Virtual Direction Routing VDR for Overlay Networks BowNan

Virtual Direction Routing (VDR) for Overlay Networks Bow-Nan Cheng (RPI), Murat Yuksel (UNR), Shivkumar Kalyanaraman (RPI) Motivation VDR Details • The explosion of peer-to-peer systems in recent years has prompted research into finding scalable and robust seeding and searching methods to support these overlay networks. Initial work relied on network flooding to find information and while robust, lacked scalability. In effort to scale large networks, many have looked at structured approaches to the problem by imposing some sort of structure to the network topology and routing based on that structure. To support search queries, a robust overlay network with routing policies must be in place as search fails to address the actual data traversal path. In much the same way, routing in overlay networks have evolved from pure flooding techniques to structured techniques. In our work, we attempt to drop the need for imposing a specific structure on the overlay network and introduce a technique to scalably route packets through an unstructured overlay network. VDR: Neighbor to Virtual Interface/Direction Mapping 30 % 8 = 6 15 % 8 = 7 30 10 % 8 = 2 10 Hierarchy/Structured neighbors connected by interfaces or neighbors under a certain RTT latency away (logical neighbors) 10 2 1 3 1 68 4 7 5 6 15 30 68 68 % 8 = 4 26 % 8 = 2 Neighbor to Virtual Interface Mapping 0 5 26 § Each neighbor ID is hashed to 160 bit IDs using SHA-1 (to standardize small or large IDs) § The virtual interface assigned to the neighbor is a function of its hashed ID (Hashed ID % number of virtual interfaces) VDR: Proactive element |10 – 1| = 9 |26 – 1| = 25 26 10 3 68 4 0 1 5 6 67 48 7 15 • Unstructured P 2 P systems make little or no requirement on how overlay topologies are established and are easy to build and robust to churn • Typical Search Technique (Unstructured Networks) Flooding Normalized Flooding • Flooding / Normalized Flooding 4 28 30 • Structured vs. Unstructured Overlay Networks 3 |5 – 1| = 4 |13 – 1| = 12 Ex: Seed Source: Node 1 5 10 5 6 0 3 7 4 • • Default Values Nodes / # of Virt Intf 50, 000 / 8 Simulation Cycles 150 Churn Percentage 0% - 50% every 5 cycles Seed/RREQ TTL 10 – 100 hops Seed Entry Expiry 10 Cycles (under churn) Number of Queries 1000 Randomly Gen. Effect of Seed/Query TTL on VDR, VDR-R, and RWR Effect of # of Virtual Interfaces on VDR, VDR-R, and RWR Effect of Average # of Neighbors on VDR, VDR-R, and RWR Effect of % Network Churn on VDR, VDR-R, and RWR Flooding Normalized Flooding Virtual Direction Routing VDR – Random NB Send (VDR-R) Random Walk Routing (RWR) 0 5 5 6 |14 – 1| = 13 |22 – 1| = 21 13 Packetized Simulations with Peer. Sim Scenarios Parameter Simulation Results 2 1 • Reachability – Percentage of nodes reachable by each node in network • State Complexity – The total state info (+ spread) maintained in the network • Average End-to-End Path Stretch – Average VDR Path vs. Shortest Path • Average Load Network-wide Random Walk Unstructured/Flat Scalable 2 1 Default Simulation Parameters • VDR-R: VDR with random neighbor forwarding (no biasing) • RWR: Data is seeded in 4 random walks and 4 walkers are sent for search 10 1 Metrics Comparison Protocols Paths from Rendezvous Nodes-to-Destination Nodes are formed by periodically sending announcement packets out orthogonal directions Trends – From Flood-based to Unstructured Scalable Flood-based Neighbors are either physical 26 15 Evaluation Metrics / Scenarios 7 55 22 14 State Seeding – State info forwarded in orthogonal directions, biasing packets toward IDs that are closer to SOURCE ID. Packets are forwarded in virtual straight lines. VDR: Reactive element • High Reach, Low path stretch, Not scalable • Random Walk • Need high TTL for high reach, Long paths, Scalable, but hard to find rare objects • Virtual Direction Routing • Globally consistent sense of direction (west is always west) Scalable interface to neighbor mapping • Routing can be done similarly to ORRP Virtual Direction Routing Paths from Source Nodes-to-Rendezvous Nodes are formed by sending route request (RREQ) packets and waiting for route reply (RREP) packets. These RREQ packets are sent on demand |10 – 12| = 2 |26 – 12| = 15 26 10 5% drop 10 1 12% drop • Focus (for now) • Small world approximations VDR Introduction VDR Primitive - Local sense of direction leads to ability to forward packets in opposite directions 180 o S 2 1 Random Walk VDR – 2 Basic Primitives A Rendezvous Points T B 1. Local directionality is sufficient to maintain forwarding of a packet on a straight line 2. Two sets of orthogonal lines in a plane intersect with high probability even in sparse, bounded networks Question: Can 1 hop neighbors in overlay networks be consistently “mapped” to a local virtual direction such that be forwarding in virtual orthogonal lines, a high chance of intersection (and search success) results? Two Components of VDR Virtual Direction Routing Neighbor to Virtual Interface Mapping Overlay Routing • Proactive Element • Reactive Element 3 68 4 0 1 5 6 67 48 3 7 15 4 28 30 Ex: Route Request: Node 12 RREQ Source: Node 1 2 1 |5 – 12| = 7 |13 – 12| = 1 5 10 5 6 0 3 7 4 13 5 6 |6 – 12| = 6 |38 – 12| = 26 13 38 0 7 Conclusions 6 Route Request – RREQ packets are forwarded in orthogonal directions, biasing packets towards REQUESTED ID 68 5 RREQ Path 30 1 Virtual view of VDR Route Request process: 46 10 26 13 RREP Path Rendezvous Node 6 38 2 RREQ: Node 12 48 Seed Path VDR Route Request Virtual View 12 Node 1 sends out a RREQ looking for Node 12. Once the RREQ intersects a rendezvous ndoe, a RREP is sent back. The virtual path of the data goes from the source node to the rendezvous to the destionation • VDR reaches 3. 5% more nodes than VDR-R and 9% more nodes than our modified random walk routing strategy (RWR) • VDR shows a 3 -4 X reach retention rate going from 0% to 50% network churn compared to VDR-R and RWR, showing itself to be much more robust to network churn • VDR increases reach with fewer number of virtual interfaces because of its biasing technique. Gains disappear if the number of neighbors is smaller than the number of interfaces • Increasing the number of neighbors generally increases reach and end-toend path stretch • VDR states are not well distributed and states and load is not spread evenly • VDR paths exhibit high path stretch compared to shortest path but good path stretch compared to pure random walk Relevant Publications • B. Cheng, M. Yuksel, S. Kalyanaraman, “Virtual Direction Routing for Overlay Networks, " Proceedings of IEEE International Conference on Peer-to-Peer Computing (P 2 P), Seattle WA, September 2009. This material is based upon work supported by the National Science Foundation under Grants 0627039, 0721452, 0721612 and 0230787. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.