A Framework to Provide Trust and Incentive in CROWN Grid for Dynamic Resource Management

A Framework to Provide Trust and Incentive in

CROWN Grid for Dynamic Resource Management

Yu Zhang1 _{, Jinpeng Huai}1 _{, Yunhao Liu}2 _{, Li Lin}1_{, Baijian Yang}3 1_{Dept. of Computer Science and Techonology}

Beihang University, Beijing, China [email protected]

2_{Dept. of Computer Science,}

Hong Kong University of Science & Technology

[email protected]

3_{Dept. of Industry and Technology,}

Ball State University [email protected]

Abstract—In order to maximize resource utilization as well as providing trust management in grid, we propose a novel framework – Trust-Incentive Resource Management (TIM). Having child model, club model, bid model and trust model, TIM dynamically manages grid resource by integrating values of prices, trust, and incentive. In this mechanism, providers set the price ac-cording to demand and supply, and consumers maximize the sur-plus upon budget and deadline. A weighted voting scheme is also proposed to secure the grid system by declining the join request from malicious nodes. A TIM prototype has been successfully im-plemented in a real grid system, CROWN grid. We evaluate the proposed approach through comprehensive experiments and achieve improved results in resource allocation efficiency, system completion time, and aggregated resource utilization.

Keywords-resource allocation; incentive; trust; CROWN

I. INTRODUCTION

A dilemma in grid computing area is that when every par-ticipating node tries to maximize its own utility, the overall util-ity of the collaboration might drop. In the worst case scenario, grid resources are easily depleted due to selfish users taking free rides without offering any sharing resource. Unfortunately, such “tragedy of the commons” phenomenon also happens in a num-ber of existing grid systems where cooperated scientific re-search grid systems emphasize on sharing resource voluntarily. Apparently, certain resource management scheme has to be im-plemented on grid systems to achieve better scalability[1-4,17].

To encourage resource sharing, several previous works adopt soft incentive schemes[5-7], which is essentially a reputa-tion system. Nodes get higher degree of trust by sharing more resources, and thus have the permission to access other re-sources. Soft incentive cannot meet the requirement of grid sys-tems in that providers not only care about the reputation, but also wish to gain benefit by providing resources. Other works adopt hard incentive scheme[1, 4], in which nodes get virtual currency by selling their resources, and then use the currency to bid for other resources. However, the assumption, wealthy nodes are more trustful, is not always valid. Simply considering

the bid price in resource allocation cannot satisfy the security concerns from different participating organizations.

In order to tackle the mentioned above issues, we combine the soft incentive and hard incentive schemes and propose a Trust-Incentive Compatible Dynamic Resource Management framework, called TIM. The primary goals of TIM are securing shared resources, promoting users to share valuable resources, maintaining the balance of supply and demand in competitive grid resource market, and finally maximizing aggregate re-source utilization. Major contributions are summarized as fol-lows.

(1) By adopting a continual exchange process and matching user resource requests with available resources, we propose the TIM model which can maximize aggregate resource utilization in an economically and computationally efficient manner. In this model, users can get more only if they are willing to share more and having higher degree of trust. As a result, it promotes collaborators to share more valuable resources and avoid mali-cious waste.

(2) In a grid system, price fluctuates when resource supply and demand changes. We separate the role of providers and consumers and apply different strategies to each role. Providers simply mark their price based on supply and demand while con-sumers offer their bids upon deadline and budget constraints. Because the resource dynamics are included in our management model, the workload of providers is balanced and the efficiency of grid system is improved.

(3) Nodes may join or leave a grid system randomly. To prevent a system from being attacked by malicious nodes, we employ a weighted voting scheme in TIM, such that a secure environment is constructed.

(4) We have successfully implemented TIM in the key pro-ject of our lab, CROWN grid. Our implementation experiences and experimental results are valuable to research peers.

II. RELATED WORK

The objective of resource incentive is to promote users to share more resources. Generally speaking, there are two incen-tive schemes: Soft Incenincen-tive[5-7] and Hard Incenincen-tive [1, 4, 8-10].

