Published on

  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide


  1. 1. V.Srilakshmi, M. Raja Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 4, July-August 2012, pp.539-543 Optimization of Overlay Networks using Self-Centric Peer Selection V.Srilakshmi 1, M. Raja Babu 2 1 Mtech Student, Dept of CSE, Aditya Engineering College, Surampalem, Andhra Pradesh, India, 2 Assoc. Professor, Dept of CSE, Aditya Engineering College, Surampalem, Andhra Pradesh, India,Abstract Changing network dynamics is a critical issue nodes may act selfishly and deviate from the defaultin many overlay networks especially in peer-to-peer protocol to maximize the benefit. Selfish behaviourfile sharing systems. Previously these network creates additional incentives for nodes to rewire, not onlyformations were approached in two ways. First by for operational purposes, but also for seizing opportunitiesimplementing practical heuristics for cooperative to incrementally maximize the local connection quality topeers, and later performing game theoretic analysis the overlay. The impact of adopting selfish connectivityfor selfish peers. Selfish Neighbour Selection (SNS) management techniques in real overlay networks has beengame theory is designed and implemented in overlay an open problem.routing that unifies the fore mentioned ways. SNS In overlay network for routing traffic, a nodelimits number of neighbours for each peer. This must select a fixed number (k) of immediate overlayapproach reduces link monitoring overhead in overlay neighbours. Previous work has considered this problemnetworks from O (n2) to O (n). Best response wirings from two perspectives: (1) Devising practical heuristicsincrease the utility of Overlays and Simple heuristic for specific applications in real deployments, such aswirings benefits non-selfish nodes. EGOIST, an SNS- bootstrapping by choosing the k closest links or byinspired prototype overlay routing system that choosing k random links in a P2P file-sharing system. (2)expresses and solves a node’s “best response” wiring Providing abstractions of the underlying fundamentalstrategy as a k-median problem on asymmetric neighbour selection problem that are analyticallydistance was developed. It’s evaluations on a variety of tractable, especially via game theoretic analysis. The bulkperformance metrics like delay, available bandwidth, of the work and main results in this area have centered onand node utilization in Planet Lab confirms its strategic games where edges are undirected, access costsefficiency over existing heuristic overlays. EGOIST are based on hop-counts, and nodes have potentially(assumes) is based on full-mesh topology which is hard unbounded degrees. While this existing body of work isto attain in network dynamics. We propose to use extremely helpful for laying a theoretical foundation andoptimized spanning tree protocol (OSTP) that enables for building intuition, it is not clear how or whether thetraffic traversing the full-mesh portion to take a guidance provided by this prior work generalizes toshortest path from source to destination. Using this situations of practical interest, in which underlyingfull-mesh the claims of EGOIST can be validated.. assumptions in these prior studies are not satisfied. Aspect not considered in previous work is the consideration ofIndex Terms: Overlay Networks, Selfish Neighbour settings in which some or even most players do not playSelection, Optimized spanning tree protocol. optimally – a setting which we believe to be typical. In this paper, the central aspects of our model areI. INTRODUCTION bounded degree, directed edges, non-uniform preference An overlay network is a virtual network of nodes vectors, and representative distance functions. Overlaysand logical links that is built on top of an existing network are dominated by selfish nodes, the resulting stablewith the purpose to implement a network service that is wirings are already so highly optimized that even non-not available in the existing network. Overlay networks selfish newcomers can extract near optimal performanceare used for a variety of popular applications including through heuristic wiring strategies.routing [1], content distribution, peer-to-peer (P2P) filesharing and streaming, data-centre applications, and Based on above observations, we design,online multi-player games. Connectivity management is a implement, and deploy EGOIST, a prototype overlayfoundational issue in overlay network applications. routing network built around best response wiringEstablishing Connection to newcomer into existing mesh strategies. EGOIST serves as a building block for theor re-establishing links between nodes is known as construction of efficient and scalable overlay applicationsconnectivity management. Overlays consist of nodes that consisting of (potentially) selfish nodes. We demonstrateare distributed across multiple administrative domains, as (1) that overlay routing EGOIST is significantly moresuch, by utilizing knowledge about the network; these efficient than systems utilizing common heuristic 539 | P a g e
