The Effect of Choosing Proper Overlay Topology on the Peer to Peer Networks Properties by ijcsiseditor

VIEWS: 105 PAGES: 4

									                                                         (IJCSIS) International Journal of Computer Science and Information Security,
                                                         Vol. 10, No. 4, April 2012




   The Effect of Choosing Proper Overlay Topology
       on the Peer to Peer Networks’ Properties
                          Mohammed Gharib                                                Amirreza Soudi
                 Department of Computer Engineering                            Department of Computer Engineering
                   Sharif University of Technology                               Sharif University of Technology
                            Tehran, Iran                                                  Tehran, Iran
                     Email: gharib@ce.sharif.edu                                   Email: soudi@ce.sharif.edu


                                                               Abstract
        P2P networks have attracted attention of many Internet users due to their ability to share large volume of data (mostly video
    and music) among people regardless of their locations. The underlay of such networks is usually based on Internet infrastructure.
    Thus a large amount of the Internet Bandwidth is allocated to transfer different data. As a result, the traffic generated by this
    type of networks is becoming one of the main problems in the cyber world. Since that most P2P networks choose their graph
    due to their algorithm, not graph’s properties, so we suggest to choose overlay graph based on graph properties itself; it cause
    enhancement in the network traffic, network time and many other properties of the P2P networks. To show this fact, we use
    Chord network, as the most renowned P2P overlay. It uses a ring graph as its overlay topology, we replace it by the more
    approperiate graph, Hypercube, then study the effects of this replacement on the network properties. We showed that this simple
    modification enhance the creation time and decrease the control traffic of the network.

    Keywords: P2P networks; Hypercube; Chord; Control traffic; Overlay topology.

                      I. I NTRODUCTION                                  one is implemented more convenient, not to the properties of
                                                                        graph itself.
   Nowadays the volume of Internet traffic mostly is generated
by different P2P networks. Also, P2P networks have attracted               The other layer is Physical layer, in which real nodes
a lot of attention because they are simple, cost effective and          (computers) connections are established. Also, actual routing
dynamic. Our goal in this paper is to improve the efficiency             is done in this layer; moreover, the delay for transferring
of such networks that use Internet as their infrastructure.             a packet from one node to another is determined in this
Currently various P2P networks exploit about 50-70% of total            layer. This layer consists of nodes, connections between them,
Internet traffic [1]. As a result, any improvement in the per-           routers, switches and etc.
formance of these networks leads to significant improvement                 P2P networks have changed over time depending on the
in the performance of Internet network. These networks are              needs and legal issues. These changes have been made over the
usually composed of two layers: Overlay and Underlay.                   years and create new generations of P2P networks. Actually
   The first layer is overlay layer, a layer which defines a              it can be said that P2P networks are composed of three
topology of the network; how the nodes are connected to                 generations [5]. The first one is P2P centralized network;
each other. This topology is not actual or physical, it is only         These P2P networks have a central server which is respon-
hypothetical arrangement to perform functions like search,              sible for adjusting of any related activity to the network. In
routing, broadcast and etc. In other words, it is a virtual             this generation of peer to peer networks, the central server
arrangement that represents placement of nodes joined to a              deals with all challenges including search, routing, network
P2P network. One of these topologies is the Ring topology in            connection style, etc. The second one is P2P Unstructured -
which each node, when joining the network, is located in a              Decentralized network; This generation of the Peer to Peer
place on a circle circumference [2] . Tree topology is another          network require no central server and nodes must themselves
topology in which each node has parents and maybe some                  meet the network challenges. This type of network is forced to
children [3]. There are other topologies like mesh topology.            use broadcast everywhere, because of the lack of any structure.
In this topology each node, is placed in the mesh graph [4].            More usage of broadcast lead to more traffic in the network.
Some topologies are constructed from combination of two or              In these types of networks, the more the nodes, the more the
more topologies such as Cube connected cycle. This kind of              connection number and this means an increase in the network
topologies are named Combinational topologies. We believe               traffic. A large rise in the network traffic will ultimately lead
that it is important to select the appropriate graph for the            to Network collapse. The third generation is P2P Structured -
overlay layer in the P2P networks. By changing the graph, we            Decentralized; This generation of P2P networks has no central
can greatly reduce the network traffic and delay exploited for           server, so to perform its actions like search, routing and etc.,
creating such networks. Currently most existed P2P networks             it doesn’t use broadcast message; instead, it employs a table
choose theirs overlay topology due to their algorithms ,which           called Distributed Hash Table (DHT) [6].


