Priority Based Load Balancing in a Self-Interested P2P Network

Size: px
Start display at page:

Download "Priority Based Load Balancing in a Self-Interested P2P Network"

Transcription

1 Priority Based Load Balancing in a Self-Interested P2P Network Xuan ZHOU and Wolfgang NEJDL L3S Research Center Expo Plaza 1, Hanover, Germany {zhou,nejdl}@l3s.de Abstract. A fundamental issue in P2P networks is that of distributing workload in a balanced way to optimize performance. Unfortunately, optimal load balance is difficult to realize, although it is easy to compute, because participants of P2P networks are usually self-interested and seek to maximize their individual utility without considering system-wide utility. In this paper, we study the influence of selfish behaviors on the load balance in P2P networks, and propose a priority based load balancing scheme to help a P2P network achieve better performance. The scheme is simple and effective, and can be easily used in a P2P environment. Besides presenting the theoretical foundation of our scheme, we also address its major implementation issues and conduct extensive experiments to verify its effectiveness and practicality. 1 Introduction A peer-to-peer network is a network that relies on the collaboration of its participants, who pool their resources to benefit everyone. A fundamental issue in a P2P network is that of distributing workload in a balanced way to optimize its performance. In particular, given a number of peers who are offering the same services, the network needs to find an assignment of jobs to these peers such that the expected response time to these jobs is minimized. While a balanced workload distribution can always be found in principle, it is difficult to be realized in a P2P network where each participant is self-interested and rational, and seeks to minimize the response time to its own jobs without considering the efficiency of the whole network. An example of the load balancing problem in a self-interested P2P network is illustrated in figure 1. Suppose peers A and B are providing the same service in a P2P network, and a number of peers wish to send n jobs to A or B to be processed. Suppose the response time for A to finish a job is always one minute, no matter how many jobs are arriving at A simultaneously. This can be represented by a function of response time, i.e. R(x) = 1, where x denotes the job arrival rate to A. In contrast, the response time for B to finish a job increases linearly with its job arrival rate, and its function of response time is R(x) = x n. If the other peers are selfish and rational, they will choose to send their jobs to B, as it should minimize the response time to their own jobs. In the end, all n jobs are jammed at B, and the average response time to these jobs ends up

2 A R(x) = 1 B R(x) = x/n Fig. 1. An Example of Load Unbalance at one minute. From the system-wide perspective, this is surely not the optimal solution: if n 2 jobs are submitted to A and the other n 2 jobs are submitted to B, the average response time could be as small as 3 4 minute. Unfortunately, no peer is willing to sacrifice by sending his jobs to A. In economics, this phenomenon is known as social dilemma, which describes the situation where all participants seek to maximize their individual utilities, but as a result get utilities which are worse than what they could achieve by cooperation. Selfish behavior in large-scale computer networks has been extensively studied in recent years. Investigated areas include network routing [1 3], resource allocation [4], quality of service [5] and others [6]. Most of this work reveals that the cost of selfish behavior is usually not ignorable. Therefore, a number of strategies, such as pricing [7] and Stackelberg routing [8], have been proposed to cope with selfish behavior and help a network achieve better performance. However, none of these strategies can be directly applied to a P2P network, where centralized control is very difficult to establish. The investigation on uncoordinated load balancing by Subhash Suri et al [9, 10] is the only work addressing the influence of selfish behavior on load balance in P2P networks. In their work they theoretically analyze the influence of selfish behaviors on load balance and give a worst-case bound to the performance loss caused by selfish behavior. However, no proposal is given for controlling selfish behavior to help a P2P network achieve better load balance. In this paper, we introduce a priority based load balancing scheme for a self-interested P2P network. In this scheme, each peer divides its jobs into two classes, and assigns one class a higher priority than the other. By adjusting the proportion between the two classes of jobs, the network can optimize its workload distribution. The scheme is simple, effective, and suitable for P2P environments. We theoretically prove that this scheme can help a P2P network achieve optimal load balance. We also address the various issues of implementing this scheme to a real P2P network, and conduct extensive experiments to demonstrate its effectiveness. The remainder of this paper is organized as follows. Section 2 defines the problem of load balancing in a self-interested P2P network, and discuss the possible solutions. Section 3 introduces our priority based load balancing scheme, and addresses the issues of implementing the scheme in real P2P networks. Section 4 experimentally evaluates the effectiveness and practicality of our scheme. Finally, we conclude this paper in section 5.

3 2 Problem Definition We will use the term uncoordinated load balancing from [9] to describe the problem of balancing workload among selfish peers. This section presents a model of uncoordinated load balancing and discusses the possible approaches to achieve load balance in a self-interested P2P network. 2.1 The Model Consider a P2P network, where n peers are providing the same kind of service, and the other peers periodically request and receive services from the n peers. We call these n peers servers, and denote them by S 0, S 1,..., S n 1. We call all the other peers in the network clients. Each server S i would display different efficiency in processing its received jobs. This efficiency can be expressed by a load-dependent function of response time, which is denoted by R i (x), where x is the job arrival rate (load) on the server. In this paper, we assume that R i (x) is always nondecreasing and convex, as it is regarded true in common real-world systems [3]. Suppose the total job arrival rate to the n servers is λ. Then, the triple (n, R, λ) forms an instance of the uncoordinated load balancing problem. A solution to this problem is an assignment of the total job arrival rate (load) over the n servers, which is denoted by a = (λ 0, λ 1,..., λ n 1 ), where λ i is the job arrival rate to S i and λ 0 + λ λ n 1 = λ. The cost C(a) of a solution a is the total response time of processing the λ jobs. That is, n 1 C(a) = R i (λ i )λ i i=0 When the clients in the P2P network are all selfish, they seek to minimize the response time to their own jobs, without regard to the overall cost of the network. As a result, each job will be sent to the peer that would incur the shortest response time. In the end, the assignment of the job arrival rate λ will end in a Nash equilibrium, where no job can improve its response time by switching to a different server. In this Nash equilibrium, the servers that are assigned a positive workload will display equal response time. This can be formally presented as follows. Lemma 1 A load assignment a = (λ 0, λ 1,..., λ n 1 ) for instance (n, R, λ) is at Nash equilibrium if and only if for every λ i > 0, R i (λ i ) R j (λ j ). According to Roughgarden s results on selfish routing [3], if R i (.) is nondecreasing, there is a unique Nash equilibrium of load assignment for every instance (n, R, λ). The solution at the Nash equilibrium is not necessarily the optimal solution that has the minimal cost C(a). The optimal solution is indeed the solution to the following nonlinear program (NLP) [11]. min : C(a) = n 1 i=0 R i(λ i )λ i subject to: n 1 i=0 λ i = λ, λ i > 0

4 It can be deduced (see [3]) that, given an instance (n, R, λ), where R(.) is convex, the optimal solution should satisfy the following conditions. Lemma 2 Suppose R i (x) = d dx (R i(x)x). A load assignment a = (λ 0, λ 1,..., λ n 1 ) is optimal if and only if for every λ i > 0, R i (λ i) R j (λ j). In other words, C(a) reaches its minimum when the marginal benefit (Ri (λ i)) of decreasing the load on server S i is at most the marginal cost (Rj (λ j)) of increasing the load on any other peer S j. We can apply the model to the example in figure 1. (Suppose n denotes the total job arrival rate instead of the number of arriving jobs.) Through Lemma 1, we find that the load assignment on Servers A and B at the Nash equilibrium is (λ A = 0, λ B = n), and the resulting cost is C(a) = n. Through Lemma 2, we find that the optimal load assignment is (λ A = n 2, λ B = n 2 ), and its cost is C(a) = 3 4 n. The maximal ratio between the cost of Nash equilibrium solution and the cost of the optimal solution is called the price of anarchy. Some earlier work [3] has shown that selfish behavior can have significant influence on the system-wide performance. If we consider the optimal load assignment as the balanced one, then the load assignment at the Nash equilibrium would be unbalanced. The objective of our research is to design a mechanism for the P2P network that could affect the clients selfish behaviors, so that the resulting load assignment will be more balanced. 2.2 Possible Solutions To achieve load balance in a P2P network, obviously we cannot count on the selfish clients to collaboratively distribute their workload in a balanced way. However, we can resort to the cooperativeness of servers. First, the servers have no incentive to disturb the workload distribution, as they benefit nothing from it. Second, as users of a P2P network, they can have improved utility if the network works more efficiently. Therefore, we regard that the servers of a P2P network are usually benign and willing to take action to balance the workload. Under this assumption, a possible approach to load balance is to calculate the proportion of workload on each server in the optimal load assignment, and restrict each server from taking more workload than that proportion. However, as the workload in a P2P network is changing frequently, it is impractical to estimate the optimal load assignment in real time. If the workload is overestimated, the approach will not work. If it is underestimated, some jobs would be starved severely. Therefore, this approach is not an appropriate solution. In this paper, we propose a priority based load balancing approach, which achieves optimal load balance by assigning priorities to workload. 3 Priority Based Load Balancing Section 3.1 introduces the priority based load balancing scheme. Sections 3.2 and 3.3 give its theoretical foundation. Section 3.4 presents implementation issues.

