Using Directionality in Mobile Routing Bow-Nan Cheng MIT

Using Directionality in Mobile Routing Bow-Nan Cheng (MIT LL) Murat Yuksel (Univ Nevada - Reno) Shivkumar Kalyanaraman (IBM IRL) (Work done at Rensselaer Polytechnic Institute) 1

Motivation Infrastructure / Wireless Mesh Networks • Characteristics: Fixed, unlimited energy, virtually unlimited processing power • Dynamism – Link Quality • Optimize – High throughput, low latency, balanced load Scalability Layer 3: Network Layer Mobile Adhoc Networks (MANET) • Characteristics: Mobile, limited energy • Dynamism – Node mobility + Link Quality • Optimize – Reachability Sensor Networks Main Issue: Scalability Introduction MORRP Key Concepts • Characteristics: Data-Centric, extreme limited energy • Dynamism – Node State/Status (on/off) • Optimize – Power consumption Simulation Results Conclusion 2

Scaling Networks: Trends in Layer 3 Flood-based Mobile Ad hoc / Fixed Wireless Networks DSR, AODV, TORA, DSDV Partial Flood: OLSR, HSLS Peer to Peer / Gnutella Overlay Networks Wired Networks Introduction OSPF, IEGRP, RIP Hierarchy/Structured Unstructured/Flat Scalable LGF, VRR, GPSR+GLS Hierarchical Routing, WSR (Mobicom 07) ORRP (ICNP 06) Kazaa, DHT Approaches: CHORD, CAN Bubble. Storm (Sigcomm 07) LMS (PODC 05) OSPF Areas MORRP Key Concepts Simulation Results Conclusion 3

Trends: Directional Communications Directional/Directive Antennas B’ B’ B A A’ Hybrid FSO / RF MANETS B D’ A D C C’ Omni-directional A’ D’ D C C’ Directional • Directional Antennas – Capacity Benefits § Theoretical Capacity Improvements - factor of 4 p 2/sqrt(ab) where a and b are the spreads of the sending and receiving transceiver ~ 50 x capacity with 8 Interfaces (Yi et al. , 2005) § Sector Antennas in Cell Base Stations – Even only 3 sectors increases capacity by 1. 714 (Rappaport, 2006) Introduction MORRP Key Concepts • Current RF-based Ad Hoc Networks: § omni-directional RF antennas § High-power – typically the most power consuming parts of laptops § Low bandwidth § Error-prone, high losses § Free Space Optics: § High bandwidth § Low Power § Dense Spatial Reuse § License-free band of operation Simulation Results Conclusion 4

ORRP Big Picture Orthogonal Rendezvous Routing Protocol ORRP Primitive 1: Local sense of direction leads to ability to forward packets in opposite directions ORRP 180 o S T 2: Forwarding along Orthogonal lines has a high chance of intersection in area Introduction A 98% • High reach (98%), O(N 3/2) State complexity, Low path stretch (~1. 2), Up to 69% high goodput, unstructured • BUT. . What happens with mobility? Wireless Mesh Networks Mobile Ad-Hoc Networks B Increasing Mobility 65% 55% 42% Overlay Networks 5

Mobile-ORRP (MORRP) Introduction What can we do? A a R B Introduction MORRP Key Concepts • Replace intersection point with intersection region. • Shift directions of send based on local movement information • Route packets probabilistically rather than based on rigid nexthop paths. (No need for route maintenance!) • Solution: a NEW kind of routing table: Directional Routing Table (DRT) Simulation Results Conclusion 6

MORRP Basic Example C R: Near Field DRT Region of Influence B S A R F Original Path S O Q N R P S: Near Field DRT Region of Influence G R’ E Original Direction (a 1) Original Path New Direction (a 2) M D L D’ I H K D: Near Field DRT Region of Influence J 1. Proactive Element – Generates Rendezvous to Dest Paths 2. Reactive Element – Generates Source to Rendezvous Paths Introduction MORRP Key Concepts Simulation Results Conclusion 7