                                                                   99                               http://sites.google.com/site/ijcsis/
                                                                                                    ISSN 1947-5500
                                                         (IJCSIS) International Journal of Computer Science and Information Security,
                                                         Vol. 10, No. 4, April 2012


   The rest of this paper is organized as follows. In section 2      Hypercube topology in different sizes.
we describe our proposed algorithm and the parameters that
are calculated. In section 3 we explain the experimental results.                     III. E XPERIMENTAL R ESULTS
Finally, in section 4, conclusion are drawn.                            We use Planetsim as our P2P network simulator. The best
                                                                     advantage of Planetsim is its separation between the overlay
               II. D ISPLACE OVERLAY G RAPH                          and the services within peer to peer networks. In the Planetsim
   As mentioned earlier any P2P network has an overlay;              simulator the services of the overlays such as DHT and DOLR
each overlay is composed of a topology. Some properties like         is completely separated from overlay topology, implying that
degree, diameter, scalability, regularity and symmetricity of        we can change the overlay topology with out any change in
the graph are very important in selecting the proper topology        services on it. so we exchange Ring graph in the Chord with
[7]. The graph with higher degree, has higher connectivity           Hypercube without any modification in the Chord algorithm.
and probably of shorter paths. Some operations such broadcast           In our simulations we map Hypercube over the Ring graph
over the higher degree graph, cause higher traffic, maybe cause       in chord network. Some advantages of Hypercube over Ring
in a network collapse. So the graph must be chosen such              topology is that the Ring degree is 2 and it leads to less
that the tradeoff between degree and number of nodes, be             neighbours and limited connectivity between nodes; since that
considered. The graph is better for the topology if it has shorter   the degree of Hypercube topology is the same as the number
diameter. Also the more scalable topology is better for the P2P      of its dimensions. Note that the very high degree topology
networks overlay.                                                    leads to more traffic too. So, the topology should be choosen
   The most famous structured (third generation) P2P networks        such that compromise between the connectivity and the traffic.
are Chord [2], CAN [8] , Pastry [9], Tapestry [5], Viceroy [10].     Also the diameter of the Ring topology is very high(half of the
The Chord network uses ring graph in its overlay network,            number of nodes within the graph) which is in the Hypercube
CAN uses Torus Graph, Pastry uses some kind of Tree graph            topology as same as the number of its dimmensions. The
which the leaves connected to each other with a ring, Tapestry       Chord topology with Ring Graph has 160 bit addresses for
uses tree graph and Viceroy uses butterfly graph as its overlay       each node, we reduce it to 32 bit to generate Hybercube
topology.                                                            with 32 dimensions and run Chord network over this 32 bit
   We want to show that choosing a proper topology for the           Hypercube. In such networks each node will have 32 neighbors
overlay affects many aspects of the network. Note that most          because the degree of each one is 32; also the diameter of the
of the existed P2P networks choosed their overlay topology           graph is 32 and it means that the distance between any pair
to the respect of theirs algorithm, not the goodness of topol-       of nodes is at most 32 hops. By variation of the graph on the
ogy properties itself. Some of the existed P2P networks are          overlay, the routing algorithm must be also changed. All these
hardly dependable on their topologies. For example, in the           have done as mentioned above by using Planetsim.
CAN network, the algorithm has rigid dependebility to the               The effects of all these changes on the properties we
torus topology or topologies like that. It means that in such        mention on part II dicussed here.
networks, changing the topology maybe lead algorithm to be              1) Network Creation Time: Network creation time is the
impractical. Another P2P networks have less dependebility on         time cosumed for creating overlay graph (in our case is
their topology, for example in the Chord network that use Ring       Hypercube with d dimension) and join specified number of
topology as its overlay topology, the ring can be displaced by       nodes. It completely isolated from the time that the Planetsim
the another topology such as Hypercube, without any serious          simulator spend for simulation operations. The simulation
change in the algorithm. In this paper we do such displacement       operation also spend some time, this time is named Simulation
and prove, by simulation, that choosing more proper graph            time. The summation of this two parameters are named Total
for the overlay layer can affect and improve many important          time. All of these times are calculated but only the Network
properties in the network, such as control traffic, creation time     creation time is useful so we don’t mention the simulation time
and etc.                                                             and total time. We compute the network creation time for the
   Chord network is very popular in researches and academic          Chord network by using both Hypercube by 32 dimensions
works because it proves that the order of network control            and Ring graphs. Fig. 1 shows the Network creation time
traffic caused by chord network is O(log2 (N )) [2]. We want          for 32-D Hypercube graph against Ring graph in the chord
to show that using proper graph for overlay can enhance many         network. As you see in the figure by using Hypercube graph
factors. So we used new graph in overlay and map this graph          the consumed time for creation of the network is much lower
over the Chord network. It leads to much lower traffic in the         than another one for the Ring graph. This time is the time
network. We will prove in this paper , by using simulation, that     that is used for finding successor and predecessor in the Chord
the order of the traffic is as same as for chord but it is about      network by using Ring graph. Hypercube graph doesn’t need
20 percent of that. We use hypercube graph for our topology          to such operations (finding successor and predecessor) because
and mapped it over Chord which is use ring graph by using            in the hypercube the degree of each node is equal to the
planetsim simulator [11].                                            number of dimensions (for this simulation it is 32). So the
   Parameters that we measure in our model are network               connectivity is very rigid in this graph, but by using Ring graph
control traffic, network creation time, and saturation point for      the connectivity for each node is held only with two nodes,


