Sunteți pe pagina 1din 8

REPORT

1) 2) 3) 4) 5) 6) 7)

Introduction Evolution of schemes over years Terminology/Algorithms Used Description of Schemes Other Topics Explored Conclusion Future Work

INTRODUCTION Compact routing scheme aims to achieve the best possible tradeoff between Routing table size at each node and the stretch i.e. ratio of distance between any two nodes in given scheme to shortest distance between those two nodes in a distributed network. There are two types of routing being followed till date to attain shortest path routing, but both of them have limitation in terms of scalability or header size for accommodating larger networks. The first way is to store next hop information of all the nodes in the routing table at every node, in which case the routing table size would be of the order of O(n), another approach is to have a source routing in which the header/label stores the path to take for routing between the two nodes. In this case the label size becomes of the order O(n) hence imposing limitation on network size.

Considering the large scale networks such as internet in the world today, the idea of compact routing emerged. The idea is to minimize storage requirement, at the cost of taking a slightly larger path, where the worst case stretch too is bounded.

The first approach was the hierarchical routing, in which a tree was created from given graph of nodes and then parent child association was employed to route. The bounds attained were.......... In the paper Hierarchical Routing for larger networks- performance evaluation and optimization by Leonard Kleinrock and Farouk Kamoun. Aim is to reduce routing table length by hierarchical partitioning of the network, hence represented as a tree. The network is divided into some m level clusters. Kth level cluster consists of all ( k-1)th level clusters recursively. The number of k-1 level clusters constituting the kth level clusters is referred to as Degree of kth level cluster.

Next philosophy was interval routing[7], where a depth first search enumeration was done for the nodes of graph and the interval of nodes under the nodes was stored, giving the range in which case the routing was to be in the said node, otherwise the data was to be forwarded to the parent node. Till the message reaches root which has information of all the nodes in its table. The bounds attained by this scheme were .................

Then came landmark based routing by Cowen [1], which was based on the idea of postal addressing system. In this scheme a set of highly connected nodes are chosen as the landmark nodes, where any local address is identified in terms of a well known landmark. A neighborhood of some fixed size is formed for all the nodes. Every node has the information of the location of the neighboring nodes and the edge to take for reaching all the landmarks nodes stored in its routing table. Though at the landmark only the information of rest all landmarks is stored, as any incoming packet has the edge information from landmark to the destination node in the label itself. Every node is identified by a label L (node_name, Landmark_name, edge_from_landmark_to_node). This works for non weighted distributed networks and achieves the bound of RT size O(n2/3 log 4/3

n) and stretch bound of 3. This scheme uses extended dominating set and nodes in maximum neighborhoods as candidates for landmark. It was the best bound achieved till the time it was published.

Thorup Zwick [2] improves on the approach of Cowen [1] by using probabilistic model for Landmark selection. It improves the bound of the landmark selection and routing table to O(n1/2logn) while stretch bound remains 3. This paper also describes ways to reduce label size while routing in a tree. The minimum label size bound achieved is (1+O(1))log2n bit. With the help of the labels of source and destination nodes, routing can be done in a tree in constant time. For improving LM selection , it uses a sample function recursively. In every iteration it passes a node which exceeds the no of nodes than a size of 4*n/s and the sample returns s no. of nodes from that set with each node assigned equal probability to be selected as landmark . Algorithm center(G,s) A; WV; While W do { AA sample(W,s); C(w) {v V|(w,v)< (A,v)} fore very w V W{wV | C(w)>4n/s}; } Return A;

The first labeling scheme described is based on assigning weight to the nodes in the tree as the number of descendents of the concerned node. Then segregating the nodes as heavy or light weight nodes based on fraction of nodes under it. The path from root to the node concerned in terms of light nodes traversed along the path is stored. The routing information at any node consists of (v, fv, hv, Hv, Pv) where v is the node, fv is the largest descendant of v, hv is the first heavy child of v, Hv is the list of heavy nodes and Pv is the port numbers of corresponding heavy nodes. Hv(0) stores the parent of the node and Pv(0) is the port of the parent of the node. Routing information stored at v =(v, fv,hv,Hv,Pv)= O(b) words Label of any node consists of (v, Lv)

where Lv is the port number of the light levels from root to the node v. Routing Algorithm Label(v)=(v , Lv) at node w If w=v, done else check if v (w , fw) a) if not a descendent f/w to parent of w using Pv[0] b) else if descendent check if v (hw , fw) search Hw and get corresponding Pw else if light descendent search Lv[lw]

