JOURNAL OF COMPUTING, VOLUME 2, ISSUE 11, NOVEMBER 2010, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 27

An implementation of the Multiagent

System for Market-based Cloud
Resource Allocation
Yee Ming Chen and Hsin-Mei Yeh

Abstract— Multiagent technique offers a promising approach for distributed allocation of Cloud resources without centralized con-
trol. The application of market mechanisms for the resource allocation of Cloud computing services is a demanding task, which re-
quires bridging economic and associated software agent technical challenges. Dynamic changes in the availability of resources over
time makes the treatment more complicated. Here we employ an allocation mechanism and market mechanism as a market-based
multiagent resource allocation model to optimal allocate resources through genetic algorithm in a cloud computing environment. Buyer
and service provider agents determine their bid and ask prices using k-pricing which sets the transaction price individually for each
matched buyer-service provider pair. These mechanisms are adaptively to meet the Cloud users/ service providers requirement and
constraints set by bundled services.

Index Terms—Multiagent System、 Cloud Resource 、Optimal Allocation

——————————  ——————————

1 INTRODUCTION

T HE last years have seen the emergence of large distri-

buted computational systems, such as the Cloud
Computing, as a new approach to solve large-scale
problems. One of the main challenges that these systems
face is the efficient allocation of Cloud resources. In
Clouds, resources e.g. processing power, memory, sto-
rage, and bandwidth, can be bundled as services, which
are offered to other Cloud users. Cloud service providers
have to plan their resource usage carefully and be aware
of dynamic changes of the incoming requests for their
services. Therefore, various resource allocation approach-
es are an exciting area of research which have been pro-
posed to address the problem[1]. Some of them are based
on analogies between the system and real economies.
Commonly called market-based, those approaches offer a
promising solution [2,3,4]. In a market, Cloud service
providers face dynamic and unpredictable users beha-
vior. The way, how prices are set in a dynamic Cloud
market environment, can influence the demand behavior
of price sensitive users [5.6]. Multiagent system (MAS)
incorporates market mechanism into system allowing the
software agents to have preferences over some attributes
of the allocation, e.g. bundle of Cloud service and its price
of the resources. The market-based multiagent system
also allow for computational and geographical distribu-
tion decentralized implementation while providing me-
chanisms to regulate the behaviors of users.
————————————————
 Corresponding author:Yee Ming Chen is with the Department of Industrial
Engineering and Management, Yuan Ze University, Taoyuan, Taiwan.
In this paper we consider the multiagent for market-
based Cloud resource allocation, which contain allocation
mechanism and market-based mechanism. The resource
allocation mechanism needs to satisfy the Cloud service
providers and the users. The market-based mechanism
that uses the concept of k-pricing that is iteratively ad-
justed to find acceptable between a set of demands and a
limited supply of Cloud resource allocation. Our goal is
to devise the market-based Cloud resource optimal allo-
cation that maximizes the utility across all Cloud users
and Cloud service providers(i.e., welfare). Then the fol-
lowing economic requirements can be stated:
1) Economic efficiency: When the allocation is eco-
nomically efficient, it is impossible to increase a partic-
ipant's welfare without decreasing another partici-
pant's welfare; i.e. there is no wasted resource. Max-
imizing the total welfare is a sufficient condition for
economic efficiency. The proposed allocation mechan-
ism employs mixed integer programming to strictly
maximize the total welfare.
2) Double-side rationality: To encourage a fair exchange
between resource providers (sellers) and users (buy-