The Directional Routing Table Use Decaying Bloom Filter (DBF) Routing Tables viewed from Node A Routing Table RT w/ Beam ID Directional RT (DRT) Dest ID Z 3 D A 2 1 B C Dest ID Next Hop Beam ID Dest IDs (% of Certainty) Beam ID B C D : Z B B Z : Z 1 1 3 : 3 B(90%), C(30%). Z(90%), D(40%). 1 2 3 4 ID 4 Next Hop ID ID set of IDs Set of IDs set of IDs • Soft State – Traditional routing tables have a hard timeout for routing entries. Soft State decreases the level of certainty with time. • Uncertainty with Distance – Nodes closer to a source will have increasingly more information about the location of the source than nodes farther away • Uncertainty with Time – As time goes on, without updates, one will have lesser amount of information about the location of a node • Uncertainty with Mobility – Neighbors can potentially be “covered” by different interfaces based on mobility speed and direction Introduction MORRP Key Concepts Simulation Results Conclusion 8

DRT Intra-node Decay Time Decay with Mobility Spread Decay with Mobility a q 2 7 x x q 3 q 1 8 As node moves in direction +x, the certainty of being able to reach nodes covered by region 8 should decay faster than of region 7 depending on speed. This information is DROPPED. Introduction q 2 > q 1 > q 3 MORRP Key Concepts a As node moves in direction +x, the certainty of being able to reach nodes covered by region 2 should be SPREAD to region 1 and 3 faster than the opposite direction. The information about a node in region 2 should be SPREAD to regions 1 and 3. Simulation Results Conclusion 9

MORRP Fields of Operation N N N N S R N N N • Near Field Operation § Uses “Near Field DRT” to match for nodes 2 -3 hops away • Far Field Operation § RREQ/RREP much like ORRP except nodes along path store info in “Far. Field DRT” Introduction MORRP Key Concepts Simulation Results N N D N N Conclusion 10

Performance Evaluation of MORRP • Metrics Evaluated § Reachability – Percentage of nodes reachable by each node in network (Hypothesis: high reachability) § Delivery Success – Percentage of packets successfully delivered network-wide § Scalability – The total state control packets flooding the network (Hypothesis: higher than ORRP but lower than current protocols out there) § Average Path Length § End to End Delay (Latency) § Aggregate Network Goodput • Scenarios Evaluated (NS 2) § Evaluation of metrics vs. AODV (reactive), OLSR (proactive), GPSR with GLS (position-based), and ORRP under various node velocities, densities, topology-sizes, transmission rates. § Evaluation of metrics vs. AODV and OLSR modified to support beamswitched directional antennas. Introduction MORRP Key Concepts Simulation Results Conclusion 11

MORRP: Aggregate Goodput Results • Aggregate Network Goodput vs. Traditional Routing Protocols § MORRP achieves from 10 -14 X the goodput of AODV, OLSR, and GPSR w/ GLS with an omni-directional antenna § Gains come from the move toward directional antennas (more efficient medium usage) • Aggregate Network Goodput vs. AODV and OLSR modified with directional antennas § MORRP achieves about 15 -20% increase in goodput vs. OLSR with multiple directional antennas § Gains come from using directionality more efficiently Introduction MORRP Key Concepts Simulation Results Conclusion 12

MORRP: Simulations Summary • MORRP achieves high reachability (93% in mid-sized, 1300 x 1300 m 2 and 87% in large-sized, 2000 x 2000 m 2 topologies) with high mobility (30 m/s). • With sparser and larger networks, MORRP performs fairly poorly (83% reach) suggesting additional research into proper DRT tuning is required. • In lightly loaded networks, MORRP end-to-end latency is double of OLSR and about 7 x smaller than AODV and 40 x less than GPSR w/ GLS • MORRP scales well by minimizing control packets sent • MORRP yields over 10 -14 X the aggregate network throughput compared to traditional routing protocols with one omnidirectional interface gains from using directional interfaces • MORRP yields over 15 -20% the aggregate network goodput compared to traditional routing protocols modified with 8 directional interfaces gains from using directionality constructively Introduction MORRP Key Concepts Simulation Results Conclusion 13

