Ce24539543

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 default
in many overlay networks especially in peer-to-peer protocol to maximize the benefit. Selfish behaviour
file sharing systems. Previously these network creates additional incentives for nodes to rewire, not only
formations were approached in two ways. First by for operational purposes, but also for seizing opportunities
implementing practical heuristics for cooperative to incrementally maximize the local connection quality to
peers, and later performing game theoretic analysis the overlay. The impact of adopting selfish connectivity
for selfish peers. Selfish Neighbour Selection (SNS) management techniques in real overlay networks has been
game 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 node
limits number of neighbours for each peer. This must select a fixed number (k) of immediate overlay
approach reduces link monitoring overhead in overlay neighbours. Previous work has considered this problem
networks from O (n2) to O (n). Best response wirings from two perspectives: (1) Devising practical heuristics
increase the utility of Overlays and Simple heuristic for specific applications in real deployments, such as
wirings benefits non-selfish nodes. EGOIST, an SNS- bootstrapping by choosing the k closest links or by
inspired 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 fundamental
strategy as a k-median problem on asymmetric neighbour selection problem that are analytically
distance was developed. It’s evaluations on a variety of tractable, especially via game theoretic analysis. The bulk
performance metrics like delay, available bandwidth, of the work and main results in this area have centered on
and node utilization in Planet Lab confirms its strategic games where edges are undirected, access costs
efficiency 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 is
to attain in network dynamics. We propose to use extremely helpful for laying a theoretical foundation and
optimized spanning tree protocol (OSTP) that enables for building intuition, it is not clear how or whether the
traffic traversing the full-mesh portion to take a guidance provided by this prior work generalizes to
shortest path from source to destination. Using this situations of practical interest, in which underlying
full-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 of
Index Terms: Overlay Networks, Selfish Neighbour settings in which some or even most players do not play
Selection, Optimized spanning tree protocol. optimally – a setting which we believe to be typical.

In this paper, the central aspects of our model are
I. INTRODUCTION bounded degree, directed edges, non-uniform preference
An overlay network is a virtual network of nodes vectors, and representative distance functions. Overlays
and logical links that is built on top of an existing network are dominated by selfish nodes, the resulting stable
with 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 performance
are used for a variety of popular applications including through heuristic wiring strategies.
routing [1], content distribution, peer-to-peer (P2P) file
sharing 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 overlay
foundational issue in overlay network applications. routing network built around best response wiring
Establishing Connection to newcomer into existing mesh strategies. EGOIST serves as a building block for the
or re-establishing links between nodes is known as construction of efficient and scalable overlay applications
connectivity management. Overlays consist of nodes that consisting of (potentially) selfish nodes. We demonstrate
are distributed across multiple administrative domains, as (1) that overlay routing EGOIST is significantly more
such, by utilizing knowledge about the network; these efficient than systems utilizing common heuristic

