Integrated Routing & MAC Scheduling for Wireless Mesh Networks. Zhibin Wu Dipankar Raychaudhuri WINLAB, ECE Dept. Rutgers University

Transcription

1 Integrated Routing & MAC Scheduling for Wireless Mesh Networks Zhibin Wu Dipankar Raychaudhuri WINLAB, ECE Dept. Rutgers University

2 Outline Our Perspective of High-Speed Wireless Mesh Networks Problem & Approaches Analytic Throughput Bounds for Integrated Routing/scheduling Protocol and Algorithms: Practical Designs Simulation Results with IRMA protocols Conclusions & Future Work 2 IRMA fro Wireless Mesh Networks December 2007

3 Emerging Broadband Radios and Throughput-Thirsty Devices Short-range, high-speed radio technologies emerging: Technology Name Link Speed (in Mbps) BW (MHz) Frequency Band n or40 2.4GHz and 5GHz WiMedia to 10.6 GHz mm-wave (WPAN) up to 2000 N/A 60GHz Device need more throughput over the air Mobiles More processing power, storage capacity and less cost Contents comes from network and share on the network Wireless Camcorders Wireless HDTV 3 IRMA fro Wireless Mesh Networks December 2007

4 New Forms of Wireless Networks CDMA2000 BTS RNC PDSN Wireless Mesh Network WiMAX BS ASN-GW Internet Internet Wi-Fi Wireless not only an access technology, but also a networking technology Wireless devices form a true wireless network Multi-hop wireless peer-to-peer transport 4 IRMA fro Wireless Mesh Networks December 2007

5 High-Speed Wireless Mesh Network Requirement: High Throughput Bulk data transfer over wireless mesh architecture. Goal: Protocol and algorithm design to achieve maximal throughput for concurrent end-to-end multihop flows. Problems: What s the achievable end-to-end capacity given a certain small number of nodes in the topology? Is there a feasible design to approximate the capacity? What s the control overhead and how it scales with the traffic demands? Conventional Approach: IEEE ad hoc routing 5 IRMA fro Wireless Mesh Networks December 2007

6 CSMA/CA MAC Problems with multi-hop radio network A simple contention-based protocol fails to handle interference Link level: Hidden terminals and exposed terminals. o Spatial reuse vs. Collision avoidance. o Mesh network simulation shows ~68% of RTS get no response Flow level: Intra and Inter-flow interference Hidden Node Self Interference Inter-flow Interference Exposed Node 6 IRMA fro Wireless Mesh Networks December 2007

7 Wireless MAC Design Techniques RTS/CTS, MACA- P, D-LSMA DCMA IRMA Reserve first, access later is a must to reach interference-free as well as throughput-optimal Flow-based scheduling design are still offline optimizations, lack of protocols End-to-End flow scheduling can be optimized jointly with L3 7 IRMA fro Wireless Mesh Networks December 2007

8 Joint Routing/Scheduling: Example (a) RTS RTS Data CTS RTS RTS RTS RTS ACK Data CTS RTS ACK MAC with designated Min-hop paths t a 4x performance gain can be achieved by carefully selecting (b) Data ACK Data Data Data ACK Data ACK Data Data ACK Data ACK ACK ACK Synchronized TDMA MAC with better paths t routing paths and MAC schedules JOINITLY. 8 IRMA fro Wireless Mesh Networks December 2007

9 Flow Based Management over TDM Slot Assignments Design both L2 & L3 protocols based on flow session, e.g. scheduling result won t change till the end of a traffic session, better end-to-end performance and less overhead Traffic flows over HS-WMN are not conventional Internet traffic Bulk data transfer & Aggregated traffic from multiple clients To serve long-lived CBR and VBR, TDMA is better. Traffic is not bursty, require deterministic bandwidth and delay By satisfying the interference constraint one can guarantee a certain fixed data rate at a given link. Global synchronization difficulty can be overcome in a small network Dedicated bandwidth can be specified for control Frame n Frame n+1 Frame n+2 9 IRMA fro Wireless Mesh Networks December 2007

10 Contributions of Our Research Offline Approach: Analytical framework to obtain optimal resource allocations and throughput bounds for end-to-end traffic over single-channel wireless mesh networks, especially for single path routing case. Online Approach: IRMA System Design Separation of control and data: Global Control Plane Centralized and distributed heuristic algorithms. 10 IRMA fro Wireless Mesh Networks December 2007