MORRP: Key Contributions • The Directional Routing Table § A replacement for traditional routing tables that routes based on probabilistic hints § Gives a basic building block for using directionality to overcome issues with high mobility in MANET and DTNs • Using directionality in layer 3 to solve the issues caused by high mobility in MANETs • MORRP achieves high reachability (87% - 93%) in high mobility (30 m/s) • MORRP scales well by minimizing control packets sent • MORRP shows that high reach can be achieved in probabilistic routing without the need to frequently disseminate node position information. • MORRP yields high aggregate network goodput with the gains coming not only from utilizing directional antennas, but utilizing the concept of directionality itself. • MORRP is scalable and routes successfully with more relaxed requirements (No need for coordinate space embedding) Introduction MORRP Key Concepts Simulation Results Conclusion 14

Thank You! • Questions and Comments? • Papers / Posters / Slides / NS 2 Code (MORRP, OLSR + AODV with Beam switched directional antennas) [ http: //networks. ecse. rpi. edu/~bownan ] • bownan@gmail. com Introduction MORRP Key Concepts Simulation Results Conclusion 15

EXTRA SLIDES 16

The Directional Routing Tables viewed from Node A Routing Table RT w/ Beam ID Directional RT (DRT) Dest ID Z 3 A 2 D 1 B C Dest ID Next Hop Beam ID Dest IDs (% of Certainty) Beam ID B C D : Z B B Z : Z 1 1 3 : 3 B(90%), C(30%). Z(90%), D(40%). 1 2 3 4 ID 4 Next Hop ID ID set of IDs Set of IDs set of IDs • Destination ID % of Certainties for each Beam ID stored within a Decaying Bloom Filter • Bloom Filter – A space-efficient probabilistic data structure that is used to test whether an element is a member of a set. § Consist of a bit array and a set of k linearly independent hash functions § Storage: IDs are hashed to each of the k hash functions stores a ``1’’ in position in the bit array for each hash function. § Search: IDs are hashed through each of the k hash functions if all positions have a “ 1”, then the ID is in the set. Otherwise, the ID is not in the set Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 17

DRT: Decaying Bloom Filter Primer ID: 1 4 Hash h (x) = (x 2 + 20) % 32 1 Funcs: h 11 h 1(2)24 24 h (6) = 21 (1) = = ID: 2 h 2(x) = x % 32 h 22 h 2(2)6= 2 h (6) = 1 (1) = ID: 6 h 3(x) = (x + 5) % 32 h 33 h 3(2)11 7 h (6) = 6 (1) = = h 4(x) = (x 3 + 25) % 32 h 44 h 4(2)17 1 h (6) = 26 (1) = = 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 Bit Array: 0 0 0 0 0 0 0 0 1 1 1 1 What policies Traditional Bloom Filter For decaying Search ID 1 – 4 of 4 bits match (IN set) bits can we Search ID 6 – 2 of 4 bit match (Not in set) employ? Decaying Bloom Filter (DBF) Search ID 1 – 4 of 4 bits match (100% chance in set) Search ID 6 – 2 of 4 bit match (50% chance in set) Introduction Wireless Mesh Networks DRT Dest Prob. Beam ID (DBF) 7 8 6 5 A 4 3 Mobile Ad-Hoc Networks 1 2 0010. . 1000 0000. . 1001 0011. . 0101. . 1001 0010. . 0000. . 0001 0011. . 1011 0111. . 1001 Overlay Networks 1 2 3 4 5 6 7 8 18