  2. 2. V.Srilakshmi, M. Raja Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 4, July-August 2012, pp.539-543neighbour selection strategies under multiple performance III. THE EGOIST OVERLAY ROUTINGmetrics, including delay, system load and available SYSTEMbandwidth. (2) We demonstrate that the performance of In previous work, it is not clear; what is theEGOIST approaches that of a full-mesh topology, while average performance gain when SNS wiring strategies areachieving superior scalability, requiring link used in highly dynamic environments, whether suchannouncements proportional to nk compared to n2 for a overlays are robust against churn, and whether they scalefull mesh topology. We propose to use optimized and whether it is practical to build overlays to supportspanning tree protocol (OSTP) that enables traffic best response and how to incorporate additional metricstraversing the full-mesh portion to take a shortest path other than delay, e.g., bandwidth. In order address thefrom source to destination. Our experimental results show problems, we implemented EGOIST, and SNS-inspiredthat EGOIST remains highly effective under significant prototype overlay routing network, serves as a buildingchurn and incurs minimal overhead. block for the distributed construction of efficient and resilient overlays where both individual and socialII. OVERLAY NETWORK MODEL performance is close to optimal. EGOIST is based on full- Previous work on overlay network creation [2], mesh topology which is hard to attain in network[3] has focused on physical telecommunication networks dynamics. We propose to use optimized spanning treeand primarily the Internet. Overlay networks are protocol (OSTP) that enables traffic traversing the full-substantially different [6] which consider the following mesh portion to take a shortest path from source tooverlay network model. The central aspects of our model destination.are:Bounded Degree: Most protocols used for implementing A. Basic Design EGOIST is a distributed system whichoverlay routing or content sharing impose hard constraints allows the creation and maintenance of an overlayon the maximum number of overlay neighbours. In network. Here every node selects and continuouslyoverlay routing systems, the number of immediate nodes updates its k overlay neighbours in a selfish manner—has to be kept small so as to reduce the monitoring and namely to minimize its (weighted) sum of distances to allreporting overhead imposed by the link-state routing destinations under shortest-path routing. For ease ofprotocol implemented at the overlay layer. Node degrees presentation, Instead of bandwidth ad node utilization, wewere implicitly bounded (as opposed to explicitly will assume that delay is used to reflect the cost of a path.constrained) by virtue of the trade-off between theadditional cost of setting up more links and the decreased In EGOIST, a newcomer overlay node vicommunication distance achieved through the addition of connects to the system by querying a bootstrap node,new links. from which it receives a list of potential overlay neighbours. The newcomer connects to at least one ofDirected Edges: Another important consideration in our these nodes, enable link state routing protocol running atwork relates to link directionality. Prior models have the overlay layer. As a result, after some time, vi obtainsgenerally assumed bi-directional [5] (undirected) links. the full residual graph G-i of the overlay. By running all-This is an acceptable assumption that fits naturally with pairs shortest path algorithm on G-i , the newcomer is ablethe unbounded node degree assumption for models that to obtain the pair-wise distance (delay) function dG-i. Intarget physical telecommunication network because actual addition to this information, the newcomer estimates d ij,wire-line communication links are almost exclusively the weight of a potential direct overlay link from itself tobidirectional. In overlay settings we consider, this node vi, for all vi∈ V-i. Using the values of dij and dG-i , theassumption needs to be relaxed since the fact that node v newcomer connects to G-i using one of the wiringforwards traffic or requests to node u does not mean that strategies. In our implementation, each node acts as anode u may also forward traffic or requests to v. server that listens to all the messages of the link stateUndirected links are created by the establishment of two protocol and propagates them only to its immediatedirected links. neighbours. In order to reduce the traffic in the system, messages are dropped that have been