5 3.1 The Scheme In our scheme, each server divides the received jobs into two classes 1st class and 2nd class. The jobs in the 1st class have strictly higher priority than the jobs in the 2nd class, in the sense that a server will always first process the 1st class jobs prior to 2nd class jobs. To differentiate between the two classes of jobs, a server selects a number of clients in the network and assigns each of them a 1st class quota, which specifies the number of 1st class jobs it can submit to the server per time unit. Later on, the jobs arriving from these clients within their 1st class quotas will be regarded by the server as the 1st class jobs, and the other jobs will fall into the 2nd class. Applying this priority based approach to our model, the job arrival rate λ i on each server S i becomes the sum of the arrival rates of the two classes of jobs, which can be represented by λ i = λ 1 i + λ2 i, where λ1 i denotes the arrival rate of the 1st class jobs and λ 2 i denotes the arrival rate of the 2nd class jobs. As we will show subsequently, by setting appropriate 1st class quotas on each server, we can guarantee that, for all the servers, the response time of any 1st class jobs is always shorter than that of 2nd class jobs. Therefore, whenever a client is issuing jobs, it will first use up its 1st class quota, so as to minimize the response time. Hence, λ 1 i is directly determined by the amount of the 1st class quota S i assigns to its clients. At the same time, each client will also seek to minimize the response time to its 2nd class jobs. As a result, the assignment of λ 2 i will end in a Nash equilibrium, where no 2nd class job can improve its response time by switching to a different server. In this scheme, the final load assignment λ 1 i + λ2 i on each server can be adjusted by adjusting the 1st class quotas. The basic idea of our priority based load balancing is to appropriately set the 1st class quotas on each server to optimize the load distribution. In the next section, we prove that we can achieve the optimal load balance through this scheme. 3.2 Finding Optimal Solution Suppose the original function of response time on server S i is R i (λ i ), where λ i is the job arrival rate to the server, i.e. λ i = λ 1 i + λ2 i. As the 1st class jobs have strictly higher priority than the 2nd class jobs, the response time for the 1st class jobs will not be affected by the 2nd class jobs present. Therefore, the response time for the 1st class jobs on server S i will be R i (λ 1 i ) no matter what λ2 i is. At the same time, the 2nd class jobs will bear all additional overhead incurred by the co-existence of the 1st and 2nd classes of jobs. Suppose the response time for the 2nd class jobs is Ri 2, this can be represented by the following equation. R i (λ 1 i + λ 2 i ) (λ 1 i + λ 2 i ) = R i (λ 1 i )λ 1 i + R 2 i λ 2 i Solving this equation, we get the response time function for the 2nd class jobs as R 2 i (λ 1 i, λ 2 i ) = R i(λ 1 i + λ2 i ) (λ1 i + λ2 i ) R i(λ 1 i )λ1 i λ 2 i As we have stated earlier, the arrival rates of the 1st class jobs (λ 1 0, λ 1 1,..., λ 1 n 1) are determined by the 1st class quotas issued by the servers. Meanwhile, each.

6 server will seek to minimize the response time to their 2nd class jobs, so that the arrival rates of the 2nd class jobs (λ 2 0, λ 2 1,..., λ 2 n 1) will end in a Nash Equilibrium, which satisfies the following criterion. Lemma 3 Given an instance (n, R, λ) and the distribution of the 1st class jobs (λ 1 0, λ 1 1,..., λ 1 n 1), a assignment of the 2nd class jobs (λ 2 0, λ 2 1,..., λ 2 n 1) is at Nash equilibrium if and only if for every λ 2 i > 0, R2 i (λ1 i, λ2 i ) R2 j (λ1 j, λ2 j ). The optimal load assignment is the assignment that incurs the smallest overall cost C(a) = n 1 i=0 R i(λ 1 i + λ2 i )(λ1 i + λ2 i ). Following the spirit of Lemma 2, the criterion of the optimal solution is: Lemma 4 Suppose R i (x) = d dx (R i(x)x). A load assignment a = (λ λ 2 0, λ λ 2 0,..., λ 1 n 1 + λ 2 0) is optimal if and only if for every λ 1 i + λ2 i > 0, R i (λ1 i + λ2 i ) R j (λ1 j + λ2 j ). The objective of our scheme is to find a assignment of the 1st class jobs (λ 1 0, λ 1 1,..., λ 1 n 1), such that the load assignment at the Nash Equilibrium is exactly the optimal load assignment. The objective can be formally represented as the following problem: find λ 1 0, λ 1 1,..., λ 1 n 1 n 1 that satisfy 1. i=0 (λ1 i + λ2 i ) = λ 2. λ 1 i 0 and λ2 i 0 3. Ri 2(λ1 i, λ2 i ) R2 j (λ1 j, λ2 j ) if λ2 i > 0 4. Ri (λ1 i + λ2 i ) R j (λ1 j + λ2 j ) if λ1 i + λ2 i > 0 (1) y R i(x)x R j(x)x b i b j aj ai 1 i 2 i Load 1 j 2 j Fig. 2. Problem Interpretation Figure 2 plots the curves of the two functions R i (x)x and R j (x)x, where R i (x) and R j (x) are the response time functions of S i and S j respectively. Four angles, which are determined by the job arrival rates λ 1 i, λ2 i, λ1 j and λ2 j, are drawn out in figure 2. a i is the slope angle of the line connecting ( λ 1 i, R ) ( i(λ 1 i ) λ1 i and λ 1 i + λ 2 i, R i(λ 1 i + λ2 i ) (λ1 i + λ2 i )). a j is the slope angle of the line connecting

