Transcript PPT
Meridian: A Lightweight Framework for Network Positioning without Virtual Coordinates Bernard Wong, Aleksandrs Slivkins and Emin Gün Sirer Cornell University http://www.cs.cornell.edu/~bwong/meridian What is Meridian? Status Central Leader Selection • • • • • • • • Meridian is a lightweight framework for performing node selection based on a set of network positioning constraints Based on a small number of query-directed measurements Meridian is organized as a loosely structured overlay – Gossip-based membership tracking – Each node keeps O(logN) state The goal of the overlay is to have each node know enough local information to authoritatively answer geographic queries for its region of the network Queries regarding distant nodes are handed off to a neighbor in that region. – Essentially, the idea is for each node-to-node query hand-off to “zoom in” to the solution space – Each hop exponentially reduces the distance to a target within the solution space Meridian solves frequently encountered network positioning problems without computing network coordinates – No embedding error – Low complexity – Scalable without CAN • • • • Find closest node – Web server, game server, etc. Construct efficient overlays – Find the most central node to a set of nodes to act as the set’s hub node – Hub nodes get promoted higher in the tree hierarchy Satisfying Service Level Agreements – Find a node that meets multiple latency constraints – Build clusters with a maximum inter-node latency Discover good one-hop indirect routes – Find lowest latency one-hop overlay route when direct route fails • • • • The protocol works as follows: 1. Client sends a “closest node discovery to target T” request to a Meridian node A 2. Node A determines its latency d to T 3. Node A probes its ring members between d/2 and 3d/2 to determine their distances to the target 4. The request is forwarded to the closest node thus discovered that is at least β times closer to the target than A 5. Process continues until no node that is β times closer can be found Minimizes average distance to a set of targets instead of one target Protocol similar to closest node discovery, but uses davg instead of d For this example, node A is the best leader out of nodes A to H, where nodes A to H are also the targets D C E A B G F H Evaluation • Closest Node Discovery Motivating problems • • We have proven in a theoretical analysis that – Meridian’s multi-hop search terminates in O(logN) hops – Error will remain constant given each node has O(logN) ring members The results hold for Internet latencies that can be modeled as growth constrained or low doubling dimension metrics – Both metrics are more general than a low dimension Euclidean metric Full implementation and deployment of three applications using the Meridian framework 1. Closest node discovery • Finds the closest (lowest latency) node to a given target 2. Central leader selection • Select the most centrally located node to a given target set 3. Multiple constraints system • Discovers a node within latency bounds of multiple targets We have evaluated these applications through both simulation parameterized by a large-scale network measurement study, and a deployment on PlanetLab. For our measurement study, we collected node-to-node round-trip latency measurements for 2500 nodes and 6.25 million node pairs on the Internet using the King measurement technique. Closet Node Discovery Light bars show the median error for discovering the closest node. Darker bars show the inherent embedding error with coordinate systems. Meridian's median closest node discovery error is an order of magnitude lower than schemes based on embeddings. Central Leader Selection 2 Client node Target node T Initial node A Closest node B 1 With larger group sizes, the central leader selection algorithm is able to find a centrally situated node more frequently. Meridian is significantly more accurate than Vivaldi for all tested configurations. Ring member of A Ring member of B Forwarding of query Latency probe Multiple Constraint System The percentage of successful multi-constraint queries is above 90% when the number of nodes that can satisfy the constraints is 0.5% or more. Multi-resolution Rings The multi-resolution rings are a set of concentric, non-overlapping rings with exponentially increasing radii • A neighbor is placed in the appropriate ring of a node depending on the peer’s latency to it Exponentially increasing ring radii provides authoritative local information A ring membership management scheme is used to further improve the diversity of nodes within each ring – The set of neighbors that form the largest hypervolume k-polytope in a non-exported local coordinate space is selected – For k = 3, three nodes from each ring that form the largest triangle are chosen • • Multiple Constraint System • • A s r=α r=αs2 • Discover a node in the solution space that satisfies the given constraints Distance to the solution space s is calculated as the sum of the square distance away from meeting each constraint – In this example, αa, αb , and αc are the latency constraints to targets A, B, and C respectively Protocol similar to closest node discovery, but uses s instead of d αa • αb Summary B A C αc • • • Meridian can solve frequently encountered network positioning problems Meridian is accurate, resilient to network dynamics, simple and lightweight – Avoids many of the pitfalls of virtual coordinates It remains to be seen whether Meridian can be applied to other network positioning problems