End-to-end path: route

Transcription

1 Multi-hop Wireless Netorks CMU CS : Internet-Scale Sensor Sstems Overvie: Large-Scale Wireless Sstems Small-Scale: Ho to build single-hop ireless LAN; ho to make TCP perform ell over it Large-Scale: Ho to build multi-hop ireless sstems (MANs? WANs?); ho to support mobile nodes and users at Internet-scale Brad Karp 18th March, 23 Multi-hop ( ad hoc ) ireless routing: Ho do e find routes hen the topolog is highl dnamic, and hen the netork diameter is great? Multi-hop ireless capacit: Ho much user traffic can e carr on a large-scale, multi-hop ireless netork? Where might multi-hop ireless netorks fit ith IrisNet? CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 1 Multi-Hop Motivation: Rooftop Netorks Metropolitan-area netork comprised of customer-oned and -operated radios: Rooftop Netorks The Routing Problem R Packet-sitched netorks S An alternative architecture to single-hop cellular sstems: Self-organizing, rapidl deploable, potentiall loer cost Great demand alread! Hardare ubiquitous; scalable algorithms for routing sorel needed End-to-end path: route Each router chooses neighbor to hich to forard received packet onard toard destination, Topolog ma be dnamic: routes change CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 2 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 3 Another Motivating Eample Vast ireless netork of mobile temperature sensors, floating on the ocean s surface: Sensor Netorks Motivation (cont d) Enable three ne classes of netorks: Ad-hoc netorks: mobile, infrastructureless, small-scale [Broch et al., 98] CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 4 Sensor netorks: mobile, large-scale Rooftop netorks: fied, large-scale, no common administrative authorit [Shepard, 96] A mi of these characteristics: Mobilit Scale (number of nodes) Lack of static hierarchical structure CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 5

2 Scalabilit Goals for Mobile, Wireless Routing Prior Work Wired, Intra-omain Internet Routing: As number of nodes increases, and mobilit rate increases: Routing protocol message cost: MINIMIZE Application packet deliver success rate: MAXIMIZE Route length: MINIMIZE Per-node state: MINIMIZE Link-State (ijkstra) and istance-vector (Bellman-Ford) routing on flat addresses to find shortest (in hops) paths escribe entire topolog to all routers (LS) or push distances across netork diameter (V), for O(N) state per router Each link change must be communicated to all routers to avoid loops and disconnection [Zaumen, Garcia-Luna Aceves, 91] Ad Hoc Routing: Algorithms target lo-bandidth, high-mobilit netorks Man proposals (SV,, TORA, AOV,, ZRP,... ) iverse approaches: V, source routing, geographic, proactive, on-demand... CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 6 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 7 Ad Hoc Routing: SV estination-sequenced istance-vector Routing: Send increasing sequence number ith route advertisements Greater seqno takes precedence over lesser metric On detecting disconnection to, router advertises route ith infinit metric and incremented seqno increments seqno on hearing advertisement ith infinit metric Use triggered updates to propagate seqno increases rapidl and eliminate potentiall looping routes namic Source Routing: Ad Hoc Routing: On-demand routing: onl generate routing protocol traffic hen forarding requires it Flood queries to learn source routes Cache replies Source routes break more frequentl as mobilit and netork diameter increase; caching steadil less effective Broch et al. evaluation of : 5-node netorks Metrics: routing protocol overhead, path optimalit, packet deliver success rate Eploration of limits of? CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 8 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 9 Prior Work: Scaling ominant factors in scaling of V, LS, algorithms: Rate of change of topolog Geograph Central Idea: Machines can kno their geographic locations. Route using GEOGRAPHY. Number of routers in the routing domain Scaling strategies: Hierarch: at AS boundaries (BGP) or on a finer scale (OSPF) Goal: Reduce number of nodes in a routing domain Assumptions: Level boundaries relativel fied; administrative authorit can choose level boundaries Caching: Store source routes overheard () Goal: Limit propagation of future source route queries Assumption: Source route remains fied hile cached Assumptions invalid for highl mobile or unstructured netorks! CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 1 Established positioning methods: GPS outdoors (single chip, lo-cost) Surveing (stationar routers) Inertial sensors (vehicles) Acoustic and radio range-finding (indoors, [AT&T Cambridge, 1997], [Priantha et al., 2]) Efficient node location lookup/registration sstem [Li et al., 2] All nodes kno on position; packet source marks packet ith destination s location CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 11