ers), the prices should only depend on the supply
demand conditions. The proposed market mechan-
ism is based on the k-pricing model [7,8] to give no
advantage on the seller's side or the buyer's side.
Therefore, the MAS resource allocation in this market
environment could be optimal either for the central entity
choosing the allocation; or with respect to a suitable ag-
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 11, NOVEMBER 2010, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
gregation of the preferences of the individual agents in
the system (e.g. agents may have preferences over the
bundles of service of they receive. in addition, they may
also have preferences over the bundles of services re-
ceived by other agents). The rest of the paper is organized
as follows. Section 2 deals with system structure related
to market-based multiagent Cloud resource allocation. In
Section 3, we describe the proposed DPSO based algo-
rithm in detail. Experimental results are presented in Sec-
tion 4 and some conclusions and future works are pro-
vided towards the end.
2 MULTIAGENT FOR MARKET-BASED FOR CLOUD
RESOURCE ALLOCATION
The two key roles driving the multiagent for market-
based Cloud resource allocation system are: Cloud Ser-
vice Providers (CSPs) providing the agents role of sellers
or supply and Cloud Users (CUs) representing buyer or
demand. The market-based MAS environments provide
the necessary infrastructure including security, informa-
tion, transparent access to remote resources, and cloud
services that enable us to bring these two entities togeth-
er. Users interact with their own brokers for managing
and scheduling their computations on the Cloud. The
CSPs make their Cloud resources enabled by running
software systems along with Cloud Trading Services
(CTS) model to enable resource trading and execution of
user requests directed through CUs. The interaction be-
tween CUs and CSPs during resource trading (service
cost establishment) is mediated through a Cloud Market
Directory (CMD) model (see Figure 1 ). They use various
welfare model, K-pricing model and interaction protocols
for deciding Cloud service access price. These models are
discussed in Section 3.
Figure 1 The multiagent for market-based Cloud resource
location system
CUs can be charged for access to various resources in-
cluding CPU cycles, storage, software, and network. The
CSPs users compose their bundles of service using inte-
28
raction protocols through CMD. The CTS (working for
the CSPs/CUs) can carry out the following steps for
Cloud resource allocation applications:
1). Identify Cloud service providers.
2). Identify suitable Cloud resources and establishes their
prices (by interacting with CSPs and CUs).
3). Select bundles of service that meet its welfare of alloca-
tion mechanism (e.g. lower cost and meet deadline re-
quirements). It uses genetic algorithm while optimiali-
zation Cloud resources and mapping requests to us-
ers.
4). Use market mechanism (working with K-pricing) ite-
ratively adjusted to find acceptable between a set of
demands and a limited Cloud supply for trading and
issues payments as agreed.
Figure 2 shows an illustration the proposed the market-
based multiagent resource allocation model which func-
tions embedded in the system as shown in the Figure 1. It
mainly functions consists of two inter-dependent me-
chanism: allocation mechanism and market mechanism.
There are two sorts of actors in a double side based market:
Cloud service provider agents and Cloud user agents (who
represents the Cloud market institution). As it shows, the
sellers and buyers submit sell orders and bid orders (buy
orders) to the market institution, respectively. The market
maker of the market institution sets pairs from these in-
coming orders according to the allocation and market
mechanisms that is applied to the market-based environ-
ment.
al-
Figure 2 The market-based multiagent resource allocation model
2.1 Allocation mechanism
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 11, NOVEMBER 2010, ISSN 2151-9617 29
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG

We formulate the allocation mechanism into the linear 1  j  N ,1  k  G ,1  t  T

mixed integer program[9,10] (d) The seller offers some service for the latest period
of all buyers'
Object function：Both buyers/sellers' greatest total wel-
fare
t  d i , k  jN1 y i , j , k ,t  0,
(9)
1  i  M ,1  k  G ,1  t  T
Maximize (e) The buyer acquisite bundles of service or not.
w  j1 k 1 t 1 vj z j,k,t qj,k  i1 j1 k 1 t 1 vi qi,k yi, j,k ,t
|N| G T |M| |N| |G| T

(1) u j  0,1,
Where z denotes whether the resource is allocated (10)
to the Cloud users. y denotes the proportion of quanti- 1 j  N
tu q allocated to the buyer accept vi is the minimum (f) The buyer accepts the service or not
price per timeslot ( 1  t  T ) at which the service pro- x j ,k  0,1,
vider wish to sell the bundled service. Allocation con- (11)
1  j  N ,1  k  G
straints divided into three types：
(g) The buyer accept this service in the time slot or
(I) The type service limiting: not

(a) Bundles of service needed by buyer z j ,k ,t  0,1,

(12)
1  j  N ,1  k  G ,1  t  T

|G |
x j , k  B j u j  0, 1  j  N
k 1 (2) (h) The seller offers the proportion of service
(b) Service in the time slot and limit 0  yi , j ,k ,t  1,
(13)

T
z j , k ,t  l j , k x j , k  0, 1  i  M ,1  j  N ,1  k  G ,1  t  T
t 1 (3)
1  j  B ,1  k  S One of the main contributions of this work concerns the
(II) The proportion of quantity of service limiting resolution approach employed. As explained earlier, we
have one objective function and thirteen constraints and
(a) Quantity of proportion of resource paid solutions should be searched considering these criteria

|N |
y i , j , k , t  1, simultaneously. To achieve this, we use a genetic algo-
j 1
(4) rithm(GA) . The evolution of the population takes place
1  i  M ,1  k  G ,1  t  T
following the general GA principles through tournament
(b) The service amount that the buyer receives is equal selection, crossover and mutation. In addition other ge-
to the service amount that all sellers offer netic operators and techniques can be used to improve

M
q j . k z j , k ,t  q y  0, the performance of the GA, such as Elitism, It is imple-
i 1 i ,k i , j ,k ,t (5) mented so that the best solution of every generation is
1  j  N ,1  k  G ,1  t  T copied to the next so that the possibility of its destruction
through a genetic operator is eliminated. The flow chart
(III) Serve in time slot and limiting of figure 3 describes the main steps of the GA procedure.

(a) The buyer begins to accept in the period earlier

than buyers and need to accept this service period .

a j , k 
 t z j ,k ,t  0 , (6)

(b) Time slot when the buyer accepted and can't be

later than buyers and need this service this service period
finally

t  d j,k z j,k ,t  0, (7)

1  j  N ,1  k  G ,1  t  T
(c) The seller offers service for the earliest period of all
buyers'

a j,k  t jN1 yi, j,k ,t  0,

(8)
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 11, NOVEMBER 2010, ISSN 2151-9617 30
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG

Start GA For instance, assume an allocation of resources from a

specific service provider to a specific buyer. The buyer
Initial Solutions
values these resources at $10 while the service provider
has a reserve price of $5. Then the (local) welfare of this
▪ Random generate a population
of bundles service allocation transaction is $10- $5 = $5, and k  $5 of the surplus is
▪ Each constraints of binary code allotted to the buyer (who thus has to pay $10 - k  $5)
solution with Eq.(2)~(13)
and ( 1- k)  $5 is allotted to the service provider (who
thus receives $5+(1-k)  $5).
generation i = i + 1 k-Pricing has two main advantages. The distribution
of welfare among users and providers can be flexibly pre-
Feasible solutions check
defined by setting the factor k, thus allowing for both
For the changed time periods, fairness and revenue considerations[14].
check the constraints

Evaluate fitness function

Eq. (1)
3. MULTIAGENT RESOURCE ALLOCATION SIMULA-
TION DEVELOPMENT
GA operation
▪ Tournament selection
▪ Crossover
▪ Mutation We are developing a simulation environment to eva-
luate the performance of the proposed market-based mul-
Elitism tiagent resource allocation model. In this section, we will
present the experimental results and comparative the
Optimal computational performance. The platform for conducting
No Iteration finish ? Yes
Allocation the experiments in a PC with Dual Core Processor
4400+2.29 GHz CPU and 1.75GB RAM. All programs are
coded in Java programming language in Borland JBuilder
Figure 3 The GA procedure for resource allocation mechanism
2006. The simulation environment consists of buy-
er/service provider agents and a market exchange. The
individual agent has its own preferences to decide the bid
2.2 Market mechanism
according to its budget and the market trend. The prices
and payments generated by proportional critical value
As introduced above, besides the allocation mechanism
pricing and k-Pricing with k = 0.5. After we know the
in the market-based multiagent resource allocation model,
basic information, the buyer agents and the service pro-
model consists another market mechanism. While the
vider agents start to trade with each other which are illu-
allocation mechanism generally accounts for the technical strated as Figure 3.
specifics of the Cloud setting, the objective of the pricing
component is to induce the selfish participants in the
market-based environment to act towards the general
objective, in this case welfare maximization. To this end,
in this section k-Pricing scheme is presented that can be
used in conjunction with the allocation mechanism. k-
Pricing is an alternative market mechanism to propor-
tional critical value pricing. It was introduced in Schnizler
et al. (2008)[11] for double-sided combinatorial auctions.
We employ the K-pricing scheme to fix the resource. Let
0  k  1 be an arbitrary proportion. The price P is then
calculated as