7 ( λ 1 j, R j (λ 1 j ) ) ( λ1 j and λ 1 j + λ 2 j, R i(λ 1 j + λ2 j ) (λ1 j + λ2 j )). b i is the slope angle of R i (x)x s tangent at point ( λ 1 i + λ2 i, R i(λ 1 i + λ2 i ) (λ1 i + λ2 i )). b j is the slope angle of R j (x)x s tangent at point ( λ 1 j + λ2 j, R i(λ 1 j + λ2 j ) (λ1 j + λ2 j )). As tan(a i ) = Ri 2(λ1 i, λ2 i ) and tan(a j) = Rj 2(λ1 j, λ2 j ), condition 3 could be represented as a i a j. As tan(b i ) = Ri (λ1 i + λ2 i ) and tan(b j) = Rj (λ1 j + λ2 j ), condition 4 could be represented as b i b j. Then, the above problem could be interpreted as: find λ 1 0, λ 1 1,..., λ 1 n 1 n 1 that satisfy 1. i=0 (λ1 i + λ2 i ) = λ 2. λ 1 i 0 and λ2 i 0 3. a i a j if λ 2 i > 0 4. b i b j if λ 1 i + λ2 i > 0 To solve this problem, we proceed as follows: 1. Determine λ i = λ 1 i + λ2 i : Find a load assignment a = (λ 0, λ 1,..., λ n 1 ) that satisfies: (a) n 1 0 λ i = λ; (b) λ i 0 for all i; (c) b i b j if λ i > Determine a i : Select the minimum a that satisfies: (a) a < b i for all i, (This guarantees that λ 2 i > 0); (b) If λ i > 0, then tan a > R i (λ i ). Then, set a i = a for all λ i > Determine λ 1 i : If λ i > 0, λ 1 i is uniquely determined by a i and λ i ; if λ i = 0, set λ 1 i = 0. As function R i (x) is nondecreasing and convex, function R i (x)x must be nondecreasing and convex too. Then, we can prove that the above three steps can always find a solution to Problem (2), as well as Problem (1). With the resulting λ 1 0, λ 1 1,..., λ 1 n 1, we can ensure that the load assignment at the Nash Equilibrium is exactly the optimal load assignment. Then the 1st class quotas on each server can be set accordingly to enforce the λ 1 0, λ 1 1,..., λ 1 n 1, so that the P2P network can achieve optimal load balance. In addition, step 2 guarantees that if λ 1 i > 0 then λ2 i > 0. That is to say, if a server receives 1st class jobs, it will certainly also receive 2nd class jobs. As all the 2nd class jobs have the same response time at the Nash equilibrium, we can guarantee that the response time of any 1st class job is shorter than that of the 2nd class jobs. This justifies our previous assumption that a client would always first use up its 1st class quota in order to minimize the received response time. 3.3 Tuning Method While we have theoretically proved that our scheme can achieve optimal load balance, it is still difficult to realize the scheme in a practical P2P system. On the one hand, it is infeasible to accurately estimate the workload in a P2P network in real time. On the other hand, it is very costly to obtain the response time functions of all peers. Both make it difficult to compute the optimal load assignment. Thus, a more practical strategy of load balancing can be composed of two steps. First, when a P2P network starts or when a new server joins the network, the server contacts some other servers to initialize its 1st class quota. (2)

8 y R (x)x i L b L a 1 2 i a i a Load i 1 b 2 i b Fig. 3. Tuning Strategy Second, at running time each server continually observes the network and tunes its workload, to draw the system-wide load assignment closer to the optimal one. During the tuning step, each peer can increase or decrease its workload by decreasing or increasing the 1st class quota assigned to its clients. This is illustrated in figure 3, which plots the curve of function R i (x)x, where R i (x) is the response time function of S i. (R i (x)x is always nondecreasing and convex.) The two points (λ 1 i, R i(λ 1 i ) λ1 i ) and (λ1 i + λ2 i, R i(λ 1 i + λ2 i ) (λ1 i + λ2 i )) on the curve are connected by a line L. The slope angle of L is actually a i in figure 2. According to the condition of the Nash equilibrium in Lemma 3, the slope of L should be the same for every R i (x)x, as long as λ 2 i > 0. For this reason, if we decrease the arrival rate of the 1st class jobs on S i from λ 1 ia to λ1 ib, line L will move parallel from L a to L b. Then, the total job arrival rate will be increased from λ 1 ia + λ2 ia to λ1 ib + λ1 ib. This reveals the following intuitive principles of load assignment tuning: 1. When a server decreases its 1st class quota, it shortens the response time of its 2nd class jobs Ri 2(λ1 i, λ2 i ), so that the server will attract more workload λ 1 i + λ2 i ; 2. When a server increases its 1st class quota, it enlarges the response time of its 2nd class jobs Ri 2(λ1 i, λ2 i ), so that the server will get less workload λ1 i +λ2 i. Such tuning is limited by an upper bound and a lower bound. Namely, when λ 1 i = 0, the server will receive the maximum workload; when λ1 i = λ i, the server will receive the minimum workload. In practice, a server can probe the status of some other peers in the network to assess whether itself is overloaded or under-loaded, and tune its workload accordingly. 3.4 Implementation Issues Figure 4 shows the architecture of a server in a P2P network. When a job is sent to the server, a router decides whether the job is in the 1st class or in the 2nd class, and routes the job to the corresponding queue. The processor always

9 Statistics 1st Class Quota Load Manager 1st Class Queue Jobs Router 2nd Class Queue Processor Fig. 4. Architecture first processes the jobs in the 1st class queue. Only when the 1st class queue is empty, it turns to process the jobs in the 2nd class queue. To prevent starvation, some 2nd class jobs will be upgraded to the 1st class if their waiting time exceed a certain threshold. However, the system should ensure that the efficiency of processing the 1st class of jobs is least affected by the 2nd class jobs. The load manager is the key component responsible for monitoring and adjusting the workload of the server. It continuously collects statistic information to learn the workload distribution in the network, and correspondingly adjusts the workload of its host server to help improve the load balance of the network. Meanwhile, it also monitors the status of its host server, and reports the information to the network. In order to achieve load balance, a server needs to assign a proper amount of 1st class quota to its clients. During the quota assignment, the server issues quota certificates to some randomly selected clients. Each quota certificate contains the identity of the target client, the time of expiration, and the quota indicating the number of 1st class jobs the client can submit in a unit time. To protect its integrity, each certificate is signed by the server s private key. When a client submits a 1st class job to the server, it attaches the quota certificate to the job, so that the router can classify it into the 1st class. If the 1st class jobs from a client exceed its quota, the extra jobs will be classified into the 2nd class. Initializing 1st Class Quota: When a P2P network is started or when a new server joins the network, the server needs to correctly initialize its 1st class quota, so as to attract an appropriate amount of workload. As it is infeasible to accurately compute the optimal load assignment of a P2P network, the initial 1st class quota is estimated through sampling, which can performed through the following steps: 1. Choose a set of sample servers, and compute the optimal load assignment over the sample servers through the process introduced in Section Estimate the optimal workload on the new server based on the load assignment over the sample servers. 3. The initial 1st class quota of the new server can be estimated through its optimal workload and the load distribution over the sample servers. Starting from this initial quota, the server can later tune its workload to further improve the load balance of the network.

10 Load Assignment Tuning: In load assignment tuning, a server probes the status of other servers to learn whether itself is overloaded or under-loaded, and increases or decreases its 1st class quota correspondingly. The information obtained in each probe will include the response time function and the current workload of each visited server. A server can conduct tuning through the following steps. 1. Calculate the local optimal load assignment over the group of servers that comprise all the probed servers and itself. 2. Compare its current workload against the optimal one obtained in step Tune its 1st class quota through the method introduced in section 3.3, such that its workload reaches the optimal workload. Intuitively, when each server keeps tuning their 1st class quotas, the load assignment in the P2P network will gradually converge to the optimal solution. This will be experimentally verified in the next section. 4 Experimental Evaluation We evaluate the proposed load balancing scheme through simulation. We simulate a P2P network that is composed by millions of participants, in which 1000 participants are offering the same service and are considered as servers. The other participants can select any server to submit their jobs. The response time function on each server follows that of M/M/1 queueing system. Namely, it is in the form 1 µ λ, where µ denotes the service rate of a server and λ denotes the job arrival rate to the server. The distribution of the service rates µ on the 1000 servers follows the Pareto distribution [12]. Due to limited space, only a part of the experiment results are reported. 4.1 Initial Load Assignment The goal of the first set of experiments is to study how much the initial load assignment of our load balancing scheme could improve the performance of the selfish P2P network. The process of initialization follows that in section 3.4. In the first experiment, we fix the sample size at 20 servers. We vary the workload on the whole network from 20% to 80%, and measure the average response time resulted by the initial load assignment. We also compare this response time against that of the selfish network without using our scheme and that of the cooperative network where optimal load assignment is achieved. The results are shown in figure 5 (a). As expected, the initial load assignment by our load balancing scheme could remarkably improve the performance of the selfish P2P network (by 10% to 30%). In the second experiment, we fix the workload of the network at 50%, and vary the sample size from 5 servers to 160 servers. Then we measure the average response time resulting from the initial load assignment. Figure 5 (b) shows the results. As expected, a larger sample could enable the servers to obtain a more complete overview of the P2P network, so that the servers could estimate their 1st class quotas more accurately and achieve better load balance. When the sample increases to 100 servers, the initial load assignment almost reaches the optimal load assignment.