DRT Inter-Node Decay DRT at Node A Strong Info B Med Info C Low Info D Noise 0 0 1 0 0 0 … 1 0 0 BEAM ID: 1 0 0 0 0 0 … 0 0 1 0 0 BEAM ID: 2 0 1 0 0 … 0 0 1 BEAM ID: 3 0 0 1 0 0 0 1 … 0 0 1 0 0 BEAM ID: 4 … 1 0 1 Merged DBF (Update DBF) … 1 0 0 0 1 Decayed DBF (50% bits dropped) B is now 100% sure A is 1 hop away while only 50% sure C can be reached through sending out interface 1 4 3 B 1 A C 2 Bitwise-OR 0 1 1 1 0 1 Decay 50% of Bits 0 0 1 0 0 My ID (A) h 1(x), h 2(x), …, hn(x) 0 0 1 1 0 1 0 0 Introduction … 1 0 1 Wireless Mesh Networks Broadcasted by A to all Neighbors Mobile Ad-Hoc Networks Overlay Networks 19

DRT Intra-node Decay Time Decay with Mobility Spread Decay with Mobility a q 2 7 x x q 2 > q 1 > q 3 q 1 8 a As node moves in direction +x, bits in DBF of region 8 should decay faster than of region 7 depending on speed As node moves in direction +x, bits in DBF of region 2 should be SPREAD to region 1 and 3 faster than the opposite direction Beam ID 1 0 0 0 0 0 0 0 0 1 1 Beam ID 2 0 1 1 0 0 0 1 1 1 0 0 0 0 1 1 0 0 0 0 0 Beam ID 3 0 0 0 0 0 0 0 0 1 1 Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 20

Conclusion / Future Work • Used Directionality to scale wireless networks (ORRP, MORRP) • Used concept of Virtual Directions to scale overlay networks (VDR) • Future Work: Extensions § Virtual direction abstraction analysis § Hybrid ORRP (that works with omnidirectional and directional antennas) § Analysis of Effect of knobs in MORRP • New Directions with Directionality § Multi-path / multi-interface diversity § Directional Network Coding § Destination-based routing based on local directions Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 21

Scaling Networks: OSI Model Transport Layer – Handles reliable transmissions end-to-end Network Layer – Manages routing from end-to-end Layers 5 -7 C F 4: Transport Layer A 3: Network Layer A 2: Link Layer 1: Physical Layer B Z E G Z H Link Layer – Manages node-to-node transmissions 1011010 Physical Layer – Handles transmission of bits through a medium Application/Presentation/Session Layers – Deal with the actual programs/data Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 22

Research Objectives • Wireless Mesh Context § Can directionality be used to address issues with scalability at higher throughput in layer 3 routing? • Mobile Ad Hoc Context § Can directionality be used to address issues with high mobility and throughput in layer 3 routing? • Overlay Network Context § Can directionality be used to scale flat, unstructured networks? Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 23

Orthogonal Rendezvous Routing Protocol ? N (4, 6) S D(X, Y)? (0, 4) W D (8, 5) (15, 5) (12, 3) (5, 1) E S By removing position information, can we still efficiently route packets? Issues in Position-based Schemes L 3: Geographic Routing using Node IDs (eg. GPSR, TBF etc. ) ORRP L 2: ID to Location Mapping (eg. GHT, GLS etc. ) N/A L 1: Node Localization Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 24

ORRP Design Considerations • Considerations: § High probability of connectivity without position information [Reachability] § Scalability O(N 3/2) total state information maintained. (O(N 1/2) per node state information) § Even distribution of state information leading to no single point of failure [State Complexity] § Handles voids and sparse networks • Trade-offs § Path Stretch § Probabilistic Reachability Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 25

ORRP Proactive and Reactive Elements Node B Fwd Table Node C Fwd Table Dest Next Hops A North A A to D A 1 Dest Next Hops 1. 2. 3. 4. B 2 Dir 230 o D A F 3 120 o D C Node F Fwd Table A North 120 o F B North Dest Next Hops North D 1 230 o North ORRP Announcements (Proactive) – Generates Rendezvous-to-Destination Routes ORRP Route Request (RREQ) Packets (Reactive) – Generates Source-to-Rendezvous Rts ORRP Route Reply (RREP) Packets (Reactive) Data path after route generation Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 26