                                                                  100                               http://sites.google.com/site/ijcsis/
                                                                                                    ISSN 1947-5500
                                                     (IJCSIS) International Journal of Computer Science and Information Security,
                                                     Vol. 10, No. 4, April 2012


successor and predecessor. So each node in Chord network         number of nodes. As mentioned in the figure the traffic become
with Ring graph need to keep the connection with previous        much lower for Hypercube overlay against the Ring. The
and next nodes to keep the connectivity of itself by other       reason of this fall in the traffic is as same as the reason for
nodes in the network. the operation such finding successor and    the Network creation time. It is the poor connectivity of the
predecessor and keep them updated consume a lot of time.         Ring graph to the respect of the Hypercube and its needs to
                                                                 find and keep updated successors and predecessors. Since that
                                                                 in the chord network no operation will be done without the
                                                                 existance of the successors and predecessors, so they should be
                                                                 always keeped updated. Such operations produce huge traffic
                                                                 in the network. Also the diameter of the ring is so high; it is an
                                                                 essential reason for producing extra traffic within the network,
                                                                 too.




                 Fig. 1.     Network Creation Time


   For more accuracy in the consumed time for network
creation in the chord network between Hypercube graph and
Ring graph, we compute the ratio between the time consumed
by Hypercube over the time consumed by the Ring. Fig.
2 shows this ratio for different number of nodes. As you                              Fig. 3.   Network Creation Traffic
see the ratio between the time consumed by the Hypercube
overlay is about 0.4 of those consumed for Ring overlay.
                                                                    For the better description of the improved traffic we calcu-
                                                                 lated the ratio between the traffic generated by the Hypercube
                                                                 overlay over the traffic generated by the Ring overlay. the
                                                                 result is shown in Fig. 4. As you say in the figure the
                                                                 traffic generated by the Hypercube is about 20% of the traffic
                                                                 generated by the Ring overlay. as mentioned earlier the traffic
                                                                 generated by the P2P networks is very important factor in such
                                                                 networks because the more traffic cause a collapse. As you saw
                                                                 choosing proper graph in the overlay of the P2P networks can
                                                                 improve many aspects of the network.




                   Fig. 2.    Creation Time Ratio


   2) Network Control Traffic: The time consumed for net-
work to be created is so important but the traffic produced in
the network is much more important for the network because
the more traffic cause the collapse in the network. Network
control traffic is number of messages that are sent by nodes in
the network for creation of the network or joining/leaving the
new nodes to/from the network. We calculate this traffic for
both the Hypercube and Ring overlays in Chord network. The
                                                                     Fig. 4.   Ratio of Network Traffic for Hypercube over Ring’s Graph
traffic is calculated for 10, 100, 1000, 2000,...,10000 nodes.
Fig. 3 shows the traffic for both overlays over the different


                                                              101                                  http://sites.google.com/site/ijcsis/
                                                                                                   ISSN 1947-5500
                                                                (IJCSIS) International Journal of Computer Science and Information Security,
                                                                Vol. 10, No. 4, April 2012