11 2 0.7 response time (s) No Balancing With Balancing Cooperative (optimal) response time (s) No Balancing With Balancing Cooperative (optimal) load (%) (a) varying load sample size (b) varying sample size Load Assignment Tuning Fig. 5. Effectiveness of Initial Load Assignment This set of experiments is intended to study how the load assignment tuning of our scheme could improve the load balance of the simulated P2P network. When conducting tuning, a server periodically probes some other servers to assess whether itself is overloaded or under-loaded, and tunes its 1st class quota correspondingly. In the first experiment, we do not use the initial load assignment, but set all the 1st class quotas to 0. Then, we conduct load assignment tuning on each server by repeating the process in section 3.4. The tuning is performed round after round. In each round, each server tunes its workload once. After each round, we set the load assignment to the Nash equilibrium defined by Lemma 3, and measure the average response time to accomplish a job. The probe size of the tuning varies from 2 servers to 8 servers. Figure 6 (a) plots the varying of average response time when the tuning proceeds. We can see that the load assignment obviously converges to the optimal one as the tuning going on. In addition, a larger probe size amplifies the effects of tuning with the probe size of 8 servers, the convergence could be twice as fast as with probe size of 2 servers. In the second experiment, we study the effects of combining initial load assignment and tuning. We fix the probe size of tuning at 2 servers, and vary the response time (s) No Balancing Cooperative (optimal) probe size = 2 probe size = 4 probe size = 8 response time (s) No Balancing Cooperative (optimal) No initial initial size = 5 initial size = 20 initial size = number of rounds (a) Without Initialization number of rounds (b) With Initialization Fig. 6. Effectiveness of Load Assignment Tuning

12 sample size of the initial load assignment from 5 servers to 100 servers. Then, we measure the average response time after each round of tuning. The results are shown in figure 6 (b). As expected, when the initial load assignment is used, the load assignment could converge much faster to the optimal load assignment. The improvement could be further magnified by using a larger sample size for initial load assignment. 5 Conclusion In this paper, we proposed a priority based load balancing scheme for a selfinterested P2P network, in which each participant is selfish and seeks to minimize its individual response time without considering the system-wide performance. We theoretically proved the correctness of our scheme, and conducted experiments to show that our scheme helps P2P networks to achieve better load balance in an efficient way. References 1. Koutsoupias, E., Papadimitriou, C.: Worst-case equilibria. In: 16th STACS. (1999) Czumaj, A., Vocking, B.: Tight bounds for worstcase equilibria. In: 13th SODA. (2002) Roughgarden, T., Tardos, E.: How bad is selfish routing. Journal of ACM 49 (2002) Lazar, A., Orda, A., Pendarakis, D.: Virtual path bandwidth allocation in multiuser networks. IEEE/ACM Transactions on Networking 5 (1997) Yaiche, H., Mazumdar, R., Rosenberg, C.: A game theoretic framework for bandwidth allocation and pricing of elastic connections in broadband networks: theory and algorithms. IEEE/ACM Transactions on Networking 8 (2000) Altman, E., Boulogne, T., Azouzi, R., Jimenez, T.: A survey on networking games in telecommunications. Computers and Operations Research 33 (2006) Cole, R., Dodis, Y., Roughgarden, T.: Pricing network edges for heterogeneous selfish user. In: 35th STOC. (2003) Roughgarden, T.: Stackelberg scheduling strategies. SIAM Journal of Computing 33 (2003) Suri, S., Toth, C., Zhou, Y.: Uncoordinated load balancing and congestion games in p2p systems. In: 3rd IPTPS. (2004) Suri, S., Toth, C.D., Zhou, Y.: Selfish load balancing and atomic congestion games. In: SPAA 04: Proceedings of the sixteenth annual ACM symposium on Parallelism in algorithms and architectures. (2004) Peressini, A.L., Sullivan, F.E., J. J. Uhl, J.: The Mathematics of Nonlinear Programming. Springer-Verlag (1993) 12. Adamic, L.A.: (Zipf, power-laws, and pareto - a ranking tutorial)

A Game-Theoretic Model and Algorithm for Load Balancing in Distributed Systems

A Game-Theoretic Model and Algorithm for Load Balancing in Distributed Systems In Proc of the 6th IEEE International Parallel and Distributed Processing Symposium (IPDPS 2002, Workshop on Advances in Parallel and Distributed Computational Models (APDCM 02, April 5, 2002, Fort Lauderdale,

More information

Tight Bounds for Selfish and Greedy Load Balancing

Tight Bounds for Selfish and Greedy Load Balancing Tight Bounds for Selfish and Greedy Load Balancing Ioannis Caragiannis 1, Michele Flammini, Christos Kaklamanis 1, Panagiotis Kanellopoulos 1, and Luca Moscardelli 1 Research Academic Computer Technology

More information

BRAESS-LIKE PARADOXES FOR NON-COOPERATIVE DYNAMIC LOAD BALANCING IN DISTRIBUTED COMPUTER SYSTEMS

BRAESS-LIKE PARADOXES FOR NON-COOPERATIVE DYNAMIC LOAD BALANCING IN DISTRIBUTED COMPUTER SYSTEMS GESJ: Computer Science and Telecommunications 21 No.3(26) BRAESS-LIKE PARADOXES FOR NON-COOPERATIVE DYNAMIC LOAD BALANCING IN DISTRIBUTED COMPUTER SYSTEMS Said Fathy El-Zoghdy Department of Computer Science,

More information

Pareto Set, Fairness, and Nash Equilibrium: A Case Study on Load Balancing

Pareto Set, Fairness, and Nash Equilibrium: A Case Study on Load Balancing Pareto Set, Fairness, and Nash Equilibrium: A Case Study on Load Balancing Atsushi Inoie, Hisao Kameda, Corinne Touati Graduate School of Systems and Information Engineering University of Tsukuba, Tsukuba

More information

Mediated Equilibria in Load-Balancing Games

Mediated Equilibria in Load-Balancing Games Mediated Equilibria in Load-Balancing Games Joshua R. Davis, David Liben-Nowell, Alexa Sharp, and Tom Wexler Carleton College; Northfield, MN Oberlin College; Oberlin, OH joshuadavis@q.com, dlibenno@carleton.edu,

More information

On the Interaction and Competition among Internet Service Providers

On the Interaction and Competition among Internet Service Providers On the Interaction and Competition among Internet Service Providers Sam C.M. Lee John C.S. Lui + Abstract The current Internet architecture comprises of different privately owned Internet service providers

More information

Multi-service Load Balancing in a Heterogeneous Network with Vertical Handover

Multi-service Load Balancing in a Heterogeneous Network with Vertical Handover 1 Multi-service Load Balancing in a Heterogeneous Network with Vertical Handover Jie Xu, Member, IEEE, Yuming Jiang, Member, IEEE, and Andrew Perkis, Member, IEEE Abstract In this paper we investigate

More information

Price of Anarchy in Non-Cooperative Load Balancing

Price of Anarchy in Non-Cooperative Load Balancing Price of Anarchy in Non-Cooperative Load Balancing U Ayesta 1,3, O Brun 2, BJ Prabhu 2 1 BCAM, Basque Center for Applied Mathematics, 48160 Derio, Spain 2 CNRS ; LAAS ; 7 avenue du colonel Roche, 31077

More information

Keywords: Dynamic Load Balancing, Process Migration, Load Indices, Threshold Level, Response Time, Process Age.