In the second approach of labeling a single heavy child is considered and rest of the light child are assigned weights. All the port numbers are stored as one string

and port numbers are encoded in varying lengths as per bits required to code them . If for vertex v, v is heavy child & v0, v1,.,vd-1 light children s.t. sv >= sv0>= sv1>=.>=svd-1 ; assign edge (v, vi ) for 0<=i< d & (v,v) port no d. Label(v)=(v, Lv, Mv) eg. lv=(2,0,5,3) =(11 101 0 10) & M=(10 100 1 10)

Masking bits are used to identify the location where a new port no starts. Eg. For port no.s (2, 3,5,6) coding will be (10,11,101,110) and the masking bits will be (10,10,100,100) where 1 indicates start of new port no. Label of any node v is (v, Lv, Mv) and required 3.42log2n bits. The length of Lv and Mv where Lv is the bit string and Mv is the corresponding masking bit string. In the third and final approach the bound of (1+O(1))log2n is achieved. Here the approach followed is to code the bits further as code(s). code(length of s). code(length of length of s). Code(s)=s.bin(||s||,|s|).bin(||s||,||s||) Label(v)=ID(v) + RT(v) ID(T; v) =ID(Tv ; v):code(i):code(sj ):code(port( =code(i):code(sj ) otherwise ; ))if v

Peleg [3] tried to solve the question of how to identify distance between nodes by the labels of the source and destination nodes. The scheme is based on Tree separators and assigning the corresponding distance to each node in given branch in the subtree. A separator splits the tree into branches having nearly n/2 members each. No separator is selected when a leaf node is obtained. Each node is then identified by concatenation labels from all its separators. For n vertex weighted trees with M bit edge weights, label size O(Mlogn+log2n) bit .

Label of a node p triples J(v)=(I(v0), dist(v,v0,T), i), Tree separator v0 labeled as (I(v0), 0, 0). If vertex v is internal to subtrees at level p-1 & becomes separator at level p then Label(v)= J1(v) . J2(v).Jp(v) The distance between any two nodes is given by sum of distances of the label with respect to a common separator in which the two nodes are not on same branch. Using the concept of Peleg and Thorup Zwick scheme of labeling the nodes in O(1+logn), Brady& Cowen [4] came up with the scheme which is meant for power law graphs(the degree distribution follows a power law) like internet. The paper achieves stretch bound of 3 and RT size of O(n1/2).. This scheme identifies a highly connected node h and builds a core of diameter d around it, named as d-core. while the remaining area is named as d-fringe. A minimum spanning tree is formed for all nodes of the graph using h as the root, while separate trees are

formed in the fringe for all connected components. Also the edges not used in the tree formation which are removed to make graph acyclic in the fringe are taken and minimum spanning tree is created using those edges. Now Peleg label is associated with every such label. For routing between source and destination all the trees are explored for presence of these two nodes( Source and Destination). The tree with minimum Peleg distance between the two nodes is chosen to route and the labeling scheme by thorup zwick is used for routing.

Conclusion

Now the role of Thorup/Zwick and Brady/Cowen scheme to be deployed in current world internet scenario is being explored. A new kind of protocol based on philosophy of BGP protocol is envisaged which supports IPv4.

Other Topics studied

1) Wormhole Routing 2) Voronoi Diagrams 3) TZ Also a tree based routing scheme was developed by TZ in [5] to identify the distance between any two nodes in a graph.

References 1) Compact Routing with Minimum stretch. Lenore J. Cowen: J. of Algorithms, 38:170-183, 2001 2) Compact Routing Schemes. Mikkel Thorup and Uri Zwick : In proceedings of the 13th Annual ACM Symposium on Parallel Algorithms and Architectures, pages 1-10. ACM, July 2001 3) Proximity-Preserving Labeling Schemes and Their Applications. David Peleg : WG99 Proceedings of the 25th International Workshop on Graph Theoretic Concepts in Computer Science, pages 30-41. 4) Compact routing on power-law graphs with additive stretch. A. Brady and L. Cowen. In ALENEX, 2006. 5) Approximate distance oracles. Mikkel Thorup and Uri Zwick: In Proc. 33rd ACM Symp. On Theory of Computing, pages 183-192, May2001. 6) Compact routing on Internet-like graphs. D. Krioukov, K. Fall, and X. Yang. In INFOCOM, 2004. 7) Labelling and implicit routing in networks. N. Santoro and R. Khatib :The Computer Journal, 28(1):5-8, 1985.

S-ar putea să vă placă și