R3: Robust Replication Routing in Wireless Networks with Diverse Connectivity Characteristics Xiaozheng Tie, Arun Venkataramani, Aruna Balasubramanian University of Massachusetts Amherst University of Washington UNIVERSITY OF MASSACHUSETTS, AMHERST • Department of Computer Science Wireless routing compartmentalized Protocols designed for well-connected meshes OLSR, ETT, ETX, EDR, … Research question: Can we design a simple routing protocol that Protocols designed for intermittentlyensures robustMANETs performance across networks connected withAODV, diverse connectivity DSDV, DSR, … characteristics all the way from well-connected meshes to mostlyProtocols designed for sparselydisconnected DTNs and everything in between? connected DTNs DTLSR, RAPID, Prophet, Maxprop, EBR, Random, … UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 2 Outline Compartmentalized design harmful Quantifying replication gain R3 design and implementation Evaluation Conclusion UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 3 Fragile performance Protocols perform poorly outside target environment Normalized delay Replication wasteful 2.1x Mesh testbed Mesh protocols perform poorly in DTNs Normalized delay DTN protocols perform poorly in mesh No contemporaneous path 2.2x DTN testbed UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 4 Spatial connectivity diversity DieselNet-Hybrid Vehicular DTN + Wifi Mesh 20 buses in Vehicular DTN 4 open AP WiFi mesh clusters < 100 contacts 100 – 200 contacts > 200 contacts UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 5 Temporal connectivity diversity Haggle Fraction of connected nodes Mobile ad hoc network 8 mobile and 1 stationary imotes 9 hour trace in Intel Cambridge Lab UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 6 Compartmentalized design harmful 1. Fragile performance under spatiotemporal diversity 2. Makes interconnection of diverse networks difficult Makes management difficult Conflates cross-layer concerns Stifles long-term innovation UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 7 Outline Compartmentalized design harmful Quantifying replication gain R3 design and implementation Evaluation Conclusion UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 8 Replication: Key difference DTN Sparsely connected Replication MANET Intermittently connected Mesh Well connected Forwarding Key question: Under what conditions and by how much replication improves performance? UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 9 Model to quantify replication gain X1 X2 Src Dst Xi Random variable denoting the delay of path i Xn • Expected delay of forwarding minE[X1],E[X2 ],...,E[Xn ] • Expected delay of replication (1) E[min{ X1, X 2,...,X n }] Replication gain (1) UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 10 Example of replication gain X1 Src Dst X2 P(X1 0.1) 90% P(X1 10) 10% P(X 2 0.3) 90% P(X 2 30) 10% E(X1) 1 E(X 2 ) 3 Replication gain depends on path delay distributions, not justdelay expected value • Expected of forwarding minE[X ],E[X ] min1,3 1 1 • Expected delay of replication Replication gain 2 (1) E[min{ X1, X 2}] 0.2 5 (1) 5x delay improvement UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 11 Replication gain vs. number of paths Trace-driven analysis on DieselNet-DTN and Haggle Two paths suffice to capture much of the gain Vehicular DTN in DieselNet Haggle UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 12 Outline Compartmentalized design harmful Quantifying replication gain R3 design and implementation Evaluation Conclusion UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 13 R3 design overview Link-state Estimate per-link delay distribution Replication Select replication paths using model Adapt replication to be load-aware Source routing along selected path(s) Dst Src Y1 Y3 Y2 UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 14 Estimate link delay distribution Link delay Link availability delay Delay to successfully transfer packet Node 2 Node 1 T=0 T=1 T=0: probe 0 unacked T=1: probe 1 unacked T=2: probe 2 acked at T=2.1 T=2 Delay = 2.1-0 = 2.1 Delay = 2.1-1 = 1.1 Delay = 2.1-2 = 0.1 Delay samples = {2.1, 1.1, 0.1} UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 15 R3 design overview Link-state Estimate per-link delay distribution Src Dst Replication Select replication paths using model Adapt replication to be load-aware Source routing along selected path(s) UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 16 Path selection using model X1 X2 Src Dst First path Xi Xn Path i s.t. E[X i ] is smallest Selected using Dijkstra’s shortest path algorithm Second path Path j s.t. E[min{X i , X j }] is smallest Selected using delay distributions and model UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 17 Adapting replication to load Problem Replication hurts performance under high load Solution Load aware replication actual_delay > t * model_estimated_delay Start Replication Forwarding actual_delay ≤ t * model_estimated_delay UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 18 R3 design overview Link-state Estimate per-link delay distribution Src Dst Replication Select replication paths using model Adapt replication to be load-aware Source routing along selected path(s) UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 19 Outline Compartmentalized design harmful Quantifying replication gain R3 design and implementation Evaluation Deployment on a DTN and mesh testbed Simulation based on real traces Emulation using mesh testbed Conclusion UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 20 R3 Deployment DieselNet DTN testbed 20 buses in a 150 sq. mile area Mesh testbed 16 nodes in one floor Simulator validation using DieselNet deployment < 10% of deployment result UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 21 R3 Trace-driven simulation Experimental settings Temporal diversity inherent in Haggle Spatial diversity inherent in DieselNet-Hybrid Varying load Compared protocols Replication: RAPID, Probabilistic Forwarding: DTLSR, AODV, OLSR Multi-configuration: SWITCH (RAPID+OLSR) UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 22 Robustness to temporal diversity Delay (min) Goodput (pkt/min) Simulation based on Haggle trace Hour Hour R3 reduces delay by up to 60% R3 increases goodput by up to 30% UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 23 Robustness to spatial diversity 6 9 1 5 8 2 4 Delay (min) Simulation based on DieselNet-Hybrid trace 3 7 Grid R3 improves median delay by 2.1x UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 24 Emulating intermediate connectivity Mesh-based emulation approach R3 reduces delay by up to 2.2x Delay (sec) Brings link up and down to vary connectivity Emulates connectivity diversity (but not mobility) Hour UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 25 Conclusion Compartmentalized design harmful R3 ensures robust performance across diverse connectivity characteristics Unified link metric based on delay distributions Replication based on delay uncertainty model Adaptive replication based on network load Thank you! UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 26 UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 27 Robustness to varying load Simulation based on DieselNet-Hybrid trace R3 reduces delay by up to 2.2x over SWITCH UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 28 More paths in DieselNet-Hybrid Average delay when R3 uses k=2, 3, 4, 5 replication paths. UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science 29