Keywords: Dynamic Load Balancing, Process Migration, Load Indices, Threshold Level, Response Time, Process Age. Volume 3, Issue 10, October 2013 ISSN: 2277 128X International Journal of Advanced Research in Computer Science and Software Engineering Research Paper Available online at: www.ijarcsse.com Load Measurement

More information

Load Balancing in Distributed Systems: An Approach Using Cooperative Games

Load Balancing in Distributed Systems: An Approach Using Cooperative Games In Proc of the 6th IEEE International Parallel and Distributed Processing Symposium (IPDPS 2002), April 5 9, 2002, Fort Lauderdale, Florida, USA, IEEE Computer Society Press, pp 50 50 Load Balancing in

More information

FOR decades, it has been the responsibility of the network

FOR decades, it has been the responsibility of the network IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 14, NO. 4, AUGUST 2006 725 On Selfish Routing in Internet-Like Environments Lili Qiu, Yang Richard Yang, Yin Zhang, and Scott Shenker Abstract A recent trend in

More information

Compact Representations and Approximations for Compuation in Games

Compact Representations and Approximations for Compuation in Games Compact Representations and Approximations for Compuation in Games Kevin Swersky April 23, 2008 Abstract Compact representations have recently been developed as a way of both encoding the strategic interactions

More information

A Load Balancing Mechanism with Verification

A Load Balancing Mechanism with Verification A Load Balancing Mechanism with Verification Daniel Grosu and Anthony T. Chronopoulos Department of Computer Science, University of Texas at San Antonio, 6900 N. Loop 1604 West, San Antonio, TX 78249 dgrosu,

More information

Workload Factoring with the Cloud: A Game-Theoretic Perspective

Workload Factoring with the Cloud: A Game-Theoretic Perspective Workload Factoring with the Cloud: A Game-Theoretic Perspective Amir Nahir Department of Computer Science Technion Israel Institue of Technology Haifa 32000 Israel Email: nahira@cstechnionacil Ariel Orda

More information

Noncooperative load balancing in distributed systems

Noncooperative load balancing in distributed systems J. Parallel Distrib. Comput. 65 (2005) 022 034 www.elsevier.com/locate/pdc Noncooperative load balancing in distributed systems Daniel Grosu a, Anthony T. Chronopoulos b, a Department of Computer Science,

More information

Game Theory Based Load Balanced Job Allocation in Distributed Systems

Game Theory Based Load Balanced Job Allocation in Distributed Systems in Distributed Systems Anthony T. Chronopoulos Department of Computer Science University of Texas at San Antonio San Antonio, TX, USA atc@cs.utsa.edu Load balancing: problem formulation Load balancing

More information

Load Balancing and Switch Scheduling

Load Balancing and Switch Scheduling EE384Y Project Final Report Load Balancing and Switch Scheduling Xiangheng Liu Department of Electrical Engineering Stanford University, Stanford CA 94305 Email: liuxh@systems.stanford.edu Abstract Load

More information

Algorithmic Mechanism Design for Load Balancing in Distributed Systems

Algorithmic Mechanism Design for Load Balancing in Distributed Systems In Proc. of the 4th IEEE International Conference on Cluster Computing (CLUSTER 2002), September 24 26, 2002, Chicago, Illinois, USA, IEEE Computer Society Press, pp. 445 450. Algorithmic Mechanism Design

More information

Competitive Analysis of On line Randomized Call Control in Cellular Networks

Competitive Analysis of On line Randomized Call Control in Cellular Networks Competitive Analysis of On line Randomized Call Control in Cellular Networks Ioannis Caragiannis Christos Kaklamanis Evi Papaioannou Abstract In this paper we address an important communication issue arising

More information

20 Selfish Load Balancing

20 Selfish Load Balancing 20 Selfish Load Balancing Berthold Vöcking Abstract Suppose that a set of weighted tasks shall be assigned to a set of machines with possibly different speeds such that the load is distributed evenly among

More information

Simple Channel-Change Games for Spectrum- Agile Wireless Networks

Simple Channel-Change Games for Spectrum- Agile Wireless Networks Simple Channel-Change Games for Spectrum- Agile Wireless Networks Roli G. Wendorf and Howard Blum Seidenberg School of Computer Science and Information Systems Pace University White Plains, New York, USA

More information

The Relative Worst Order Ratio for On-Line Algorithms

The Relative Worst Order Ratio for On-Line Algorithms The Relative Worst Order Ratio for On-Line Algorithms Joan Boyar 1 and Lene M. Favrholdt 2 1 Department of Mathematics and Computer Science, University of Southern Denmark, Odense, Denmark, joan@imada.sdu.dk

More information

CMSC 858T: Randomized Algorithms Spring 2003 Handout 8: The Local Lemma

CMSC 858T: Randomized Algorithms Spring 2003 Handout 8: The Local Lemma CMSC 858T: Randomized Algorithms Spring 2003 Handout 8: The Local Lemma Please Note: The references at the end are given for extra reading if you are interested in exploring these ideas further. You are

More information

International journal of Engineering Research-Online A Peer Reviewed International Journal Articles available online http://www.ijoer.

International journal of Engineering Research-Online A Peer Reviewed International Journal Articles available online http://www.ijoer. RESEARCH ARTICLE ISSN: 2321-7758 GLOBAL LOAD DISTRIBUTION USING SKIP GRAPH, BATON AND CHORD J.K.JEEVITHA, B.KARTHIKA* Information Technology,PSNA College of Engineering & Technology, Dindigul, India Article

More information

Price of Anarchy in Non-Cooperative Load Balancing

Price of Anarchy in Non-Cooperative Load Balancing Price of Anarchy in Non-Cooperative Load Balancing Urtzi Ayesta, Olivier Brun, Balakrishna Prabhu To cite this version: Urtzi Ayesta, Olivier Brun, Balakrishna Prabhu. Price of Anarchy in Non-Cooperative

More information

LOAD BALANCING MECHANISMS IN DATA CENTER NETWORKS

LOAD BALANCING MECHANISMS IN DATA CENTER NETWORKS LOAD BALANCING Load Balancing Mechanisms in Data Center Networks Load balancing vs. distributed rate limiting: an unifying framework for cloud control Load Balancing for Internet Distributed Services using

More information

Decentralized Utility-based Sensor Network Design

Decentralized Utility-based Sensor Network Design Decentralized Utility-based Sensor Network Design Narayanan Sadagopan and Bhaskar Krishnamachari University of Southern California, Los Angeles, CA 90089-0781, USA narayans@cs.usc.edu, bkrishna@usc.edu

More information

The Economics of the Cloud: Price Competition and Congestion

The Economics of the Cloud: Price Competition and Congestion The Economics of the Cloud: Price Competition Congestion JONATHA ANSELMI Basque Center for Applied Mathematics BCAM DANILO ARDAGNA Dip. di Elettronica e Informazione, Politecnico di Milano JOHN C.S. LUI

More information

Unraveling versus Unraveling: A Memo on Competitive Equilibriums and Trade in Insurance Markets

Unraveling versus Unraveling: A Memo on Competitive Equilibriums and Trade in Insurance Markets Unraveling versus Unraveling: A Memo on Competitive Equilibriums and Trade in Insurance Markets Nathaniel Hendren January, 2014 Abstract Both Akerlof (1970) and Rothschild and Stiglitz (1976) show that

More information

Load Balancing. Load Balancing 1 / 24

Load Balancing. Load Balancing 1 / 24 Load Balancing Backtracking, branch & bound and alpha-beta pruning: how to assign work to idle processes without much communication? Additionally for alpha-beta pruning: implementing the young-brothers-wait

More information

Differenciated Bandwidth Allocation in P2P Layered Streaming

Differenciated Bandwidth Allocation in P2P Layered Streaming Differenciated Bandwidth Allocation in P2P Layered Streaming Abbas Bradai, Toufik Ahmed CNRS-LaBRI University of Bordeaux- 5, Cours de la libération. Talence, 45 {bradai, tad} @labri.fr Abstract There

More information

Router Scheduling Configuration Based on the Maximization of Benefit and Carried Best Effort Traffic

Router Scheduling Configuration Based on the Maximization of Benefit and Carried Best Effort Traffic Telecommunication Systems 24:2 4, 275 292, 2003 2003 Kluwer Academic Publishers. Manufactured in The Netherlands. Router Scheduling Configuration Based on the Maximization of Benefit and Carried Best Effort

More information

On Selfish Routing in Internet-Like Environments

On Selfish Routing in Internet-Like Environments 1 On Selfish Routing in Internet-Like Environments Lili Qiu Yang Richard Yang Yin Zhang Scott Shenker UT Austin Yale University UT Austin UC Berkeley lili@cs.utexas.edu yry@cs.yale.edu yzhang@cs.utexas.edu

More information

A New Nature-inspired Algorithm for Load Balancing

A New Nature-inspired Algorithm for Load Balancing A New Nature-inspired Algorithm for Load Balancing Xiang Feng East China University of Science and Technology Shanghai, China 200237 Email: xfeng{@ecusteducn, @cshkuhk} Francis CM Lau The University of