Soft incentive includes two models, Peer-Approved and Service-Quality. Peers in [5] are allowed to access resources only from others with a lower or equal ratings, and the QoS provided to these peers also can be differentiated. Feldman [6] proposes a Reciprocative decision function, and introduces the notion of generosity. Richard et al. [7] present an allocation mechanism for bandwidth resource and introduce the notion of contribution to maximize utilization. In essence, soft incentive in above work is a reputation system where the reputation [5] (or generosity[6], contribution[7]) of a peer is consistent with the utility and quantity of resources supplied by the peer, but no price mechanism is involved in above systems.

There have been many researches in resource management approach which are based on economic models [1, 4, 8-10]. Re-source management based on price scheme can be treated as hard incentive, which adopts a Token-Exchange approach. Each first-time user might be allotted a fixed number of tokens, but once these run out, the user has to serve resources to earn to-kens. Chun [1] allocates resources using a combinatorial auc-tion that allows users to express preferences with complemen-tarities. Feldman [8] presents a price-anticipating resource allo-cation mechanism. In their work, each user can reach the Nash equilibrium by iteratively applying a best response algorithm to adapt his bids. Resource allocation in above researches mainly considers the bid price of consumers, that is, the higher price the node bids, the more resources the node gets. However, wealthy nodes are not always with higher degree reputation. Only considering the bid price in resource allocation cannot sat-isfy the security concerns from different participating organiza-tions.

The TIM we proposed combines features from both soft in-centive approach and hard inin-centive approach. The framework is dynamic oriented and completely distributed. Such design allows better scalability when managing practical grid resources. The weighted voting scheme included in our work can decline suspicious nodes from joining the grid system and therefore maintaining a secure and balanced grid environment.

III. SYSTEM MODEL

The key project in our lab, CROWN (China R&D Environ-ment Over Wide-area Network), is aiming at empowering in-depth integration of resources and cooperation of researchers nationwide and worldwide. As illustrated in Fig. 1, the distrib-uted resource management in CROWN grid adopts a two-tier peer-to-peer architecture[11-13]. The super layer is the back-bone consisting of CDRSes (CROWN Distributed Resource Server); the child layer includes all clients and resource provid-ers. Each club consists of one CDRS and multiple child nodes,

such as Clubj in Fig. 1. Each CDRS will periodically publish the provision resources. A child node selects the desired resource, generates the corresponding bids, and transmits them to the se-lected CDRS by its own parent CDRS.

We first propose our TIM model, including four building blocks, i.e. CHILD model, CLUB model, BIDS model, and TRUST model. Suppose that there are p clubs (that is, p CDRSes), and each club may have n types of resources. Let

Clubj denote the jth Club (j=1,", p), and Childijdenote the ith child node in Clubj.

CHILD Model Childij is defined as ( i, i, i

j j j S E Slot c c c 1 2 1 2 , , ) : i ( i, i, , i), ( i, i, , i), j j j j j j j n n c c c c i c c c j i i directtrust revenue_{j j} S s s s E e e e c = " = " 1 2 ( , , , ) i i i i j j j j n c c c c

Slot = slot slot "slot to denote supply vector, excess vector, and time slot vector of i

j Child . i, i, i j j j t t t c c c s e slot (t =1,", n) denotes supply quantity, demand quantity, and time slot of the

tth resource type; a direct trust value directtrustij∈[0,1] in Clubj;

a node revenue i

j

revenue ∈ R.

