the

LESS and MORE values of all the possible

coalitions the agent needs to decide which coalition

formation satisfies all or most of the requirements of

the Job J1. As we can see the coalition formation of

ABC seems to match all the criteria’s of the Job J1

and the next best coalition that satisfies most of the

requirements is AC that is ranked in the agents table

as a potential coalition in case ABC cannot be

formed. Each coalition then selects a leader to

represent the coalition to optimize on

communications with the RSD and the originator R.

The leaders associated with every coalition formed

starts to contact R and bids for the Job J1. Hence we

have computed the utility value for potential

coalitions and the best coalition available to the

agent is used for bidding. The list of potential

coalitions after accepting the proposals will form

coalition to bid for the Job J1 and after negotiations

with respect to payoffs the agents will perform the

Job J1. Coalitions are dispersed once the selected

coalition is allocated for job processing and when

the Job is finished the Agents break the Coalition

formation and become autonomous as they join the

RSD again. Two or more Coalitions can also be

maintained by the remote peer R to achieve fault

tolerance with respect to job processing. Thus

dynamic coalition in agents tries to provide a good

solution to Agent based service oriented Grid

computing systems where best possible coalitions

are computed to provide optimal job processing

among agents.

6 LIMITATIONS

The primary limitations posed by dynamic coalition

in agents is that coalition formation might take more

time than expected due to the negotiations among

the agents. The other problem associated with agents

is the possibility of not bidding for a Job and doing

Job processing at all due to their autonomous

properties such as different goals or dissimilarities

observed in agents.

7 CONCLUSION

Applying Dynamic Coalition methodologies to

Multi Agent based High performance Grid

computing systems has lead to a new perspective to

the usage of intelligent agents in Grid computing

systems. The A

viGrid tries to provide solutions to

minimizing the usage of Resource discovery models

and process analyzers by utilizing directory services

such as the APM and the effective usage of dynamic

coalition schemes. Local coalitions are dynamically

formed by intelligent agents having a commonality

of goals based on a service oriented schematic.

REFERENCES

Andrew S. Tanenbaum and Robbert Van Renesse; ACM

Comput. Surv. 17, 4 (Dec. 1985), Distributed

Operating Systems, Pages 419 - 470.

Avinash Shankar, Daniel Saffioti, Ian Piper, Ashwin

Shankar, June 21st – 24th 2004, “Service Oriented

Web Based Meta Resource Sharing Platform - The

CBWeB

Portal “. The 2005 International Conference on Parallel

and Distributed Processing Techniques and

Applications, CSREA Press, Pages 744-749

Justin R.D. Dyson, Nathan E. Griffiths, Hélène N. Lim

Choi Keung, Stephen A. Jarvis, Graham R. Nudd,

Trust in Agents for Grid Computing.

Condor Cycle Stealing Technology – Condor / Condor -

G, www.cs.wisc.edu/condor/, Last accessed 6/3/2004

File Sharing applications - Bearshare, Kazza, Gnutella,

www.zeropaid.com, Last visited 5/01/2005

Grid Computing Definition - Source:

http://www.huihoo.com/grid/grid_computing_info_ce

ntre.htm, Last accessed: 5/01/2005

Open LDAP directory service protocol - Source:

www.openldap.org, Last accessed: 24/6/2004

P2PJXTA Framework - www.openp2p.com Last visited

5/01/2005

Sun Microsystems JXTA Framework - www.jxta.org Last

accessed: 5/01/2005

The TeraGrid Project - www.teragrid.org

UK’s E-Science Project - Source: www.escience-

grid.org.uk, Last accessed: 5/01/2005

DYNAMIC COALITION IN AGENT AWARE ADHOC VIRTUAL P2P INTERCONNECT GRID COMPUTING

SYSTEM – A3PVIGRID

175