More information

How To Balance In A Distributed System

How To Balance In A Distributed System 6 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 11, NO. 1, JANUARY 2000 How Useful Is Old Information? Michael Mitzenmacher AbstractÐWe consider the problem of load balancing in dynamic distributed

More information

Moral Hazard. Itay Goldstein. Wharton School, University of Pennsylvania

Moral Hazard. Itay Goldstein. Wharton School, University of Pennsylvania Moral Hazard Itay Goldstein Wharton School, University of Pennsylvania 1 Principal-Agent Problem Basic problem in corporate finance: separation of ownership and control: o The owners of the firm are typically

More information

Routing and Peering in a Competitive Internet

Routing and Peering in a Competitive Internet Routing and Peering in a Competitive Internet Ramesh Johari Stanford University rjohari@stanford.edu John N. Tsitsiklis MIT jnt@mit.edu Abstract Today s Internet is a loose federation of independent network

More information

SIMULATION OF LOAD BALANCING ALGORITHMS: A Comparative Study

SIMULATION OF LOAD BALANCING ALGORITHMS: A Comparative Study SIMULATION OF LOAD BALANCING ALGORITHMS: A Comparative Study Milan E. Soklic Abstract This article introduces a new load balancing algorithm, called diffusive load balancing, and compares its performance

More information

A Dynamic Load Balancing Algorithm For Web Applications

A Dynamic Load Balancing Algorithm For Web Applications Computing For Nation Development, February 25 26, 2010 Bharati Vidyapeeth s Institute of Computer Applications and Management, New Delhi A Dynamic Load Balancing Algorithm For Web Applications 1 Sameena

More information

8 Modeling network traffic using game theory

8 Modeling network traffic using game theory 8 Modeling network traffic using game theory Network represented as a weighted graph; each edge has a designated travel time that may depend on the amount of traffic it contains (some edges sensitive to

More information

Bargaining Solutions in a Social Network

Bargaining Solutions in a Social Network Bargaining Solutions in a Social Network Tanmoy Chakraborty and Michael Kearns Department of Computer and Information Science University of Pennsylvania Abstract. We study the concept of bargaining solutions,

More information

Scheduling Allowance Adaptability in Load Balancing technique for Distributed Systems

Scheduling Allowance Adaptability in Load Balancing technique for Distributed Systems Scheduling Allowance Adaptability in Load Balancing technique for Distributed Systems G.Rajina #1, P.Nagaraju #2 #1 M.Tech, Computer Science Engineering, TallaPadmavathi Engineering College, Warangal,

More information

A Hybrid Scheduling Approach for Scalable Heterogeneous Hadoop Systems

A Hybrid Scheduling Approach for Scalable Heterogeneous Hadoop Systems A Hybrid Scheduling Approach for Scalable Heterogeneous Hadoop Systems Aysan Rasooli Department of Computing and Software McMaster University Hamilton, Canada Email: rasooa@mcmaster.ca Douglas G. Down

More information

A Game-Theoretic Analysis to Resolve the Tussle in Cloud Storage Services

A Game-Theoretic Analysis to Resolve the Tussle in Cloud Storage Services Appl. Math. Inf. Sci. 8, No. 3, 1361-1367 (214) 1361 Applied Mathematics & Information Sciences An International Journal http://dx.doi.org/1.12785/amis/8353 A Game-Theoretic Analysis to Resolve the Tussle

More information

Game Theory: Supermodular Games 1

Game Theory: Supermodular Games 1 Game Theory: Supermodular Games 1 Christoph Schottmüller 1 License: CC Attribution ShareAlike 4.0 1 / 22 Outline 1 Introduction 2 Model 3 Revision questions and exercises 2 / 22 Motivation I several solution

More information

Fairness in Routing and Load Balancing

Fairness in Routing and Load Balancing Fairness in Routing and Load Balancing Jon Kleinberg Yuval Rabani Éva Tardos Abstract We consider the issue of network routing subject to explicit fairness conditions. The optimization of fairness criteria

More information

Shareability and Locality Aware Scheduling Algorithm in Hadoop for Mobile Cloud Computing

Shareability and Locality Aware Scheduling Algorithm in Hadoop for Mobile Cloud Computing Shareability and Locality Aware Scheduling Algorithm in Hadoop for Mobile Cloud Computing Hsin-Wen Wei 1,2, Che-Wei Hsu 2, Tin-Yu Wu 3, Wei-Tsong Lee 1 1 Department of Electrical Engineering, Tamkang University

More information

Resource Allocation Schemes for Gang Scheduling

Resource Allocation Schemes for Gang Scheduling Resource Allocation Schemes for Gang Scheduling B. B. Zhou School of Computing and Mathematics Deakin University Geelong, VIC 327, Australia D. Walsh R. P. Brent Department of Computer Science Australian

More information

Stochastic Inventory Control

Stochastic Inventory Control Chapter 3 Stochastic Inventory Control 1 In this chapter, we consider in much greater details certain dynamic inventory control problems of the type already encountered in section 1.3. In addition to the

More information

A Game Theoretical Framework on Intrusion Detection in Heterogeneous Networks Lin Chen, Member, IEEE, and Jean Leneutre

A Game Theoretical Framework on Intrusion Detection in Heterogeneous Networks Lin Chen, Member, IEEE, and Jean Leneutre IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL 4, NO 2, JUNE 2009 165 A Game Theoretical Framework on Intrusion Detection in Heterogeneous Networks Lin Chen, Member, IEEE, and Jean Leneutre

More information

Week 7 - Game Theory and Industrial Organisation

Week 7 - Game Theory and Industrial Organisation Week 7 - Game Theory and Industrial Organisation The Cournot and Bertrand models are the two basic templates for models of oligopoly; industry structures with a small number of firms. There are a number

More information

A New Forwarding Policy for Load Balancing in Communication Networks

A New Forwarding Policy for Load Balancing in Communication Networks A New Forwarding Policy for Load Balancing in Communication Networks Martin Heusse Yvon Kermarrec ENST de Bretagne BP 83, 985 Brest Cedex, France Martin.Heusse@enst-bretagne.fr Abstract We present in this

More information

Two Papers on Internet Connectivity and Quality. Abstract

Two Papers on Internet Connectivity and Quality. Abstract Two Papers on Internet Connectivity and Quality ROBERTO ROSON Dipartimento di Scienze Economiche, Università Ca Foscari di Venezia, Venice, Italy. Abstract I review two papers, addressing the issue of

More information

Performance Issues of a Web Database

Performance Issues of a Web Database Performance Issues of a Web Database Yi Li, Kevin Lü School of Computing, Information Systems and Mathematics South Bank University 103 Borough Road, London SE1 0AA {liy, lukj}@sbu.ac.uk Abstract. Web

More information

Approximated Distributed Minimum Vertex Cover Algorithms for Bounded Degree Graphs

Approximated Distributed Minimum Vertex Cover Algorithms for Bounded Degree Graphs Approximated Distributed Minimum Vertex Cover Algorithms for Bounded Degree Graphs Yong Zhang 1.2, Francis Y.L. Chin 2, and Hing-Fung Ting 2 1 College of Mathematics and Computer Science, Hebei University,

More information

Pareto Efficiency and Approximate Pareto Efficiency in Routing and Load Balancing Games

Pareto Efficiency and Approximate Pareto Efficiency in Routing and Load Balancing Games Pareto Efficiency and Approximate Pareto Efficiency in Routing and Load Balancing Games Yonatan Aumann Yair Dombb Abstract We analyze the Pareto efficiency, or inefficiency, of solutions to routing games

More information

Distributed Dynamic Load Balancing for Iterative-Stencil Applications

Distributed Dynamic Load Balancing for Iterative-Stencil Applications Distributed Dynamic Load Balancing for Iterative-Stencil Applications G. Dethier 1, P. Marchot 2 and P.A. de Marneffe 1 1 EECS Department, University of Liege, Belgium 2 Chemical Engineering Department,

More information

Minimizing Probing Cost and Achieving Identifiability in Probe Based Network Link Monitoring