CLUB Model Clubj is defined as (numj, Sj, Dj, Ej, Pj, Pjnode,

Bidsj, RecTrustj): a children number num_j∈ ;N S_j=( ,s s1_j 2_j,

,sn_j) " , _{( ,}1 2_{, ,} n₎ j j j j D = d d " d , _{( , , , )}1 2 n j j j j E = e e " e , _{( ,}1 j j P = p 2_{, ,} n₎ j j

p " p to denote supply vector, demand vector, excess vec-tor, and price vector of the resources in Clubj,respectively, where t, t, ,t t

j j j j

s d e p (t=1, , )" n denote supply quantity, demand quantity, excess quantity and price of the tth resource type; an average price of child node

1 n node t t j j j j t P s p num = =

∑

vec-tor Bidsj=(bids bids1_j, 2_j, ," bidsm_j) received by Clubj, and bidsq_j is the qth bid; a recommendation trust value vector

1 2

( _j , _j , , _jp)

j

RecTrust = rectrust rectrust " rectrust , rectrust_ji [0,1]

∈ is the recommendation trust value from CDRS in Clubj to CDRS in Clubi.

BIDS Model Bids is defined as (cid, pid, et, Q, (q1_{, q}2_{, …, q}n_),

directtrust, C, V): cid∈{1, " , p}, pid N∈ to denote the bid

coming from pid

cid

Child ; a estimated job execution time et; a de-sired quantity Q≥ 1 of nodes; q1_{, q}2_{, … , q}n _{to denote the tth}

(t= "1, , )n type resource quantity required by the consumer; a direct trust value directtrust∈[0,1] of the pid

cid

Child ; a set of constraints C, such as deadline and budget, etc; and a bidding price V that the consumer is willing to pay.

Figure 1.The two-tier architecture of TIM

TRUST Model We adopt the trust model used in [14]. If T1

de-notes a recommendation trust value from A to B, and T2 denotes a direct trust value from B to C, then the trust value from A to C is _{1－（ －1 T2)}T1_{. We suppose that nodes in the same club have}

direct trust relations, nodes in different clubs have indirect trust relations, and CDRSes have recommendation trust relations. We get the following trust inference.

(1) Direct trust computing. After nodes in Clubi use the

re-sources of k

i

Child to execute jobs, the nodes will report positive or negative experiences to the CDRS in Clubi. A direct trust re-lationship will set up only if all experiences with k

i

Child that the CDRS in Clubi knows about are positive experiences. Let q be the number of positive experiences, then the direct trust value from the CDRS to Childik is directtrustik = −1 λq , whereλ is the probability of reliability with a single task;

(2) Recommendation trust computing. After nodes in Clubj have used the resources of nodes in Clubi to execute jobs, the nodes in Clubj will report positive or negative experiences to the CDRS in Clubj. Given numbers of positive and negative experi-ences p and n, the recommendation trust value from the CDRS of Clubj to the CDRS of Clubi is：

rectrustji(p, n)= 1 0 p n_{if p n} else λ −  − >    ;

(3) The trust value from Clubj to Childik is :

1 (1 ) ji

k k

i i

rectrust

trust = − −directtrust .

IV. DESIGN OF TIM

In this section, we present our TIM framework to manage shared resources in CROWN.

A. Provider price strategy

Smale Theorem [15]: For a market with n (n 2_{≥ ) types of}

commodity, P = (p1," , pn) is the price vector at some moment, where 0(pi≥ i= "1, , )n is the price of the ith commodity. Let D, S, E denote supply vector, demand vector, and excess vector

of commodities (E=D-S), and D, S, E be the functions on the price vector P. If the equilibrium point of market exists, then E(P*) = 0, where P* denotes the equilibrium price. By default, the equilibrium point cannot be reached automatically. Rather, it can be reached only by continually adjusting the price accord-ing toD P_E( ) dP E p( ) dt µ × = , where ( ) ( l ) E n n m D P e p × ∂ = ∂ , l m e p ∂ ∂ denotes the partial derivative of the excess quantity of the lth type commodity to the price of the mth type commodity, and

µ denotes a coefficient with the same sign as ( )D PE .

In particular, it is impossible to use Smale’s method directly [9] because grid economy is inherently discrete and the partial derivatives that the method requires do not exist. However, we are able to get good approximations for the partials at a given price vector. Starting with a price vector, the preferences at price vectors can be obtained by fixing all but one price and varying the remaining price slightly. Once achieving a “secant” approximation for each commodity, we substitute these ap-proximations for the values of the partial derivatives in the ma-trixD PE( ), discretize with respect to time, solve for a price vec-tor, and iterate.