received more thanNon-uniform preference vectors: In our model, a vector once and propagates unique messages. There are also twois supplied to each node that captures its local preference threads, one for estimating dij, and one responsible forfor all other destinations. In overlay routing such estimating the new wiring and propagating the wiring topreference may capture the percentage of locally the immediate neighbours. In order to minimize the loadgenerated traffic that a node routes to each destination, in the system, a node propagates its wiring to itsand then the aggregation of all preference vectors would immediate neighbours only if this changes.amount to an origin/destination traffic matrix. In P2Poverlays such preference may amount to speculations B. optimized spanning tree protocolfrom the local node about the quality of, or interest in, the An optimized spanning tree protocol (OSTP)content held by other nodes. minimizes latency and provides high throughput in a full- 540 | P a g e
  3. 3. V.Srilakshmi, M. Raja Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 4, July-August 2012, pp.539-543mesh portion of a network, and is compatible with connects to the nodes taken by adding (modulo n) its id toexternal networks where a standard spanning tree protocol each offset. If k2 is small then the nodes will need tois used. The OSTP enables traffic traversing the full-mesh monitor (e.g., ping) the backbone links closely so as toportion to take a shortest path from source to destination quickly identify and restore disconnections. With higherthrough use of full-mesh connectivity. In some k2 the monitoring can be more relaxed due to theembodiments, a cluster includes a plurality of servers existence of alternative routes through other cycles.connected in a full mesh, and the OSTP is used on Computing BR using k1 links granted the existence of theinternal ports of the servers. In some embodiments, the k2 links can be achieved by restricting the set candidateOSTP is configured on a per-VLAN basis. In some immediate neighbours for swapping.embodiments, the servers exchange special messagesenabling determination of full-mesh connectivity. In We have implemented HybridBR in EGOIST.further embodiments, sending of the special messages is As hinted above, donated links are monitoredsuppressed on certain port types, such as external ports. In aggressively so as to recover promptly from anysome embodiments, determination of the full-mesh disconnections in the connectivity backbone through theconnectivity disables use of a standard spanning tree use of frequent heartbeat signalling. On the other hand,protocol and/or enables use of OSTP on the full-mesh the monitoring and upkeep of the remaining BR linksportion. could be done lazily, namely by measuring link costs, and recomputing BR wirings at a pace that is convenient toC. Dealing with Churn the node—a pace that reduces probing and computational EGOIST’s BR (best response) neighbour overheads without risking global connectivity.selection strategy assumes that existing nodes never leavethe overlay. Therefore, even in an extreme case in which To differentiate between these two types of linksome nodes are reachable through only a unique path, a monitoring strategies (aggressive versus lazy), innode can count on this path always being in place. EGOIST we allow rewiring of a dropped link to beOverlay routing networks (e.g., RON [3]) are not performed in one of two different modes: immediate andinherently prone to churn to the extent that file-sharing delayed. In immediate mode, re-wiring is done as soon asP2P networks. Nonetheless, nodes may occasionally go it is determined that the link is dropped, whereas indown, or network problems may cause transient delayed mode re-wiring is only performed (if necessary)disconnections until successive re-wirings establish new at the preset wiring epoch T. unless otherwise specified,paths. One could re-formulate the BR objective function we assume a delayed re-wiring mode is in use.used by a node to take into account the churningbehaviour of other nodes. This, however, requires D. Cost Metricsmodelling of the churn characteristics of various nodes in The choice of an appropriate ―cost‖ of traversing a link depends largely on the application. In EGOIST wean overlay, which is not feasible in large networks. consider the following metrics: In EGOIST, we follow a different approach Link and Path Delays: Delays are natural cost metricsreminiscent of how k-Random and k-Closest strategies for many applications, including real-time ones. To obtainensure overlay connectivity. We introduce a hybrid wiringstrategy (Hybrid BR), in which each node uses k1 of its k the delay cost metric, a node needs to obtain estimates forlinks to selfishly optimize its performance using BR, and its own delay to potential neighbours, and