Minimizing Probing Cost and Achieving Identifiability in Probe Based Network Link Monitoring Minimizing Probing Cost and Achieving Identifiability in Probe Based Network Link Monitoring Qiang Zheng, Student Member, IEEE, and Guohong Cao, Fellow, IEEE Department of Computer Science and Engineering

More information

Analysis of Approximation Algorithms for k-set Cover using Factor-Revealing Linear Programs

Analysis of Approximation Algorithms for k-set Cover using Factor-Revealing Linear Programs Analysis of Approximation Algorithms for k-set Cover using Factor-Revealing Linear Programs Stavros Athanassopoulos, Ioannis Caragiannis, and Christos Kaklamanis Research Academic Computer Technology Institute

More information

OPRE 6201 : 2. Simplex Method

OPRE 6201 : 2. Simplex Method OPRE 6201 : 2. Simplex Method 1 The Graphical Method: An Example Consider the following linear program: Max 4x 1 +3x 2 Subject to: 2x 1 +3x 2 6 (1) 3x 1 +2x 2 3 (2) 2x 2 5 (3) 2x 1 +x 2 4 (4) x 1, x 2

More information

Network Formation and Routing by Strategic Agents using Local Contracts

Network Formation and Routing by Strategic Agents using Local Contracts Network Formation and Routing by Strategic Agents using Local Contracts Elliot Anshelevich 1 and Gordon Wilfong 2 1 Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY. 2 Bell Labs,

More information

IN CURRENT distributed systems such as computational

IN CURRENT distributed systems such as computational IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART B: CYBERNETICS, VOL. 34, NO. 1, FEBRUARY 2004 77 Algorithmic Mechanism Design for Load Balancing in Distributed Systems Daniel Grosu, Student Member,

More information

Load Balancing in MapReduce Based on Scalable Cardinality Estimates

Load Balancing in MapReduce Based on Scalable Cardinality Estimates Load Balancing in MapReduce Based on Scalable Cardinality Estimates Benjamin Gufler 1, Nikolaus Augsten #, Angelika Reiser 3, Alfons Kemper 4 Technische Universität München Boltzmannstraße 3, 85748 Garching

More information

arxiv:1112.0829v1 [math.pr] 5 Dec 2011

arxiv:1112.0829v1 [math.pr] 5 Dec 2011 How Not to Win a Million Dollars: A Counterexample to a Conjecture of L. Breiman Thomas P. Hayes arxiv:1112.0829v1 [math.pr] 5 Dec 2011 Abstract Consider a gambling game in which we are allowed to repeatedly

More information

Analysis of a Production/Inventory System with Multiple Retailers

Analysis of a Production/Inventory System with Multiple Retailers Analysis of a Production/Inventory System with Multiple Retailers Ann M. Noblesse 1, Robert N. Boute 1,2, Marc R. Lambrecht 1, Benny Van Houdt 3 1 Research Center for Operations Management, University

More information

The Student-Project Allocation Problem

The Student-Project Allocation Problem The Student-Project Allocation Problem David J. Abraham, Robert W. Irving, and David F. Manlove Department of Computing Science, University of Glasgow, Glasgow G12 8QQ, UK Email: {dabraham,rwi,davidm}@dcs.gla.ac.uk.

More information

P2P File Sharing - A Model For Fairness Versus Performance

P2P File Sharing - A Model For Fairness Versus Performance 1 The Design Trade-offs of BitTorrent-like File Sharing Protocols Bin Fan John C.S. Lui Dah-Ming Chiu Abstract The BitTorrent (BT) file sharing protocol is very popular due to its scalability property

More information

Quality of Service versus Fairness. Inelastic Applications. QoS Analogy: Surface Mail. How to Provide QoS?

Quality of Service versus Fairness. Inelastic Applications. QoS Analogy: Surface Mail. How to Provide QoS? 18-345: Introduction to Telecommunication Networks Lectures 20: Quality of Service Peter Steenkiste Spring 2015 www.cs.cmu.edu/~prs/nets-ece Overview What is QoS? Queuing discipline and scheduling Traffic

More information

Parallel Scalable Algorithms- Performance Parameters

Parallel Scalable Algorithms- Performance Parameters www.bsc.es Parallel Scalable Algorithms- Performance Parameters Vassil Alexandrov, ICREA - Barcelona Supercomputing Center, Spain Overview Sources of Overhead in Parallel Programs Performance Metrics for

More information

How Much Can Taxes Help Selfish Routing?

How Much Can Taxes Help Selfish Routing? How Much Can Taxes Help Selfish Routing? Richard Cole Yevgeniy Dodis Tim Roughgarden July 28, 25 Abstract We study economic incentives for influencing selfish behavior in networks. We consider a model

More information

Energy Constrained Resource Scheduling for Cloud Environment

Energy Constrained Resource Scheduling for Cloud Environment Energy Constrained Resource Scheduling for Cloud Environment 1 R.Selvi, 2 S.Russia, 3 V.K.Anitha 1 2 nd Year M.E.(Software Engineering), 2 Assistant Professor Department of IT KSR Institute for Engineering

More information

Capacity Management and Equilibrium for Proportional QoS

Capacity Management and Equilibrium for Proportional QoS IEEE/ACM TRANSACTIONS ON NETWORKING 1 Capacity Management and Equilibrium for Proportional QoS Ishai Menache and Nahum Shimkin, Senior Member, IEEE Abstract Differentiated services architectures are scalable

More information

Deployment of express checkout lines at supermarkets

Deployment of express checkout lines at supermarkets Deployment of express checkout lines at supermarkets Maarten Schimmel Research paper Business Analytics April, 213 Supervisor: René Bekker Faculty of Sciences VU University Amsterdam De Boelelaan 181 181

More information

On the Traffic Capacity of Cellular Data Networks. 1 Introduction. T. Bonald 1,2, A. Proutière 1,2

On the Traffic Capacity of Cellular Data Networks. 1 Introduction. T. Bonald 1,2, A. Proutière 1,2 On the Traffic Capacity of Cellular Data Networks T. Bonald 1,2, A. Proutière 1,2 1 France Telecom Division R&D, 38-40 rue du Général Leclerc, 92794 Issy-les-Moulineaux, France {thomas.bonald, alexandre.proutiere}@francetelecom.com

More information

Using Generalized Forecasts for Online Currency Conversion

Using Generalized Forecasts for Online Currency Conversion Using Generalized Forecasts for Online Currency Conversion Kazuo Iwama and Kouki Yonezawa School of Informatics Kyoto University Kyoto 606-8501, Japan {iwama,yonezawa}@kuis.kyoto-u.ac.jp Abstract. El-Yaniv

More information

5 Market Games For Teaching Economics

5 Market Games For Teaching Economics 5 Market Games For Teaching Economics Progression 5 Market Games from website economics-games.com To be played separately or as a sequence: Market Game 1: Sunk costs, monopoly, and introduction to the

More information

Adaptive Online Gradient Descent

Adaptive Online Gradient Descent Adaptive Online Gradient Descent Peter L Bartlett Division of Computer Science Department of Statistics UC Berkeley Berkeley, CA 94709 bartlett@csberkeleyedu Elad Hazan IBM Almaden Research Center 650

More information

4. Fixed-Priority Scheduling

4. Fixed-Priority Scheduling Simple workload model 4. Fixed-Priority Scheduling Credits to A. Burns and A. Wellings The application is assumed to consist of a fixed set of tasks All tasks are periodic with known periods This defines

More information

Game Theory Based Iaas Services Composition in Cloud Computing

Game Theory Based Iaas Services Composition in Cloud Computing Game Theory Based Iaas Services Composition in Cloud Computing Environment 1 Yang Yang, *2 Zhenqiang Mi, 3 Jiajia Sun 1, First Author School of Computer and Communication Engineering, University of Science

More information

Oscillations of the Sending Window in Compound TCP

Oscillations of the Sending Window in Compound TCP Oscillations of the Sending Window in Compound TCP Alberto Blanc 1, Denis Collange 1, and Konstantin Avrachenkov 2 1 Orange Labs, 905 rue Albert Einstein, 06921 Sophia Antipolis, France 2 I.N.R.I.A. 2004