539 | P a g e

neighbour selection strategies under multiple performance III. THE EGOIST OVERLAY ROUTING
metrics, including delay, system load and available SYSTEM
bandwidth. (2) We demonstrate that the performance of In previous work, it is not clear; what is the
EGOIST approaches that of a full-mesh topology, while average performance gain when SNS wiring strategies are
achieving superior scalability, requiring link used in highly dynamic environments, whether such
announcements proportional to nk compared to n2 for a overlays are robust against churn, and whether they scale
full mesh topology. We propose to use optimized and whether it is practical to build overlays to support
spanning tree protocol (OSTP) that enables traffic best response and how to incorporate additional metrics
traversing the full-mesh portion to take a shortest path other than delay, e.g., bandwidth. In order address the
from source to destination. Our experimental results show problems, we implemented EGOIST, and SNS-inspired
that EGOIST remains highly effective under significant prototype overlay routing network, serves as a building
churn and incurs minimal overhead. block for the distributed construction of efficient and
resilient overlays where both individual and social
II. 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 tree
and 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 to
overlay 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 which
overlay routing or content sharing impose hard constraints allows the creation and maintenance of an overlay
on the maximum number of overlay neighbours. In network. Here every node selects and continuously
overlay 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 all
reporting overhead imposed by the link-state routing destinations under shortest-path routing. For ease of
protocol implemented at the overlay layer. Node degrees presentation, Instead of bandwidth ad node utilization, we
were 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 the
additional cost of setting up more links and the decreased In EGOIST, a newcomer overlay node vi
communication 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 of
Directed Edges: Another important consideration in our these nodes, enable link state routing protocol running at
work relates to link directionality. Prior models have the overlay layer. As a result, after some time, vi obtains
generally 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 able
the unbounded node degree assumption for models that to obtain the pair-wise distance (delay) function dG-i. In
target 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 to
bidirectional. In overlay settings we consider, this node vi, for all vi∈ V-i. Using the values of dij and dG-i , the
assumption needs to be relaxed since the fact that node v newcomer connects to G-i using one of the wiring
forwards traffic or requests to node u does not mean that strategies. In our implementation, each node acts as a
node u may also forward traffic or requests to v. server that listens to all the messages of the link state
Undirected links are created by the establishment of two protocol and propagates them only to its immediate
directed links. neighbours. In order to reduce the traffic in the system,
messages are dropped that have been received more than
Non-uniform preference vectors: In our model, a vector once and propagates unique messages. There are also two
is supplied to each node that captures its local preference threads, one for estimating dij, and one responsible for
for all other destinations. In overlay routing such estimating the new wiring and propagating the wiring to
preference may capture the percentage of locally the immediate neighbours. In order to minimize the load
generated traffic that a node routes to each destination, in the system, a node propagates its wiring to its
and then the aggregation of all preference vectors would immediate neighbours only if this changes.
amount to an origin/destination traffic matrix. In P2P
overlays such preference may amount to speculations B. optimized spanning tree protocol
from 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

mesh portion of a network, and is compatible with connects to the nodes taken by adding (modulo n) its id to
external networks where a standard spanning tree protocol each offset. If k2 is small then the nodes will need to
is used. The OSTP enables traffic traversing the full-mesh monitor (e.g., ping) the backbone links closely so as to
portion to take a shortest path from source to destination quickly identify and restore disconnections. With higher
through use of full-mesh connectivity. In some k2 the monitoring can be more relaxed due to the
embodiments, 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 the
internal ports of the servers. In some embodiments, the k2 links can be achieved by restricting the set candidate
OSTP is configured on a per-VLAN basis. In some immediate neighbours for swapping.
embodiments, the servers exchange special messages
enabling 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 monitored
suppressed on certain port types, such as external ports. In aggressively so as to recover promptly from any
some embodiments, determination of the full-mesh disconnections in the connectivity backbone through the
connectivity 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 links
portion. could be done lazily, namely by measuring link costs, and
recomputing BR wirings at a pace that is convenient to
C. 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 leave
the overlay. Therefore, even in an extreme case in which To differentiate between these two types of link
some nodes are reachable through only a unique path, a monitoring strategies (aggressive versus lazy), in
node can count on this path always being in place. EGOIST we allow rewiring of a dropped link to be
Overlay routing networks (e.g., RON [3]) are not performed in one of two different modes: immediate and
inherently prone to churn to the extent that file-sharing delayed. In immediate mode, re-wiring is done as soon as
P2P networks. Nonetheless, nodes may occasionally go it is determined that the link is dropped, whereas in
down, 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 churning
behaviour of other nodes. This, however, requires D. Cost Metrics
modelling 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 we
an 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 metrics
reminiscent of how k-Random and k-Closest strategies
for many applications, including real-time ones. To obtain
ensure overlay connectivity. We introduce a hybrid wiring
strategy (Hybrid BR), in which each node uses k1 of its k the delay cost metric, a node needs to obtain estimates for
links 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 delays
be used for assuring basic connectivity under churn. We
using two different methods: an active method based on
call this wiring ―hybrid‖ because, in effect, two wiring
strategies are in play – a selfish BR strategy that aims to ping, and a passive method using the pyxida virtual
maximize 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. Using
redundant 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 more
use 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 and
complexities 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 is
avoid 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 in
for the creation of a single bidirectional cycle. For higher
which all outgoing links from a node are assigned the
k2, the system decides k2/2 offsets and then each node
541 | P a g e

same cost, which is set to be equal to the measured load memory consumption is close to 0%, and the bandwidth
of 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 a
given interval.