11 Obtaining throughput Bounds from LP Optimizations Maximize end-to-end throughput : multi-commodity network flow problem (linear programming) Analytical throughput bounds with protocol interference model for following scenarios: Min-hop routing + link Scheduling othe path for each <s, d> is known as the min-hop path. Joint routing/scheduling (route is uncertain) o Multipath Routing: traffic from s to d is split in multiple routes More optimal than single-path routing Linear programming, but NP-hard to get all constraints o Single path routing. Mixed integer programming problem Using maximal utilization rounding to avoid integer variables 11 IRMA fro Wireless Mesh Networks December 2007

12 Numerical Results Gaps between the throughput bounds obtained by LP optimizations and a system simulation shows huge potential for integrated protocol design [Gupta 00] :Given n nodes randomly located in a unit disk, transport capacity with interference-free protocols is Network capacity decreases with increasing number of flows 12 IRMA fro Wireless Mesh Networks December 2007

13 IRMA System Model Global control plane and data plane All control signaling on a separate plane Parameters of IRMA component in data plane is determined by control algorithms Control Algorithms Global Control Plane (GCP) Control Message IRMA Algorithms IRMA Control Agent Routing CSMA MAC PHY Control Plane Application IRMA Component Controlled Routing TDMA MAC PHY Data Plane Data Packet Protocol stacks in IRMA system 13 IRMA fro Wireless Mesh Networks December 2007

14 IRMA System Control Cycle Control Cycle: Detection and report of new or changed traffic demands. IRMA optimization determine the paths and conflict-free TDMA schedules for each node. IRMA components (Routing and MAC) transform and work with the new working parameters to ensure QoS. Bootstrap Topology Discovery Load default IRMA parameters Traffic Variation Detection IRMA Working Optimization Path/Schedule Adjust 14 IRMA fro Wireless Mesh Networks December 2007

15 Heuristic Algorithms for IRMA Centralized Algorithms Existing a central entity collect global information (topology, BW, interference relationship) and run optimization algorithms CIRMA-MH: solves min-hop routing, then optimizes link scheduling based on routing results and real-time flow demands. CIRMA-BR: attempts to optimize routing and scheduling decisions simultaneously, using available MAC bandwidth information to route around congested areas. Distributed Algorithms (DIRMA-MH and DIRMA-BR) each node needs to figure out the best schedule and route for its traffic in a distributed manner, with local information exchange only (in the control plane). Optimization for the whole flow is conducted hop by hop (source to sink) 15 IRMA fro Wireless Mesh Networks December 2007

16 CIRMA-MH MH (Min-hop) routing + TDMA link scheduling based on the path selection Inputs of the algorithm: Topology (G(V,E)) Traffic Profile (source-destination, bandwidth requirement) and Interference-index k TDMA frame length T (Number of slots in a frame) Output: route selection and MAC TDMA slot assignments CIRMA-MH Algorithm: 1. Find the shortest route with hop metric 2. For each link e in each flow F i, assign earliest available slot x for this link as long as it does not conflict with the links already scheduled in this slot x 3. Repeat step 2 until all bandwidth requirements are fulfilled or no more slots are available. 16 IRMA fro Wireless Mesh Networks December 2007

17 CIRMA-BR Min-hop routes are not optimal, cause local congestions Better paths can be found and yield higher throughput than MH paths Joint TDMA Link Scheduling and Bandwidth Aware Routing (BR) CIRMA-BR Algorithm: 1. Sorting the flow in ascending order by bandwidth requirements 2. For each flow F i, i= 1,2, M a) Generate link Metric based on available free TDMA Slots b) finding shortest path for flow F i with the bandwidth metric. and assign conflict-free TDMA slots for this flow A B D C (a) A B D C (b) Different routes used by (a) CIRMA-MH and (b) CIRMA-BR in a 6x6 grid for two vertical flows 17 IRMA fro Wireless Mesh Networks December 2007

18 DIRMA-MH Send SREQ, Receive SAPP Update TX-OK Free SREQ+SR EJ D SREQ SREJ Update Slot State Transitions C A B F SUPD Reserved Update Prohibited SREQ SAPP SCAN Receive SREQ, Send SAPP Updat e SREQ+SR EJ Update SUPD RX-OK SREQ+SR EJ E SREQ SAPP Distributed TDMA link schedule setup: Lock and reserve (too ideal) Schedule first, correct later SUPD messages Coordinate neighbors/interferers to mark the slot in Free/TxOK/RxOK/Prohibited States. Periodically broadcasted or triggered SREQ/SAPP/SREJ Reserve the slot to schedule a link transmission SCAN (Schedule Cancel) Release slot if no longer needed for this link Correct schedule errors if collision is detected Solve Hidden Node and Exposed Problems 18 IRMA fro Wireless Mesh Networks December 2007