   3) Saturation Point for Hypercube Topology: In this paper                      As mentioned in the Fig. 5 the saturation point for different
we say that the Hypercube has better degree and diameter                       number of dimensions is between about 60% to about 90%
to the respect of Ring, but we don’t say anything about the                    of theoretical number of nodes that can be contained in the
third property of the topologies that is also very important                   network. So we can conclude that however the scalability of
in choosing the proper topology for P2P network. The third                     Hypercube is less than the Ring but it is not bad. Also by
property is the scalability. The topology has good scalability                 choosing 32-D graph that can contain theoretically about 4
if the number of nodes can changed easily. Ring topology                       billion nodes the saturation point is at least about 2.5 billion
is very scalable. It means that any number of nodes can be                     nodes. so it is very better choice for the P2P network to the
putted on the Ring without any problem. It is one of the major                 respect to the Ring graph.
properties of the Ring topology. Hypercube is less scalable.
                                                                                                           IV. C ONCLUSION
The number of nodes in the hypercube is directly related to
the number of dimensions. The number of nodes can be placed                       The P2P’s are popular networks and are used extensively.
on the Hypercube are calculated as 2 to the power of the                       However, the designer of this network did not pay enough
number of dimensions. The problem is that in the real world                    attention to choose proper topology for overlay of these
reach such numbers are impossible. For example in the 8-                       networks. In this paper, we showed that the selection of proper
D Hypercube overlay, theoretically we can put 256 node but                     graph for overlay can effect many factors such as traffic and
in real world when a node want to joining the network and                      time and enhance them. Also we analyze some properties of
assigned to an address, when it see that the address is filled                  Hypercube topology in the P2P networks. In this analysis we
previously with another node, it try to join another time. It is               found the saturation point in different number of dimensions
named a failing in the join operation. In the real world the                   that lead to fail in the network. So we can conclude that
node when fail in joining for several times, it will consider                  not only choosing the topology is important problem but also
the overall joining operation as fail. The failing probability                 choose of the specific graph is very important.
will increase with the increment of number of nodes joined to                                            ACKNOWLEDGMENT
the network. The Ring topology does not related with this
problem because it is fully scalable but in the Hypercube                         The authors would like to thank Dr. M. Kharrazi for her
graph it cause important problem. For avoiding the problem                     insightful comments and Ms. F. Javanmard for pre editing this
of failing in join operation we use 32-D Hypercube graph that                  paper.
theoretically can contain about 4 billion nodes. In addition                                                  R EFERENCES
to this we compute the saturation point for the Hypercube
                                                                                [1] C.-H. Wang and Y.-T. Wu, “Network locality positioning system in
topology for different number of dimensions. We define the                           p2p networks,” in Second International Conference on Ubiquitous and
saturation point in Hypercube as the maximum number of                              Future Networks (ICUFN),, (2010), pp. 182–187.
nodes that can join the network with a certain probability. We                  [2] I. Stoica, R. Morris, D. Karger, M. F. Kaashoek, and H. Balakrishnan,
                                                                                    “Chord: A scalable p2p lookup service for internet applications,” in
set this probability to be 30%, meaning if we try to join K+1                       SIGCOMM, August (2001), pp. 149–160.
nodes, 10 times, to a hypercube with d dimensions (that can                     [3] P. Limin and X. Wenjun, “A binary-tree based hierarchical load balanc-
contain 2d nodes, also k+1 ≤ 2d ), and joining operation failed                     ing algorithm in structured peer-to-peer systems,” Covergence Informa-
                                                                                    tion Technology, vol. 6, no. 4, pp. 42–49, (2011).
in more than 3 times, the saturation point for the hypercube                    [4] Lobb, R. John, C. da Silva, A. Paula, Leonardi, Emilio, Mellia,
with d dimensions considered as k. We calculate saturation                          Marco, Meo, and Michela, “Adaptive overlay topology for mesh-based
point for 5, 6, ... , 14 dimensions Hypercube. The result is                        p2p-tv systems,” in Proceedings of the 18th international workshop on
                                                                                    Network and operating systems support for digital audio and video,
shown in Fig. 5.                                                                    ser. NOSSDAV ’09. New York, NY, USA: ACM, (2009), pp. 31–36.
                                                                                    [Online]. Available: http://doi.acm.org/10.1145/1542245.1542253
                                                                                [5] B. Y. Zhao, J. D. Kubiatowicz, and A. D. Josephl, “Tapestry: An
                                                                                    infrastructure for fault-tolerant wide-area location and routing,” in Tech.
                                                                                    Rep. CSD-01-1141, April (2001).
                                                                                [6] F. Dabek, B. Zhao, P. Druschel, J. Kubiatowicz, and I. Stoica, “Toward
                                                                                    a common api for structured p2p overlays,” in IPTPS, Feb. (2003), pp.
                                                                                    33–44.
                                                                                [7] M. Gharib, Z. Barzegar, and J. Habibi, “A novel method for supporting
                                                                                    locality in p2p overlays using hypercube topology,” in International
                                                                                    Conference on Intelligent Systems, Modelling and Simulation, (2010),
                                                                                    pp. 391–395.
                                                                                [8] S. Ratnasamy, P. Francis, M. Handley, R. Karp, and S. Schenkerl, “A
                                                                                    scalable content-addressable network,” in SIGCOMM, Aug. (2001), pp.
                                                                                    384–389.
                                                                                [9] A. Rowstron and P. Druschel, “Pastry: Scalable, distributed object
                                                                                    location and routing for large-scale p2p systems,” in Middleware, Nov.
                                                                                    (2001), pp. 329–350.
                                                                               [10] D. Malkhi, M. Naor, and D. Ratajczak, “viceroy: A scalable and dynamic
                                                                                    emulation of the butterfly,” in PODC, (2002), pp. 183–192.
                                                                               [11] [Online]. Available: http://planetsim.sourceforge.net/
Fig. 5.   Saturation Point for different number of dimensions in Hypercube




                                                                             102                                  http://sites.google.com/site/ijcsis/
                                                                                                                  ISSN 1947-5500

								
To top