Reachability Numerical Analysis 2 1 Probability of Unreach highest at perimeters and corners P{unreachable} = P{intersections not in rectangle} 3 NS 2 Simulations with MAM show 4 Possible Intersection Points around 92% reachability 57% 98. 3% 99. 75% 67. 7% Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 27

Path Stretch Analysis Average Stretch for various topologies x = 1. 255 x = 1. 15 • • Square Topology – 1. 255 Circular Topology – 1. 15 25 X 4 Rectangular – 3. 24 Expected Stretch – 1. 125 x = 3. 24 Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 28

State Complexity Analysis/Simulations GPSR DSDV XYLS ORRP Node State O(1) O(n 2) O(n 3/2) Reachability High 100% High (99%) Name Resolution O(n log n) O(1) Invariants Geography None Global Comp. Local Comp. ORRP state scales with Order N 3/2 Introduction Wireless Mesh Networks ORRP states are distributed fairly evenly in an unstructured manner (no single point of failure) Mobile Ad-Hoc Networks Overlay Networks 29

ORRP: Simulation Results Summary • Base Case § Reach – 99% for Square topologies, 92% for Rectangular topologies (MAM helped) § Path Stretch – Roughly 1. 2 § Goodput – About 30 x more aggregate network goodput than AODV, 10 x more aggregate network goodput than OLSR and 35 x more aggregate network goodput than GPSR with GLS (due to better usage of medium) • Network Voids § Average path length fairly constant (Reach and State not different) • Additional Lines § Reach/Path Stretch – All showed large gains from 1 to 2 lines but diminishing returns thereafter § Goodput – Higher average network throughput with additional lines (better paths and higher reach) but not by much • Varying Number of Interfaces § Significant increase in reachability from 4 to 8 interfaces, but gains trail off Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 30

ORRP: Summary • ORRP achieves high reachability in random topologies • ORRP achieves O(N 3/2) state maintenance – scalable even with flat, unstructured routing • ORRP achieves low path stretch (Tradeoff for connectivity under relaxed information is very small!) • ORRP achieves roughly 30 X in aggregate network goodput compared to AODV, 10 X the aggregate network goodput compared to OLSR, and 35 X the aggregate network goodput compared to GPSR with GLS. Relevant Papers • • • B. Cheng, M. Yuksel, and S. Kalyanaraman, Rendezvous-based Directional Routing: A Performance Analysis, In Proceedings of IEEE International Conference on Broadband Communications, Networks, and Systems (BROADNETS), Raleigh, NC, September 2007. (invited paper) B. Cheng, M. Yuksel, and S. Kalyanaraman, Directional Routing for Wireless Mesh Networks: A Performance Evaluation, Proceedings of IEEE Workshop on Local and Metropolitan Area Networks (LANMAN), Princeton, NJ, June 2007. B. Cheng, M. Yuksel, and S. Kalyanaraman, Orthogonal Rendezvous Routing Protocol for Wireless Mesh Networks, Proceedings of IEEE International Conference on Network Protocols (ICNP), pages 106 -115, Santa Barbara, Nov 2006. Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 31

Wireless Nets: Key Concepts to Abstract • Directionality CAN be used to provide high reach, high goodput, low latency routing in wireless mesh (ORRP) and highly mobile adhoc networks (MORRP) • Primitives: § Local directionality is enough to maintain forwarding along a straight line § Two sets of orthogonal lines intersect with a high probability in a bounded region • Overlay Networks: § Can we take these concepts to scale unstructured, flat, overlay networks? Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 32

Virtual Direction Routing Introduction • Structured vs. Unstructured Overlay Networks § 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 • High Reach • Low path stretch • Not scalable § Random Walk Virtual Direction Routing • 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 Random Walk • Focus (for now) § Small world approximations Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 33

