Roadmap-Based End-to-End Traffic Engineering for Multi-hop Wireless Networks Mustafa O. Kilavuz Ahmet Soran Murat Yuksel University of Nevada Reno
Outline Introduction Framework Simulation Results Conclusion and future work
Introduction
Motivation Why load balance the traffic (i.e., traffic engineering) in multi-hop wireless networks? – Mitigate hotspots – Attain higher throughput (aggregate throughput is maxed) – Lifetime of the network (load on nodes/routers is evenly distributed)
Desired Properties Flexible: End-to-end route selection capability (like MPLS) – Source application can control paths the traffic takes Scalable: Do not want to store – Global topology information – Flow state Can we achieve both by in a feasible and scalable manner?
Flexibility: Source-Based E2E Trajectories Defining E2E paths require topology info – hard to get Idea: Decouple the E2E path from the underlying topology, control plane costs could be reduced significantly! Ideal Trajectory Approximate Trajectory Actual Trajectory
Void Area Trajectory-Based Forwarding (TBF) Source Destination Data D. Niculescu B. Nath
Scalability: Roadmaps Need to summarize the congestion state of the global network – hard to gather Idea: Use the adaptive roadmaps concept from robotics S. Bhattacharya, et al
Current Schemes Mostly shortest path – Greedy – Not suitable for load balancing – E.g. GPSR Mostly topology dependent – Not scalable against network changes/dynamics – E.g. DSR
Overall Framework
Routing Framework with Roadmap Roadmap Trajectory Approximator Application-Specific Constraint (e.g., path accuracy, max delay) send(dest, data, constraint) Network Packets with approximate trajectory to the network send(dest, apprx_traj, data) Congestion indications as link weight updates to the roadmap Shortest path on the roadmap as ideal trajectory Path Selection for E2E TE at Routing Layer
Void Area Building the Roadmap
Generating Ideal Trajectory Void Area Source Destination
Feedback: Void Areas Void Area Source Destination Data Feedback
Feedback: Congested Areas Congestion causes packet drops Broadcast feedback – High priority – Small size 50% probability to reroute
Load Balancing Roadmap edge weights are increased as they are being used. Unused edges’ weights are gradually decreased. Change trajectory after sending n packets over it.
Simulation
Simulation Setup Goal: Maximum throughput TBR vs. Greedy Perimeter Stateless Routing (GPSR) Why GPSR? – Similar properties with TBR Geographic Scalable Topology-independent – Good reference for benchmarking Shortest path No end-to-end
Void Area Greedy Perimeter Stateless Routing (GPSR) Source Destination Greedy Forwarding Perimeter Forwarding Greedy Forwarding B. Karp, H. Kung
Simulation Setup Field size1500 x 1500 pixel 2 Wireless node range150 pixels Runtime20s Traffic rate160 Kbps Network density10, 15, 20, 25 – Number of nodes114, 171, 229, 286 Number of traffic flows3, 5, 10 Packet queue size5, 10, …, 50 Reruns16
Simulation: Trajectories Source Destination
Simulation: Roadmap
Results
Work Load Heat Map GPSR Roadmap based TBR
Throughput QDFQDF Q – Packet queue size of nodes D – Network density (Average number of neighbors) F – Number of traffic flows (Source – destination pairs) TBR has higher throughput overall GPSR has good throughput on sparse networks High number of flows increases congestion, reduces throughput High queue size increases throughput
Hop Count QDFQDF Q – Packet queue size of nodes D – Network density (Average number of neighbors) F – Number of traffic flows (Source – destination pairs) TBR has longer routes to avoid congestion and to do load balancing Network density is not a major factor but causes GPSR spikes because of perimeter mode
Packet Delay QDFQDF Q – Packet queue size of nodes D – Network density (Average number of neighbors) F – Number of traffic flows (Source – destination pairs) TBR packet delay increases within acceptable amounts Large queue size causes more delay
Conclusion and Future Work
Conclusion Mobile scenarios Algorithms optimization Improvements to roadmaps – Construction (regular patterns) – Better methods for ideal trajectory – Local vs. global
Questions & Answers
Backup Slides
Void Area Trajectory-Based Routing (TBR) Data Source Destination M. Yuksel et al. Ideal Trajectory Approximate Trajectory Special Intermediate Node (SIN)
Work Load distribution QDFQDF Q – Packet queue size of nodes D – Network density (Average number of neighbors) F – Number of traffic flows (Source – destination pairs) TBR distributes load better Load is more balanced in dense networks High number of flows puts more load on central nodes
Contributions The concept of minimizing routing state under application-based constraints. Formulation of the trajectory approximation problem minimizing the routing state. Proof that the trajectory approximation problem is NP- hard. Solutions to solve the trajectory approximation problem. Customized the trajectory approximation problem for power-scarce networks. A roadmap-based mechanism for end-to-end traffic engineering for multi-hop wireless networks.
Roadmap Simulations Source Node Destination Node Void Area Data Packet Approximate Trajectory
Roadmap Simulations Roadmap Edges Ideal Trajectory Roadmap Vertices Source Node Destination Node Approximate Trajectory
Roadmap Simulations
Routing Protocols Destination-Sequenced Distance Vector (DSDV) Ad hoc On Demand Distance Vector (AODV) Greedy Perimeter Stateless Routing (GPSR) Distance Routing Effect Algorithm for Mobility (DREAM) Dynamic Source Routing (DSR) Trajectory-Based Forwarding (TBF) Trajectory-Based Routing (TBR) Roadmaps in robotics
Destination-Sequenced Distance Vector (DSDV) Destination Next hop Source Destination Routing table
Ad hoc On Demand Distance Vector (AODV) RREQ Source Destination RREP
Distance Routing Effect Algorithm for Mobility (DREAM) Source Destination
Dynamic Source Routing (DSR) | 2 1 | 3 1 | 2 | 4 1 | 2 | 4 | 6 1 | 2 | 4 | 6 | 7 Source Destination 1 | 2 | 4 | 5 1 | 2 | 4 | Data 1 | 2 | 4 | 6 | 7 | 5
Comparison DSDVAODVGPSRDREAMDSRTBR Flexibility Scalability (State) Scalability (Messaging) Reachability Computation Type ProactiveReactive
Cost Comparison Error tolerance: 5% GA performs pretty close to the exhaustive search Longest representation heuristic is not bad Exhaustive Search Equal error heuristic did not do well Equal Error Longest Representation M. Kilavuz et. al. Minimizing multi-hop wireless routing state under application-based accuracy constraints, MASS 2008
Equal Error Longest Representation Time Comparison Equal Error heuristic runs in no time Exhaustive search takes too much time These run in reasonable amount of time Error tolerance: 5% M. Kilavuz et. al. Minimizing multi-hop wireless routing state under application-based accuracy constraints, MASS 2008