for the delay between pairs of overlay nodes already in the network. In―donates‖ the remaining k2 = k – k1 links to the system to EGOIST, we estimate directed (one-way) link delaysbe used for assuring basic connectivity under churn. We using two different methods: an active method based oncall this wiring ―hybrid‖ because, in effect, two wiringstrategies are in play – a selfish BR strategy that aims to ping, and a passive method using the pyxida virtualmaximize local performance and a selfless strategy that coordinate system. Using ping, one-way delay is estimated to be one half of the measured ping round-trip-aims to maintain global connectivity by providing times (RTT) averaged over enough samples. Usingredundant routes. pyxida, delay estimates are available through a simple There are several ways in which a system can query to the pyxida system. Using ping produces moreuse the k2 donated links of each node to build a accurate estimates, but subjects the overlay to added load,connectivity backbone. Using k-MST (a centralized whereas using pyxida produces less accurate estimates,construction) to maintain connectivity is problematic, as it but consumes much less bandwidth.must always be updated, not to mention the overhead andcomplexities involved in establishing (k2/2)-MSTs. To Node Load: For many overlay applications, it may be the case that the primary determinant of the cost of a path isavoid these complexities, EGOIST uses a simpler solution the performance of the nodes along that path. In EGOIST,that forms k2/2 bidirectional cycles. For k2 = 2, it allows we allow the use of a variation of the delay metric infor the creation of a single bidirectional cycle. For higher which all outgoing links from a node are assigned thek2, the system decides k2/2 offsets and then each node 541 | P a g e
  4. 4. V.Srilakshmi, M. Raja Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 4, July-August 2012, pp.539-543same cost, which is set to be equal to the measured load memory consumption is close to 0%, and the bandwidthof the node. In EGOIST, we did this by querying the CPU consumption per node is negligible.load of the local node, and computing an exponentially-weighted moving average of that load calculated over agiven interval.Available Bandwidth: Another important cost metric,especially for content delivery applications, is theavailable bandwidth on overlay links. In EGOIST, weused pathChirp[10], a light-weight, fast and accurate tool,which fits well with specific constraints, namely: it doesnot impose a high load on nodes since it does not requirethe transmission of long sequences of packet trains, and itdoes not exceed the maxburst limits of Planet lab.IV. PERFORMANCE EVALUATION OFEGOIST To facilitate comparisons between variousneighbour selection strategies, we often report thenormalized routing cost , which is the ratio of the costachievable using a given strategy to that achievable usingBR.Control Variables: In our first set of experiments, ouraim is to identify for the three metrics of interest thepayoff from adopting a selfish neighbour selectionstrategy, i.e., using a BR strategy in EGOIST. This payoffwill depend on many variables. While some of thesevariables are not within our control (e.g., the dynamicnature of the Internet as reflected by variability inobserved Planet Lab conditions), others are within ourcontrol, e.g., n, T, and the various settings for our activemeasurement techniques. Fig. 2. Normalized individual costs and 95th-percentileFig 1: Percentage of CPU and memory consumption, and confidence intervals with respect to BR cost, underbandwidth consumption in EGOIST. The metric is delay different metrics, of various neighbour selection strategiesvia ping. in a 50-node EGOIST overlay. One control variable that is particularly Results for Delay Metric: Figures 1(a) and (b) show theimportant is the number of direct neighbours, k that an performance of the various neighbour selection strategiesEGOIST node is allowed to have. In any ways, k puts a in EGOIST normalized with respect to that achievablepremium on the significance of making a judicious choice using BR when the metric of interest is the overlay pathof neighbours. For small values of k, choosing the right delay over a range of values for k (using ping andset of neighbours has the potential of making a bigger pyxida). These results show that BR outperforms all theimpact on performance, when compared to the impact for other wiring strategies, especially when k is small. Thelarger values of k. In Fig 1, we show the average CPU and performance advantage of BR in terms of routing delaymemory utilization, along with the average bandwidth stands, even for a moderate number of neighbours. Theseconsumption to maintain the overlay per node. CPU and results confirm the superiority of BR relative to other strategies, but do not give us a feel for how close is the 542 | P a g e