Definition Let Sj, Dj, Ej, Pj denote supply vector, demand vector, excess vector, and price vector in Clubj, respectively. Obviously, Ej = Dj-Sj, and Sj, Dj, Ej are the functions of Pj. IfE Pj( ) 0j* = , then Clubj reaches the balance, whereP*jis the equilibrium price. The unit of price is given in “grid dollars” (G$).

In the grid, let the coefficient _{µ be 1, dt be the step length} of the time for adjusting per unit price, and dP be the difference of two continues price. For example, suppose there are four types of resource in Clubj, CPU, memory, disk, and bandwidth. The current price and excess vector are Pj=(2G$, 6G$, 9G$, 4G$) and Ej =(200, 0, 300, 100). With the resource exchange records, the CDRS in Clubj find that when the price of some resource is increased one unit, the increased supply quantity is 20 units, and the decreased demand quantity is 30 units. The decreased demand quantity of other types of resource is 10 units.

1 2 3 4 -20-30 -10 -10 -10 200 -10 -20-30 -10 -10 0 -10 -10 -20-30 -10 300 -10 -10 -10 -20-30 100 p p p p _∆   _{ }    _{∆ }    _{ =}   _{∆ } ₋   _{ }    _{∆ }   Thus, _{∆ = − ∆ =p}1 _{5, 0, 7.5, 2.5p}2 _{∆ =p}3 _{∆ = −p}4 _{. We use}

the following formula _max{ ,_{ε p}k _{+ ∆p}k} _{to determine a new} resource price, whereε > is a small constant preventing prices 0 to approach zero value.

Clubi Clubj k child layer super layer _CDRS CDRS CDRS CDRS CDRS

B. Consumer price strategy

Each grid user generates the combination of resources for its tasks according to their requirements, and submits the corre-sponding combination to a selected CDRS that the user will bid for. The goal of each grid user is to maximize its own surplus upon deadline and budget constraints. Given the average price

node j

P of each node in the Clubj and the completion time con-straint T, the utility function of each grid user can be expressed as follows:

( ) ( _jnode )

U V =K T−L P V −V

in which L is the length of the job, V is the payment value of the user, LPjnode V is the estimated job execution time, and K is a constant coefficient defined by the user. Thus, the utility opti-mization problem above can be written as:

( ) ( ) ) . . ( ) 0 node j Max U V g V T LP V s t g V   ≥  ( =

-By using the multiplicator method, we obtain the approximative optimal solution V∗ =3L P_jnode (2T−K) as the bidding price, that is, the value V in BIDS model.

C. Trust-incentive compatible resource allocation algo-rithm

After collecting the consumers’ bids, the CDRS of each club uses the TIM mechanism to allocate resources as the following greedy algorithm.

TABLE Ⅰ. TIM ALGORITHM

1: Calculate the per unit valuation bidij=Vji (Qij×etij) for each bid in

Clubj

2: Scale each bidi_jusing the formulabi_j=bid bidi_j max_j

3: Calculate the evaluation value of each bid byevlbidi_j=

(1 )

i pid

j trustcid

b

α× + −α × , α∈[0,1] is risk degree of Clubj 4: Sort all bids in descending order according to_evlbidij 5: for all bids in the sorted bid list do

6: if the resource request for a bid can be fulfilled with the remaining node resources then

7: allocate the resources to the bid. 8: end if

9: end for

D. Dynamic management for nodes

Nodes may join and leave the collaboration dynamically, or transfer from one club to another club. Thus, some CDRSes maybe have the trust records for a child node. In our TIM ap-proach, the more resources a

node provides, the higher degree of trust a node has. Obvi-ously, every node is willing to cooperate with a node with higher trust value.

