TopologyAware Overlay Construction and Server Selection Sylvia Ratnasamy

Topology-Aware Overlay Construction and Server Selection Sylvia Ratnasamy Mark Handley Richard Karp Scott Shenker Infocom 2002

Connections of a node

Introduction l Problem: Inefficient routing in large-scale networks l l l In large-scale overlay networks, each node is logically connected to a small subset of other participants. Due to the lack of effort to ensure that application-level connectivity is congruent with underlying IP-level network topology Basic Idea: Optimize routing paths in network l l Define a binning scheme whereby nodes partition themselves into bins Nodes that fall within a given bin are relatively close to one another in terms of network latency

Outline l l l Introduction Distributed Binning Topologically-aware construction of overlay networks Topologically-aware server selection Conclusion

Extracting proximity information l Measuments that can be used to derive topological information: l traceroute: s l l l 2 sec a BGP routing table: l l l intended for network diagnostic purposes, too heavy-weight, excessive load on the network, disabled ICMP at some sites for security not directly available for end users, requires privilege or third party service Network latency: l l 7 sec b often a direct indicator of network performance, light-weight, end-to-end measurement, non-intrusive manner 5 sec c t

Distributed Binning l Goal: l l l Have a set of nodes independently partition themselves into disjoint “bins” Nodes within a single bin are relatively closer to one another than to nodes not in their bin Scheme: l l A well-known set of machines that act as landmarks on the Internet Form a distributed binning of nodes based-on their relative distances l A node measures round-trip-time (RTT) to each landmark and orders landmarks in order of increasing RTT l Every node has an associated ordering of landmarks(or bin)

Distributed Binning l Scheme: (Cont. ) l After finding ordering, we calculate absolute values of each RTT in ordering as follows l l l We divide the range of possible latency values into a number of levels. Convert RTT values into level number and obtain a level vector Example: Level 0 0 -100 ms Level 1 100 -200 ms Level 2 > 200 ms l 2 l 1 57 ms 232 ms l 3 Node A’s bin becomes “l 2 l 3 l 1: 0 1 2” A l 117 ms Topologically close nodes likely to have same ordering and belong to same bin

Distributed Binning Scheme

Performance of Distributed Binning l l Even though it is clearly scalable, does it do a reasonable job? Metric used: average inter-bin latency = average latency from a given node to all nodes not in its bin average intra-bin latency = average latency from a given node to all nodes in its bin l A higher gain ratio indicates a higger reduction in latency, hence more desirable

Performance of Distributed Binning l Datasets or test topologies: l TS-10 K and TS-1 K: l Transit-Stub topologies with 10000 and 1000 nodes respectively. l 2 -level hierarchy l PLRG 1 and PLRG 2: l Power-Law Random graph with 1166 and 1779 nodes l Edge latencies assigned randomly NLANR: l Distributed network of over 100 active monitors l Systematically perform scheduled measurement between each other l

Performance of Distributed Binning l Other binning algorithms used in experiments: l l Random Binning: l Each nodes selects a bin at random l acts as a lower bound for the gain ratio Nearest Neighbor clustering: l Each node is initially assigned to a cluster itself. l At each iteration, two closest clusters are merged into a single cluster. l The algorithm terminated when the required number of clusters is obtained _

Performance of Distributed Binning l Experiments: Effect of number of landmarks (#level=1) Effect of number of levels (#landmarks=12)

Performance of Distributed Binning l Experiments: Comparison of different binning techniques(#levels=1)

Topologically-aware construction of overlay networks l Two types of overlay networks l l l Structured: l Nodes are interconnected in some well-defined manner(Application-level) Unstructured: l Much less structured like Gnutella, Freenet Metric for evaluation:

Topologically-sensitive CAN construction l Content-Addressable Network l l Scalable indexing system for large-scale decentralized storage applications on the Internet Built around a virtual multi-dimensional Cartesian coordinate space Entire coordinate space is dynamically partitioned among all the peers, i. e. every peer possesses its individual, distinct zone within the overall space A CAN peer maintains a routing table that holds the IP address and virtual coordinate zone of each of its neighbor coordinates

2 D CAN Example State of the system at time t Peer Resource Zone x In this 2 dimensional space, a key is mapped to a point (x, y)

Routing in CAN d-dimensional space with n zones y • • Routing length: (x, y) Peer Q(x, y) Query/ Resource path of • Algorithm: Q(x, y) Choose the neighbor nearest to the destination key

Contribution to CAN l l Construct CAN topologies that are congruent with underlying IP topology Scheme: l With m landmarks, m! such ordering is possible l l For example, if m=2, then possible orderings are “ab” and “ba” We partion the coordinate space into m! equal sized portions, each corresponding to a single ordering l l l Divide the space along first dimension into m portions Each portion is then sub-divided along the second dimension into m 1 portions Each of these are divided into m-2 portion and so on… l When a node joins CAN at a random point, the node determines its associated bin based-on delay measurement l According to its landmark ordering, it takes place in the correspanding portion of CAN

Gain in CAN using Distributed Binning Stretch for a 2 D CAN; topology TS-1 K; #levels=1 Stretch for a 2 D CAN; topology PLRG 2; #levels=1

Topologically-aware construction of unstructured overlays l l Aims much less structured overlay such as Gnutella, Freenet Focusing on the following general problem in unstructured overlays: “Given a set of n nodes on the Internet, have each node picks any k neighbor nodes from this set so that the average routing latency on the resultant overlay is low” l Optimal overlay is NP-hard, so used some heuristic called Short-Long

Topologically-aware construction of unstructured overlays l Short-Long Heuristic l A node picks its k neighbors by picking k/2 nodes closest to itself and then picks another k/2 nodes at random l Well-connected pocket of closest nodes and inter-connections to far pockets with random picks Current Node Nearby Nodes Distant Nodes Other Nodes l Bin. Short-Long (Contribution) : l A node picks k/2 neighbors at random from its bin and picks remaining k/2 at random

Gain in Unstructured Overlay using Distributed Binning Unstructured overlays; TS-10 K; #levels=1; #landmarks=12

Topology-aware server selection l Replication of content over Internet gives rise to the problem of server selection l Parameter: Server load and distance(in term of Network Latency) l _Replication Server Client

Topology-aware server selection l Server selection process with distributed binning works as follows: l l l Compared performance to 3 schemes: l l If there exist one or more servers within same bin as client, then client is redirected to a random server from its own bin If no server exists within same bin as client, then an existing server whose bin is most similar to client’s bin is selected at random Random: Client selects server at random Hotz Metric: Uses RTT measure from a node to well known landmarks to estimate internode distance (Triangle inequality) Cartesian Distance: Calculates Euclidean distance using level vector of node and selects the server with minimum distance Measurement for evaluation:

Topology-aware server selection Comparison of different schemes under following conditions: • 12 landmarks and 3 levels • 1000 servers for TS-10 K, 100 servers for TS-1 K, PLRG 1 and PLRG 2 and 10 for NLANR

Topology-aware server selection-Node Perspective CDF of latency stretch for TS-10 K data CDF of latency stretch for NLANR data

Conclusion l l l _ Described a simple, scalable, binning scheme that can be used to infer network proximity information Nature of the underlying network topology affects behavior of the scheme It is applied to the problem of topologically-aware overlay construction and server selection domains Three applications of distributed binning is given: l Structured Overlay l Unstructured Overlay l Server selection A small number of landmarks yields significant improvements. Can be referred as network-level GPS system

Happy end! Thank you for your patience!