Pi , j ,k ,t  K ( v j z j ,k ,t )  (1  K )( vi yi , j ,k ,t ) (14)

The basic idea is to distribute the welfare generated by

the allocation mechanism between buyers and service
providers according to a factor k  [0,1] . K-pricing
Figure 3 Trade exchange for buyer agents and the Cloud service
enables us to disentangle the risk aspect of discounting provider agents
from the time aspect and allows us to quantify exactly
how much of an asset’s current price is due to time dis-
counting and how much is due to risk discounting. From figure 4 to 7 illustrate a series of trading among
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 11, NOVEMBER 2010, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
seven participants: three Cloud service provider agents
offer three bundled services. Buyer agent 1, 2, 3 and 4
get their need different bundled service simultaneously.
Figure 4 Buyer agent 1 with all service provider agents re-
source allocation after trading
Figure 5 Buyer agent 2 with all service provider agents re-
source allocation after trading
31
Figure 6 Buyer agent 3 with all service provider agents re-
source allocation after trading
Figure 7 Buyer agent 4 with all service provider agents re-
source allocation after trading
As buyer's number increases, the degree of competition
among the buyers also increases. When the competition is
the stronger , will also be high buyer's bid. They would
like to bid by higher than the lowest bid price offered at
that timing, we call that jump bidding in this behavior.
That is the ratio of jump bidding(ROJB):
JumpBidding  Cleaning Pr ice
ROJB  (15)
Cleaning Pr ice
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 11, NOVEMBER 2010, ISSN 2151-9617 32
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG

The higher in k value, the cleaning price Inclined to buy- tions show that the model works properly. We are inter-
er's bid price. This shows that it is comparatively favora- ested in the behavior of the trading exchange, particularly
ble to the seller. therefore both buyer/sellers' greatest the interaction between Cloud user agents and Cloud
welfare will increase as shown in Table 1. service provider agents. Market –based model model
shows promise for enhancing resource allocation and
pricing. We are now working toward GA and K-pricing
Table 1 Different ROJB and k value are produced welfare (w) schemes find an interesting match between scalability
ROJB and individual behavior. By means of MAS simulations,
the effect of multiple Cloud users and providers strategic
K value behavior might be investigated. Extensions of the heuris-
1 1.1 1.2 1.3 1.4 tic allocation scheme, such as the
use of more sophisticated norms in the sorting phase, as
0.3 61.52 73.35 116.01 143.97 144.33
well as the study of their impact on the mechanisms’
0.5 65.21 82.79 120.21 135.16 146.75 strategic properties, might be further promising areas for
0.7 67.52 97.35 128.39 148.87 203.08 future research.

Different k value and ROJB value were produced wel- ACKNOWLEDGEMENTS

fare as the figure 8 . The buyer would like to offer it with
the higher price, then welfare(w) rising also becomes This research work was sponsored by the National
more and more obvious. Safeguard one's own interests Science Council, R.O.C., under project number NSC99-
and incline to choose higher k value. Not only can make 2221-E-155-022.
welfare value improve like this , but also make the trade
succeed.