As illustrated in Fig. 1, when the CDRS in Clubj receives the join request, it will propagate the join request to other CDRSes. Since each club may have different trust records and recommendation trust value, it may have different opinions about the requesting node. In our proposal, this is solved by employing a weighted voting scheme to decide whether to ac-cept the requesting node or not. After receiving the vote request, the CDRSes make their own decisions as:

1 > 0 -1 < i

trust

vote no trust value record trust τ τ  =  

where τ ∈[0,1]is the configuration threshold value used by each CDRS. The result will then be returned to the voting spon-sor. According to the majority principle, the CDRS in Clubj uses the constraint in equation

1 0

p

ji i

i∑= rectrust vote ≥ to make the

final decision.

V. IMPLEMENTATION

We have implemented the TIM approach in CROWN sys-tem with Java. The cooperation facility among nodes is pro-vided by CROWN grid, a fully decentralized grid middleware infrastructure. In the setup phase, we created five clubs [16], and each club has a CDRS and 40-50 virtual child nodes.

A. Efficiency of allocation mechanism

We first conduct an experiment with a set of five peers with varying currency and trust value, bidding for some portion of 200 unit resources. In the experiment, the amount of resource that each bid i can obtain is determined by

i i

i

TotalResources evlbid× ∑evlbid . We adopt this distribution policy to protect light users against starvation from heavy users when the demand is over the supply. To evaluate TIM, we first define a metric named allocation_ratio for each bid as follows.

= _ _ _

allocation_ratio allocation quantity total request quantity

In this experiment, we assume that the total request quantity of each node is 100 units. There are five curves in Figures 2 and 3, where x-axis represents the bids times, and y-axis represents

0 1 2 3 4 5 6 7 8 9 10 11 0 20 40 60 80 100 Bids Times A llo c a tion R a ti o( % ) S=90% S=30% S=60% S=30% S=90% 0 1 2 3 4 5 6 7 8 9 10 11 0 20 40 60 80 100 Bids Times Al loc a tion R a tio (% ) S=90% T=0.2 S=30% T=0.8 S=60% T=0.6 S=30% T=0.1 S=90% T=0.9 0 10 20 30 40 40 60 80 20 40 60 80 100 X-Completion Time(s ) Y -Sy st em Load (%) Z -A cc u m u la te d C om p le ti on R a ti o (% )

TIM Price Strategy Robin Strateg y

83% 64%

Figure 2. .Incentive compatible allocation Figure 3.Trust incentive allocation Figure 4.Completion ratio vs. dynamic system load

the allocation ratio. We first set the risk degree _α=1 in Fig. 2 and consider the security factor and set the risk degree α=0.2

in Fig. 3. The symbol S in both figures denotes the ratio of pro-viding resources, for example, the first node provides 90 per-cent of local resources to other nodes. The symbol T in Fig. 3 denotes the trust value of bidding nodes. In the first three peri-ods, only two nodes request for resources and the supply meets the demand. Thus the allocation ratios for the two nodes are all 100 percent. After the third period, the increasing bids outnum-bered the supply.

As shown in Fig. 2, without the security consideration, the nodes providing the same resources nearly obtain the same allo-cation ratio. However, if considering the trust value in Fig. 3, we can see that the provision ratio of node 3 (60%) is less than that of node 1 (90%), but node 3 has a higher trust value (T=0.6) and thus obtains more resources than node 1. Similarly, node 2 and node 4 have the same provision ratio: 30%. Node 2 obtains more resources than node 4 at the sixth bid in Fig. 3, because the trust value of node 2 (T=0.8) is far greater than that of node 4 (T=0.1).