VDR: Neighbor to Virtual Interface Map Example: Neighbor IDs used Instead Of SHA-1 Hashes 30 % 8 = 6 15 15 % 8 = 7 30 10 % 8 = 2 8 Virtual Interfaces 26 10 2 10 1 3 68 26 26 % 8 = 2 68 1 0 1 4 5 7 6 15 30 68 % 8 = 4 • Neighbors are either physical neighbors connected by interfaces or neighbors under a certain RTT latency away (logical neighbors) • Neighbor to Virtual Interface Mapping § 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) Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 34

VDR: State Seeding and Route Request State Seeding – |10 – 1| = 9 |26 – 1| = 25 10 26 State info forwarded in orthogonal 2 1 directions, biasing 3 0 packets toward IDs 1 that are closer to 4 7 68 SOURCE ID. Packets 5 6 are forwarded in virtual straight lines. 30 Ex: Seed Source: Node 1 Route Request – RREQ packets are forwarded in orthogonal directions, biasing packets towards REQUESTED ID 26 3 68 1 0 1 4 5 2 67 48 3 28 15 10 4 5 |5 – 1| = 4 5 |13 – 1| = 12 6 28 15 30 |5 – 12| = 7 5 |13 – 12| = 1 Wireless Mesh Networks 0 3 7 4 6 2 5 13 0 5 7 6 |14 – 1| = 13 |22 – 1| = 21 55 14 22 10 1 10 4 1 5 1 3 7 2 13 67 48 10 1 |10 – 12| = 2 |26 – 12| = 15 2 Ex: Route Request: Node 12 RREQ Source: Node 1 Introduction 10 1 2 0 3 7 4 6 1 13 5 |6 – 12| = 6 |38 – 12| = 26 Mobile Ad-Hoc Networks 0 7 6 38 6 Overlay Networks 35

VDR: Simulation Parameters RREQ Path 30 1 5 10 13 26 48 Flooding 46 68 RREP Path 6 Rendezvous Node 38 2 RREQ: Node 12 Seed Path VDR Route Request Virtual View 12 § VDR-R: VDR with random neighbor forwarding (no biasing) § RWR: Data is seeded in 4 random walks and 4 walkers are sent for search • Peer. Sim – 50, 000 Nodes, Static + Dynamic Network § Reach Probability – High (98% w/ TTL of 100) § Average Path Stretch – High (16) § State and Load Spread – Not evenly distributed Wireless Mesh Networks Virtual Direction Routing VDR – Random NB Send (VDR-R) • Simulation of VDR vs. RWR, VDR-R Introduction Normalized Flooding Random Walk Routing (RWR) Random Walk Mobile Ad-Hoc Networks Overlay Networks 36

VDR: Robustness Results • State Distribution Network-wide § Average States maintained relatively equal for VDR, VDR-R and RWR at 350 -390 § VDR States are not very evenly distributed, with some nodes having more state than others. This is due to the sending bias • Robustness to Network Churn § VDR drops only 5% compared to VDR-R and RWR which drop 1215% reach when going from 0% to 50% network churn § Even with a TTL of 50, VDR reaches a good amount of the network Introduction Wireless Mesh Networks 5% drop 12% drop Mobile Ad-Hoc Networks Overlay Networks 37

VDR: Key Contributions • Introduction of the concept of Virtual Directions to eliminate need for structure (coordinate space, DHT structures) to scale flat, unstructured overlay networks • A flat, highly scalable, and resilient to churn routing algorithm for overlay networks • VDR provides high reach (98% even only for a TTL of 100 in a 50, 000 node network) • VDR drops only 2 -5% going from 0% churn to 50% churn Relevant Papers • B. Cheng, M. Yuksel, and S. Kalyanaraman, Virtual Direction Routing for Overlay Networks, In preparation for submission to IEEE Peer to Peer Computing (P 2 P) 2008. Introduction Wireless Mesh Networks Mobile Ad-Hoc Networks Overlay Networks 38