More information

How To Compare Load Sharing And Job Scheduling In A Network Of Workstations

How To Compare Load Sharing And Job Scheduling In A Network Of Workstations A COMPARISON OF LOAD SHARING AND JOB SCHEDULING IN A NETWORK OF WORKSTATIONS HELEN D. KARATZA Department of Informatics Aristotle University of Thessaloniki 546 Thessaloniki, GREECE Email: karatza@csd.auth.gr

More information

Performance of networks containing both MaxNet and SumNet links

Performance of networks containing both MaxNet and SumNet links Performance of networks containing both MaxNet and SumNet links Lachlan L. H. Andrew and Bartek P. Wydrowski Abstract Both MaxNet and SumNet are distributed congestion control architectures suitable for

More information

NOVEL PRIORITISED EGPRS MEDIUM ACCESS REGIME FOR REDUCED FILE TRANSFER DELAY DURING CONGESTED PERIODS

NOVEL PRIORITISED EGPRS MEDIUM ACCESS REGIME FOR REDUCED FILE TRANSFER DELAY DURING CONGESTED PERIODS NOVEL PRIORITISED EGPRS MEDIUM ACCESS REGIME FOR REDUCED FILE TRANSFER DELAY DURING CONGESTED PERIODS D. Todinca, P. Perry and J. Murphy Dublin City University, Ireland ABSTRACT The goal of this paper

More information

On the Existence of Optimal Taxes for Network Congestion Games with Heterogeneous Users

On the Existence of Optimal Taxes for Network Congestion Games with Heterogeneous Users On the Existence of Optimal Taxes for Network Congestion Games with Heterogeneous Users Dimitris Fotakis 1, George Karakostas 2, and Stavros G. Kolliopoulos 3 1 School of Electrical and Computer Engineering,

More information

A Load Balancing Algorithm based on the Variation Trend of Entropy in Homogeneous Cluster

A Load Balancing Algorithm based on the Variation Trend of Entropy in Homogeneous Cluster , pp.11-20 http://dx.doi.org/10.14257/ ijgdc.2014.7.2.02 A Load Balancing Algorithm based on the Variation Trend of Entropy in Homogeneous Cluster Kehe Wu 1, Long Chen 2, Shichao Ye 2 and Yi Li 2 1 Beijing

More information

Index Terms : Load rebalance, distributed file systems, clouds, movement cost, load imbalance, chunk.

Index Terms : Load rebalance, distributed file systems, clouds, movement cost, load imbalance, chunk. Load Rebalancing for Distributed File Systems in Clouds. Smita Salunkhe, S. S. Sannakki Department of Computer Science and Engineering KLS Gogte Institute of Technology, Belgaum, Karnataka, India Affiliated

More information

Distributed Load Balancing for Machines Fully Heterogeneous

Distributed Load Balancing for Machines Fully Heterogeneous Internship Report 2 nd of June - 22 th of August 2014 Distributed Load Balancing for Machines Fully Heterogeneous Nathanaël Cheriere nathanael.cheriere@ens-rennes.fr ENS Rennes Academic Year 2013-2014

More information

DECENTRALIZED LOAD BALANCING IN HETEROGENEOUS SYSTEMS USING DIFFUSION APPROACH

DECENTRALIZED LOAD BALANCING IN HETEROGENEOUS SYSTEMS USING DIFFUSION APPROACH DECENTRALIZED LOAD BALANCING IN HETEROGENEOUS SYSTEMS USING DIFFUSION APPROACH P.Neelakantan Department of Computer Science & Engineering, SVCET, Chittoor pneelakantan@rediffmail.com ABSTRACT The grid

More information

A MPCP-Based Centralized Rate Control Method for Mobile Stations in FiWi Access Networks

A MPCP-Based Centralized Rate Control Method for Mobile Stations in FiWi Access Networks A MPCP-Based Centralized Rate Control Method or Mobile Stations in FiWi Access Networks 215 IEEE. Personal use o this material is permitted. Permission rom IEEE must be obtained or all other uses, in any

More information

VoIP Network Dimensioning using Delay and Loss Bounds for Voice and Data Applications

VoIP Network Dimensioning using Delay and Loss Bounds for Voice and Data Applications VoIP Network Dimensioning using Delay and Loss Bounds for Voice and Data Applications Veselin Rakocevic School of Engineering and Mathematical Sciences City University, London, UK V.Rakocevic@city.ac.uk

More information

0.0.2 Pareto Efficiency (Sec. 4, Ch. 1 of text)

0.0.2 Pareto Efficiency (Sec. 4, Ch. 1 of text) September 2 Exercises: Problem 2 (p. 21) Efficiency: p. 28-29: 1, 4, 5, 6 0.0.2 Pareto Efficiency (Sec. 4, Ch. 1 of text) We discuss here a notion of efficiency that is rooted in the individual preferences

More information

A Comparison of General Approaches to Multiprocessor Scheduling

A Comparison of General Approaches to Multiprocessor Scheduling A Comparison of General Approaches to Multiprocessor Scheduling Jing-Chiou Liou AT&T Laboratories Middletown, NJ 0778, USA jing@jolt.mt.att.com Michael A. Palis Department of Computer Science Rutgers University

More information

Joint Optimization of Overlapping Phases in MapReduce

Joint Optimization of Overlapping Phases in MapReduce Joint Optimization of Overlapping Phases in MapReduce Minghong Lin, Li Zhang, Adam Wierman, Jian Tan Abstract MapReduce is a scalable parallel computing framework for big data processing. It exhibits multiple

More information

Dynamic Resource Pricing on Federated Clouds

Dynamic Resource Pricing on Federated Clouds Dynamic Resource Pricing on Federated Clouds Marian Mihailescu and Yong Meng Teo Department of Computer Science National University of Singapore Computing 1, 13 Computing Drive, Singapore 117417 Email:

More information

D1.1 Service Discovery system: Load balancing mechanisms

D1.1 Service Discovery system: Load balancing mechanisms D1.1 Service Discovery system: Load balancing mechanisms VERSION 1.0 DATE 2011 EDITORIAL MANAGER Eddy Caron AUTHORS STAFF Eddy Caron, Cédric Tedeschi Copyright ANR SPADES. 08-ANR-SEGI-025. Contents Introduction

More information

Truthful and Non-Monetary Mechanism for Direct Data Exchange

Truthful and Non-Monetary Mechanism for Direct Data Exchange Truthful and Non-Monetary Mechanism for Direct Data Exchange I-Hong Hou, Yu-Pin Hsu and Alex Sprintson Department of Electrical and Computer Engineering Texas A&M University {ihou, yupinhsu, spalex}@tamu.edu

More information

AN EFFICIENT DISTRIBUTED CONTROL LAW FOR LOAD BALANCING IN CONTENT DELIVERY NETWORKS

AN EFFICIENT DISTRIBUTED CONTROL LAW FOR LOAD BALANCING IN CONTENT DELIVERY NETWORKS Available Online at www.ijcsmc.com International Journal of Computer Science and Mobile Computing A Monthly Journal of Computer Science and Information Technology IJCSMC, Vol. 3, Issue. 9, September 2014,

More information

Windows Server Performance Monitoring

Windows Server Performance Monitoring Spot server problems before they are noticed The system s really slow today! How often have you heard that? Finding the solution isn t so easy. The obvious questions to ask are why is it running slowly

More information

A Topology-Aware Relay Lookup Scheme for P2P VoIP System

A Topology-Aware Relay Lookup Scheme for P2P VoIP System Int. J. Communications, Network and System Sciences, 2010, 3, 119-125 doi:10.4236/ijcns.2010.32018 Published Online February 2010 (http://www.scirp.org/journal/ijcns/). A Topology-Aware Relay Lookup Scheme

More information

The Price War in Wireless Network Marketing

The Price War in Wireless Network Marketing Price War in Heterogeneous Wireless Networks Patrick Maillé a, Bruno Tuffin b a Institut Telecom; Telecom Bretagne 2 rue de la Châtaigneraie CS 17607, 35576 Cesson-Sévigné Cedex, FRANCE b INRIA Rennes

More information