B. Impact of TIM price strategy

This experiment is to study characteristics of price setting strategy with Round-Robin strategy in terms of job completion time, which is measured from accessing the requested grid re-sources till task is accomplished. Three clubs are the resource

providers, with each having 200 unit resources. Resource re-quests are generated by the child nodes and the bid is generated at an interval of 350 time units. We change the system load from 0.1 to 0.9 with a step of 0.1, where system load is defined as a ratio of aggregate bids load to aggregated capability of pro-viders. The initial value of the resource price is 50G$, and each CDRS re-publishes the resource price with an interval of 500 time units. In Fig. 4, we contrast the performance between TIM price direction strategy and Round-Robin strategy with system load at 0.4, 0.6, and 0.8, respectively. From Fig.4, we can see that TIM price setting strategy has better efficiency and spends less time to complete tasks compared with the Round-Robin strategy, especially at higher bids. At 10,000 time units, the completion ratio for Round-Robin strategy is only around 64%, while our price direction strategy can score 83% of the comple-tion ratios.

C. Evaluation of trust model

Malicious nodes may exist in grid environments to disturb resource exchange. We consider the security problem from both sides, including malicious consumers and malicious providers. On one hand, when bidding for resources, a malicious consumer can either set a higher bit value arbitrarily or does not give the corresponding payment. Such behaviors adversely affect the interests of resource providers. On the other side, a malicious provider can boast of having more resources to get more cur-rency. It makes good providers losing the bids.

0 10 20 30 40 50 60 70 80 90 0 200 400 600 800 1000 1200 1400 1600 1800 2000

% of Malicious Peers in 100 Bids

C lub R e v e nue (G $ ) risk defree =1 risk deg ree =0.5 risk deg ree =0

10 20 30 40 50 60 70 80 90100 0 200 400 600 800 1000 1200 1400

% of Malicious Nodes in 40 Joining Peers

Clu b Re v e n u e (G $ )

club with TIM club without TIM

0 10 20 30 40 50 60 70 80 90 100 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

% of Malicious Nodes in 40 Joining Peers

R e cTr u st o f C D S R

club with TIM club without TIM

We first study how malicious consumers affect the grid sys-tem. In this experiment, there are three clubs and 40 nodes in each club are resource providers. 100 nodes from the other two clubs act as consumers and generate bids. We let the three pro-viding clubs receive the 100 bids each time by varying the per-centage of malicious consumers from 0 to 0.9 with a step of 0.1. For each scenario, a set of risk degree_{α in TIM are configured:} 0, 0.5, and 1. As shown in Fig. 5, the club considering both benefit and security factors (α =0.5)obtains more revenue af-ter the percentage of malicious consumers is over 20%. When the percentage of malicious consumers is over 80%, the club considering only the security factors will get more revenue.

We then study how malicious providers affect the grid sys-tem. In the simulations, Club1 implements admission control based on the TIM weighted voting scheme, while Club2 dose not. Both clubs have 40 resource providers and are handling 40 bids. Consider the case when 40 nodes from the other three clubs want to join Club1and Club2. Still, we set the percentage of malicious join peers from 0 to 0.9 with a step of 0.1. The CDRS in Club1set the admission policy τ =0.3, which means that the node whose trust value is lower than 0.3 will not be ac-cepted. As a result, few malicious nodes are able to join Club1, while many of them can join Club2.

As Shown in Fig. 6, when the percentage of malicious join increases, the revenue of both clubs decreases. But the revenue of Club2 without TIM is suffering much more than Club1 with TIM. At the extreme case when all the new nodes are malicious nodes, the revenue of Club2 is only 60 percent of the revenue of

Club1. Fig. 7 shows that as more malicious nodes enter Club2, the recommendation trust value of the CDRS in Club2 without TIM de crease rapidly.

VI. CONCLUSION

Proving trust and incentive in grid computing environments are of great importance. This paper presents a Trust-Incentive Compatible Dynamic Resource Management, TIM, on the basis of economy model and trust model. By introducing the price strategy, trust-incentive compatible resource allocation mecha-nism and the weighted voting scheme, TIM encourges nodes to share more resources, ensures the balance of supply and de-mand, enhances the aggregated resource utilization, and main-tains the secure environment of a grid computing system.

TIM scheme has been successfully implemented in our key project, CROWN grid environment. We evaluate our proposed approach by comprehensive experiments and achieved much improved results in resource allocation, system completion time and aggregated resource utilization.