3 Assumptions Bi-directional radio links (e.g., IEEE ith link-level acknoledgements) Netork nodes placed roughl in a plane Radio propagation in free space; distance from transmitter determines signal strength at receiver (to-ra ground reflection model) Greed Forarding Nodes learn immediate neighbors positions through beacons/piggbacking on data packets Locall optimal, greed forarding choice at a node: Forard to the neighbor geographicall closest to the destination Fied, uniform radio transmitter poer CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 12 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 13 In Praise of Geograph Greed Forarding Failure Self-describing Greed forarding not alas possible! Consider: As node densit increases, shortest paths through ireless netorks correspond increasingl to Euclidean straight line beteen source and destination v z Each node s state concerns onl immediate neighbors: Tin per-node state Routing protocol pushes state onl one hop tin routing protocol overhead Local forarding decisions robust to topolog changes Compare ith lookup in O(N) table under V, LS CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 14 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 15 Voids Node ensit and Voids When the intersection of a node s circular radio range and the circle about the destination on hich the node sits is empt of nodes, greed forarding is impossible Such a region is a void: v z void CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks Number of nodes Fraction of paths Eisting and Found Paths, 134 m 134 m Region Fraction eisting paths Fraction paths found b greed The probabilit that a void region is empt of nodes increases as nodes become more sparse CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 17

4 Void Traversal: The Right-Hand Rule Well-knon graph traversal: right-hand rule: When arriving at from, move to the net verte counterclockise about from 2. z Planar vs. Non-Planar Graphs The right-hand rule ma not tour enclosed faces on graphs ith edges that cross (non-planar graphs) u z v 4. Traverses interior faces in clockise edge order; eterior faces in counterclockise edge order Seek a distributed algorithm that removes crossing edges ithout partitioning the netork, using onl neighbors positions as input CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 18 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 19 Planarized Graphs Relative Neighborhood Graph (RNG) [Toussaint, 8] and Gabriel Graph (GG) [Gabriel, 69] are long-knon planar graphs Planarized Graphs: Eample 2 nodes, placed uniforml at random on a 2-b-2-meter region; radio range 25 meters Assume an edge eists beteen an pair of nodes separated b less than a threshold distance (i.e., the nominal radio range) RNG and GG can be constructed using onl neighbors positions, and can be shon not to partition the netork! u v u v Full Netork GG Subset RNG Subset RNG GG CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 2 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 21 Full Greed Perimeter Stateless Routing All packets begin in greed mode Upon greed failure, node marks its current location in packet, and marks packet in perimeter mode Perimeter Mode Forarding Eample Perimeter mode packets follo simple planar graph traversal: Forard along successivel closer faces b right-hand rule, until reaching destination, or node closer to it than perimeter mode entr point Return packets to greed mode hen the reach a node closer than their perimeter mode entr point CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 22 Traverse face closer to along b right-hand rule, until reaching the edge that crosses Repeat ith the net closer face along, &c. Record first edge on face to detect disconnection CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 23