Figure 8 Different welfare (w) in various k and ROJB
Therefore, if can accept higher k value with the attitude of
compromise, both buyers/sellers' greatest total welfare
will also increase.
4. CONCLUSIONS AND FUTURE WORK
In this paper we proposed the market-based multia-
gent resource allocation model on cloud computing envi-
ronment. It allows buyers to order an arbitrary composi-
tion of services to different service providers. The pro-
posed allocation mechanism and market mechanism de-
velop MAS simulation environment to evaluate the per-
formance of the proposed model. The preliminary simula-
REFERENCES
[1] K. L. Mills and C. D. Dabrowski. “Can economics- based re-
source allocation prove effective in a computation marketplace?
“,Journal of Grid Computing, Vol.6, pp.:291-311,Sept, 2008.
[2] Z. Yang, N. Ye and Y.-C. Lai, “QoS model of a router with feed-
back control”, Quality and Reliability Engineering Int’l, vol. 22(4),
pp. 429-444, 2006.
[3] R. Buyya, C. S. Yeo, and S. Venugopal, “Market Oriented Cloud
Computing: Vision, Hype, and Reality for Delivering IT Services
as Computing Utilities”, Proc. 10th IEEE Int’l Conf. on High Per-
formance Computing and Communications, pp. 234-242, 2008.
[4] S. S. Yau, Y. Yin and H. G. An, “An Adaptive Model for Tradeoff
between Service Performance and Security in Service-based En-
vironments”, Proc. Int’l Conf. Web Services (ICWS 2009), pp. 287-
294, 2009.
[5] C. T. Yang, P. C. Shih, C. F. Lin and S. Y. Chen, “A Resource
Broker with an Efficient Network Information Model on Grid
Environments”, The Journal of Supercomputing, vol. 40(3), pp. 249-
267,2007.
[6] S. S. Yau, N. Ye, H. Sarjoughian and D. Huang, “Developing
Service-based Software Systems with QoS Monitoring and
Adaptation”, Proc. 12th Int’l Workshop on Future Trends of Distri-
buted Computing Systems, pp. 74-80, 2008.
[7] J. Stoesser, C. Roessle and D. Neumann, “Decentralized Online
Resource Allocation for Dynamic Web Service Applications”,
Proc. 4th Int’l Conf. Enterprise Computing, E-Commerce, and E-
Service, pp. 425-428, 2007.
[8] R. Buyya, C. S. Yeo, and S. Venugopal, “Market Oriented Cloud
Computing: Vision, Hype, and Reality for Delivering IT Services
as Computing Utilities”, Proc. 10th IEEE Int’l Conf. on High Per-
formance Computing and Communications, pp. 234-242, 2008.
[9] I. Griva, S. G. nash and A. Sofer, “Linear and Nonlinear Optima-
zation (2nd ed.)”, Society for Industrial Mathematics, Philadel-
phia, PA, 2008.
[10] Ikki Fujiwara, Kento Aida and Isao Ono, “Market-based Service
Allocation for Distributed Computing”, The institute of electron-
ics information and communication engineers,2009.
[11] B. Schnizler, D. Neumann, D. Veit, C. Weinhardt,” Trading grid
JOURNAL OF COMPUTING, VOLUME 2, ISSUE 11, NOVEMBER 2010, ISSN 2151-9617 33
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG

services – A multi-attribute combinatorial approach’. European

Journal of Operational Research 187 (3), 943– 961,2008.
[12] M. Kallits, G. Michailidis and M. Devetsikiotis, “Pricing and
Optimal Resource Allocation in Next Generation Network Ser-
vices”, Proc. IEEE Sarnoff Symposium, pp. 1-5, 2007.
[13] C. Bae and D. Cho, “Fairness-Aware Adaptive Resource Alloca-
tion Scheme in Multihop OFDMA Systems”, IEEE Communica-
tions Letters, vol. 11(2), pp. 134-136,2007.
[14] J. Stoesser, C. Roessle and D. Neumann, “Decentralized Online
Resource Allocation for Dynamic Web Service Applications”,
Proc. 4th Int’l Conf. Enterprise Computing, E-Commerce, and E-
Service, pp. 425-428,2007.

Biography of the authors:

Yee Ming Chen is a professor in the Department

of Industrial Engineering and Management at Yu-
an Ze University, where he carries out basic and
applied research in agent-based computing. His
current research interests include soft computing,
supply chain management, and pattern recogni-
tion.

Hsin-Mei Yeh was a graduated student in the De-

partment of Industrial Engineering and Manage-
ment at Yuan Ze University, where she was study-
ing basic and applied research in Cloud computing
and heuristic methods. She now works in AU Op-
tronics Corporation as Engineering.