In the future, we will widely deploy our TIM approach in the CROWN grid to construct a more secure and balanced col-laborative grid environment.

ACKNOWLEDGMENT

This work is supported by grants from the China National Science Foundation (Project No.90412011, No.60573053), China 863 High-tech Programme (Project No.2005AA115420) and China Major State Basic Research Development Program (Project No. 2005CB321803).

REFERENCES

[1] B. Chun, J. Albrecht, D. Parkes, and A. Vahdat, “Computational Resource Exchanges for Distributed Resource Allocation,” Technical Report, 2004. [2] Z. Zhu and X. Zhang, “Look-Ahead Architecture Adaptation to Reduce

Processor Power Consumption,” IEEE Micro, 2005.

[3] K. Shen, H. Tang, T. Yang, and L. Chu, “Integrated Resource Manage-ment for Cluster-based Internet Services,” In Proceedings of the 5th sym-posium on Operating Systems Design and Implementation, 2002. [4] C. Yeo and R. Buyya, “Pricing for Utility-driven Resource Management

and Allocation in Clusters,” In Proceedings of the 12th International Con-ference on Advanced Computing and Communication, 2004.

[5] K. Ranganathan, M. Ripeanu, A. Sarin, and I. Foster, “To Share or not to Share: an Analysis of Incentives to Contribute in Collaborative File Shar-ing Environments,” In ProceedShar-ings of Workshop on Economics of Peer to Peer Systems, 2003.

[6] M. Feldman, K. Lai, I. Stoica, and J. Chuang, “Robust Incentive Tech-niques for Peer-to-Peer Networks,” In Proceedings of ACM E-Commerce Conference, 2004.

[7] T. Ma, S. Lee, J. Lui, and D. Yau, “A Game Theoretic Approach to Pro-vide Incentive and Service Differentiation in P2P Networks,” In Proceed-ings of ACM SIGMETRICS/PERFORMANCE, 2004.

[8] M. Feldman, K. Lai, and L. Zhang, “A Price-Anticipating Resource Allo-cation Mechanism for Distributed Shared Clusters,” In Proceedings of ACM E-Commerce Conference, 2005.

[9] R. Wolski, J. Brevik, J. S. Plank, and T. Bryan, “Grid Resource Allocation and Control Using Computational Economies,” In Proceedings of Grid Computing: Making the Global Infrastructure a Reality, 2003.

[10] A. Das and D. Grosu, “Combinatorial Auction-Based Protocols for Re-source Allocation in Grids,” In Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium, 2005.

[11] Z. Zhuang, Y. Liu, and L. Xiao, “Dynamic Layer Management in Super-peer Architectures,” In Proceedings of the 33th International Conference on Parallel Processing, 2004.

[12] S. Banerjee, C. Kommareddy, K. Kar, B. Bhattacharjee, and S. Khuller, “Construction of an Efficient Overlay Multicast Infrastructure for Real-time Applications,” In Proceedings of INFOCOM, 2003.

[13] R. Ranjan, A. Harwood, and R. Buyya, “Grid Federation: An Economy Based, Scalable Distributed Resource Management System for Large-Scale Resource Coupling,” Grid Computing and Distributed Systems Laboratory, University of Melbourne, Australia, Technical Report, 2004. [14] T. Beth, B. Malte, and K. Birgit, “Valuation of Trust in Open Networks,”

In Proceedings of the Conference on Computer Security, 1994.

[15] S.Smale, “Dynamic in General Equlibrium Theory,” American Economic Review, 1976.

[16] J. Huai, Y. Zhang, X. Li, and Y. Liu, “Distributed Access Control in CROWN Groups,” In Proceedings of the 34th International Conference on Parallel Processing, 2005.

[17] Y. Liu, X. Liu, L. Xiao , L. Ni, and X. Zhang, "Location-Aware Topology Matching in P2P Systems," IEEE INFOCOM, Hong Kong, March 2004.