Available Bandwidth: Another important cost metric,
especially for content delivery applications, is the
available bandwidth on overlay links. In EGOIST, we
used pathChirp[10], a light-weight, fast and accurate tool,
which fits well with specific constraints, namely: it does
not impose a high load on nodes since it does not require
the transmission of long sequences of packet trains, and it
does not exceed the maxburst limits of Planet lab.

IV. PERFORMANCE EVALUATION OF
EGOIST
To facilitate comparisons between various
neighbour selection strategies, we often report the
normalized routing cost , which is the ratio of the cost
achievable using a given strategy to that achievable using
BR.
Control Variables: In our first set of experiments, our
aim is to identify for the three metrics of interest the
payoff from adopting a selfish neighbour selection
strategy, i.e., using a BR strategy in EGOIST. This payoff
will depend on many variables. While some of these
variables are not within our control (e.g., the dynamic
nature of the Internet as reflected by variability in
observed Planet Lab conditions), others are within our
control, e.g., n, T, and the various settings for our active
measurement techniques.

Fig. 2. Normalized individual costs and 95th-percentile
Fig 1: Percentage of CPU and memory consumption, and confidence intervals with respect to BR cost, under
bandwidth consumption in EGOIST. The metric is delay different metrics, of various neighbour selection strategies
via ping. in a 50-node EGOIST overlay.

One control variable that is particularly Results for Delay Metric: Figures 1(a) and (b) show the
important is the number of direct neighbours, k that an performance of the various neighbour selection strategies
EGOIST node is allowed to have. In any ways, k puts a in EGOIST normalized with respect to that achievable
premium on the significance of making a judicious choice using BR when the metric of interest is the overlay path
of neighbours. For small values of k, choosing the right delay over a range of values for k (using ping and
set of neighbours has the potential of making a bigger pyxida). These results show that BR outperforms all the
impact on performance, when compared to the impact for other wiring strategies, especially when k is small. The
larger values of k. In Fig 1, we show the average CPU and performance advantage of BR in terms of routing delay
memory utilization, along with the average bandwidth stands, even for a moderate number of neighbours. These
consumption 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

performance of EGOIST using BR wiring to the ―best boost when compared to currently used heuristics, and
possible‖ performance. To do so, we note that by that they scale much better simple heuristics. Optimized
allowing nodes to connect to all other nodes in the spanning tree protocol (OSTP) enables traffic traversing
overlay, we would be creating a complete overlay graph the full-mesh portion to take a shortest path from source
with O(n2) overlay links, obviating the need for a to destination. Finally, for achieving real-time
neighbour selection strategy. With respect to the other requirements and carrying the traffic generated by an
heuristics, 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 a
k-Random ends up outperforming k-Closest for slightly building block for efficient routing in overlay
larger values of k. This can be explained by noting that k- applications.
Random ends up creating graphs with much smaller
diameters than the grid-like graphs resulting from the use
of k-Closest, especially as k gets larger. In all VI REFERENCES
experiments, k- Regular performed the worst. In [9] we [1] D. Andersen, H. Balakrishnan, F. Kaashoek, and
also 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 Selfishly
is the sum of the loads of all nodes in the path. These Constructed Overlay Routing Networks,‖ in
results 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 Overlay
delivering 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 Selfish
Results 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, and
metric. Recall that, here, the objective is to get the highest J. W. Byers, ―Implications of Selfish Neighbor
possible aggregate bandwidth to all destinations (again, Selection in Overlay Networks,‖ in Proc. IEEE
assuming 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 of
similar to those obtained for the delay metric, with BR Peer-to-peer Networks: Service Capacity and
outperforming 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, 2006
wide range of values of k. [7] A. Young, J. Chen, Z. Ma, A. Krishnamurthy, L.
L. Peterson, and R. Wang, ―Overlay Mesh
V. 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 Search
neighbour budgets and has come up with a series of Heuristics for k-Median and Facility Location
findings 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 Available
beneficial for all. Secondly, through the development, Bandwidth Estimation for Network Paths,‖ in
implementation, and deployment of EGOIST, we have Proc. PAM’03.
established that Best-Response neighbour selection
strategies that they provide a substantial performance

543 | P a g e

Ce24539543

More Related Content

What's hot

Viewers also liked

Similar to Ce24539543

Ce24539543