  5. 5. V.Srilakshmi, M. Raja Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 4, July-August 2012, pp.539-543performance of EGOIST using BR wiring to the ―best boost when compared to currently used heuristics, andpossible‖ performance. To do so, we note that by that they scale much better simple heuristics. Optimizedallowing nodes to connect to all other nodes in the spanning tree protocol (OSTP) enables traffic traversingoverlay, we would be creating a complete overlay graph the full-mesh portion to take a shortest path from sourcewith O(n2) overlay links, obviating the need for a to destination. Finally, for achieving real-timeneighbour selection strategy. With respect to the other requirements and carrying the traffic generated by anheuristics, the results in Figures 2 (a) and (b) show that k- online multi-player P2P game, we use EGOIST prototype.Closest outperforms k-Random when k is small, but that Egoist incurs minimal overhead and it can be used as ak-Random ends up outperforming k-Closest for slightly building block for efficient routing in overlaylarger values of k. This can be explained by noting that k- applications.Random ends up creating graphs with much smallerdiameters than the grid-like graphs resulting from the useof k-Closest, especially as k gets larger. In all VI REFERENCESexperiments, k- Regular performed the worst. In [9] we [1] D. Andersen, H. Balakrishnan, F. Kaashoek, andalso show that BR wiring strategy is robust to cheating. R. Morris, ―Resilient Overlay Networks,‖ in Proc. ACM SOSP’01.Results for Node Load: Figure 2 (c) shows the results we [2] B.-G. Chun, R. Fonseca, I. Stoica, and J.obtained using the node load metric, where the path cost Kubiatowicz, ―Characterizing Selfishlyis the sum of the loads of all nodes in the path. These Constructed Overlay Routing Networks,‖ inresults show clear delineations, with BR delivering the Proc. IEEE INFOCOM’04.best performance over all values of k, k-Random [3] B. G. Rocha, V. Almeida, and D. Guedes,delivering the second-best performance, and k-Closest ―Improving Reliability of Selfish Overlaydelivering the worst performance as it fails to predict Networks,‖ in Proc. WWW’06.anything beyond the immediate neighbour, especially in [4] G. Smaragdakis, V. Lekakis, N. Laoutaris, A.light of the high load variance in Planet Lab. Bestavros, J. W. Byers, and M. Roussopoulos, ―EGOIST: Overlay Routing using SelfishResults for Available Bandwidth: Figure 2(d) shows the Neighbor Selection,‖ in Proc. ACM CoNEXT’08.results we obtained using available bandwidth as the cost [5] N. Laoutaris, G. Smaragdakis, A. Bestavros, andmetric. Recall that, here, the objective is to get the highest J. W. Byers, ―Implications of Selfish Neighborpossible aggregate bandwidth to all destinations (again, Selection in Overlay Networks,‖ in Proc. IEEEassuming a uniform preference for all destinations) – thus, INFOCOM’07.larger is better. These results show trends that are quite [6] X. Yang and G. de Veciana, ―Performance ofsimilar to those obtained for the delay metric, with BR Peer-to-peer Networks: Service Capacity andoutperforming all other strategies—delivering a two-fold Role of Resource Sharing Policies,‖ Perform.to four-fold improvement over the other strategies, over a Eval., vol. 63, no. 3, pp. 175–194, 2006wide range of values of k. [7] A. Young, J. Chen, Z. Ma, A. Krishnamurthy, L. L. Peterson, and R. Wang, ―Overlay MeshV. CONCLUSION Construction Using Interleaved Spanning Trees,‖ Selfish neighbour selection in a richer model that in Proc. IEEE INFOCOM’04.captures the nuances of overlay applications more [8] V. Arya, N. Garg, R. Khandekar, A. Meyerson,faithfully than previous work and strictly enforced K. Munagala, and V. Pandit, ―Local Searchneighbour budgets and has come up with a series of Heuristics for k-Median and Facility Locationfindings with substantial practical value for real overlay Problems,‖ SIAM J. on Computing, vol. 33, no.networks. First, when compared with simple random and 3, pp. 544–562, 2004.myopic heuristics, we have shown that a best response [9] G. Smaragdakis, ―Overlay Network Creation and(i.e., selfish) selection of neighbours leads to the Maintenance with Selfish Users,‖ Ph.D.construction of overlays with much better performance. Dissertation, Boston University, 2009.The reason is that by being selfish, nodes embark on a [10] V. Ribeiro, R. Riedi, R. Baraniuk, J. Navratil,distributed optimization of the overlay that turns out to be and L. Cottrell, ―pathChirp: Efficient Availablebeneficial for all. Secondly, through the development, Bandwidth Estimation for Network Paths,‖ inimplementation, and deployment of EGOIST, we have Proc. PAM’03.established that Best-Response neighbour selectionstrategies that they provide a substantial performance 543 | P a g e