5 Pause Time : Protocol Techniques for namic Netorks Use of MAC-laer failure feedback: As in [Broch, Johnson, 98], interpret retransmit failure reports from the MAC as indication a neighbor has gone out-of-range Interface queue traversal and packet purging: Upon MAC retransmit failure for a neighbor, alk the interface queue and remove packets to that neighbor to avoid head-of-line blocking of transmitter during retries on those packets Promiscuous netork interface: Reduce beacon load and keep positions stored in neighbor tables current b tagging all packets ith the forarding node s position Planarization triggers: Re-planarize upon acquisition of a ne neighbor and ever loss of a former neighbor, to keep planarization up-to-date as topolog changes Simulation Environment ns-2 ith ireless etensions [Broch et al., 1998]: full MAC, phsical propagation; allos comparison of results Topologies and Workloads: Nodes Region ensit CBR Flos 5 15 m 3 m 1 node / 9 m m 6 m 1 node / 9 m m 134 m 1 node / m 2 3 Simulation Parameters: Pause Time:, 3, 6, 12 s Motion Rate: [1, 2] m/s Beacon Interval: 1.5 s ata Packet Size: 64 btes CBR Flo Rate: 2 Kbps Simulation Length: 9 s CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 24 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 25 Packet eliver Success Rate (5, 2; ense) (2 nodes) (2 nodes), B = 1.5 (5 nodes) (5 nodes), B = CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 26 Fraction data pkts delivered Routing Protocol Overhead (5, 2; ense) 1e+7 1e (2 nodes) (2 nodes), B = 1.5 (5 nodes) -Broch (5 nodes) (5 nodes), B = CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 27 Path Length (2; ense) Fraction of delivered pkts >= 5 Hops Longer than Optimal Routing Algorithm State Size (2; ense) Routing algorithm Routing algorithm requires state proportional to node densit; stores state at each router proportional to the sums of the lengths of source routes CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 28 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 29 Per-node state (nodes) Per-node state (btes) Routing protocol overhead (pkts)

6 Packet eliver Success Rate (5; Sparse) (5 nodes) (5 nodes), B = 1.5 (5 nodes), B = 1.5, Greed Onl CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 3 Fraction data pkts delivered Routing Protocol Overhead (5; Sparse) Routing protocol overhead (pkts) (5 nodes) (5 nodes), B = 1.5 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 31 Path Length (5 nodes, Sparse) Fraction of delivered pkts >= 5 Hops Longer than Optimal CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 32 Pause Time Routing Algorithm Capacit of Ad Hoc Wireless Netorks Contet: Mobicom 21, on the heels of ears of ad hoc routing research, nearl eclusivel in simulation One commercial real sstem: Metricom (R.I.P., 2) Goals: Eplain details of MAC hen used for forarding, as regards netork capacit Provide simple model for capacit of ad hoc netorks, as related to traffic matri Fundamental phenomenon: nodes use their on one-hop transmission capacit not onl for data the originate, but also for data the forard CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 33 Some depressing intuition: Ad Hoc Capacit: Intuition Spatial reuse lets distant radios transmit simultaneousl, as the don t interfere Forarding and The energ required to garble another s transmission is far less than that required to be received properl Interfering range is 55 m, hile transmission range is 25 m For constant node densit, one-hop capacit, sum of all single-hop transfer rates possible in the netork, gros as O(n) As netork diameter gros, for random source/destination pairs, average path length gros as O( n) Total end-to-end capacit: O(n/ n), and so per-node capacit is O(1/ n). Throughput per node approaches zero as number of nodes increases! CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 34 What s the best throughput e can epect in a chain A B C, if ranges ere equal? A and B can t transmit simultaneousl; nor can B and C; nor can A and C Best throughput: 1/3 link rate With 55 m interference range, best throughput drops to 1/4; no interferes ith A s transmissions to B In simulations of greed senders arranged in a chain, throughput is closer to 1/7 than 1/4; boundar effect CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 35

7 Forarding and Backoff Traffic Matri and Multi-Hop Wireless Capacit A B C E F Capacit available to each node inversel related to epected flo phsical path length Traffic matri tpicall studied in ad hoc routing: uniforml randoml selected flo endpoints Epected path length for a uniform random traffic pattern on a netork of area A: 2 A/3 For n nodes and fied node densit, A n E ill clobber A B Yet A doesn t kno of s transmission Result: repeated eponential backoff b A So the capacit available to each node is O(1/ n) Perhaps this is h published ad hoc routing studies use ca. 6 Kbps total application traffic orkloads! CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 36 CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 37 Poer-La Traffic Patterns and Capacit Poer-la traffic patterns, here probabilit of communication ith node distance aa is given b α, offer constant per-node capacit For α = 2, epected communication distance scales as O(log2 A) A useful design rule for sstems for multi-hop ireless netorks, e.g., GLS location database [Li et al., ] Poer-la construct makes analsis tractable; meaning is intuitivel useful Evidence of poer-la communication patterns in the ild? CMU CS : Internet-Scale Sensor Sstems Brad Karp Lecture 18 Multi-hop Wireless Netorks 38