19 DIRMA-BR Dynamic hop-by-hop routing/scheduling 1. Put forwarders in a candidate pool Interference-aware direction C 2. Candidates are sorted based on the combination of two metrics: A B Min-hop direction o M1:Distance to d o M2: Available BW for this hop and next D hop Interfered Area 3. Select the candidate with smallest metric as forwarder and reserve schedules 4. Choose the next candidate if fails 5. Fall back to MH path and reserve if all fails 19 IRMA fro Wireless Mesh Networks December 2007

20 Performance Evaluation Implement the GCP and IRMA algorithms in ns-2.28 A separate control radio and channel in GCP Ns-2 Simulation Parameters Compare the performance IRMA algorithms Baseline approaches o DSDV o AODV IRMA fro Wireless Mesh Networks December 2007

21 Simulation Results 40-node mesh topology Varying number of source-destination pairs Comparing IRMA algorithms with conventional layered approaches Results show that our IRMA schemes can provide up to ~200% gains. DIRMA-MH algorithm achieves results amount to 90% of CIRMA-MH Both BR algorithms behave better than MH algorithms in most of the cases. CIRMA-BR vs. CIRMA-MH 22.4% more, DIRMA-BR vs. DIRMA-MH 2.37% more The DIRMA-BR algorithm is not very good because per-hop selection is based upon volatile local information 21 IRMA fro Wireless Mesh Networks December 2007

22 Evaluation of the Control Overhead Overhead Statistics Baseline: RTS/CTS + routing overhead IRMA: All control signaling in GCP Simulation Topology 4x4 grid 10 traffic sessions with random start/end Traffic duration: exponential distributed. Results normalized by end-to-end throughput IRMA approach reduce signaling overhead because per-flow control is much more efficient than per-packet control 22 IRMA fro Wireless Mesh Networks December 2007

23 Centralized Protocols vs. Distributed Protocols CIRMA Route and scheduling need to be disseminated when each traffic session begins/ends Don't scale with increasing number of flows and size of network DIRMA Most overhead comes from system bootstrap, discovery and interference characterization Negligible overhead for scheduling and path selection 23 IRMA fro Wireless Mesh Networks December 2007

24 Conclusion and Future Work Fundamental need for cross-layer integration and joint optimization for superior network performance in WMN. Joint Routing and Scheduling show up to 300% performance gain in numerical results We proposed IRMA for wireless mesh networks and discussed: Flow-based Interference-free scheduling Realistic system model heuristic, promising online algorithms Simulation results show that IRMA design improve end-to-end throughput significantly with modest signaling overhead. How much gains when channel assignment is also incorporated Guarantee Fairness & QoS. 24 IRMA fro Wireless Mesh Networks December 2007

25 Please visit 25 IRMA fro Wireless Mesh Networks December 2007

26 Related Publications 1. Z. Wu, S. Ganu and D. Raychaudhuri, ''IRMA: Integrated routing and MAC scheduling in multihop wireless mesh networks'', in Proceedings of the Second IEEE Workshop on Wireless Mesh Networks, Reston VA, Sept Z. Wu, S. Ganu, I. Seskar and D. Raychaudhuri, ''Experimental investigation of PHY layer rate control and frequency selection in based ad-hoc networks'', in Proceedings of ACM SIGCOMM Workshops, August, S. Zhao, Z. Wu, A. Acharya and D. Raychaudhuri, ''PARMA: A PHY/MAC aware routing metric for ad-hoc wireless networks with multi-rate radios'', in Proceedings of IEEE WoWMoM'05, June Z. Wu, D. Raychaudhuri, ''D-LSMA: Distributed link scheduling multiple access protocol for QoS in ad-hoc networks'', in Proceedings of IEEE GLOBECOM '04, November 2004, pp IRMA fro Wireless Mesh Networks November 2007

27 Single-Path Routing vs. Multi-path Routing Capacity/Throughput Single Path Routing Less throughput bounds compare to multi-path routing Multi-path Routing Load balancing over more routes, utilizing more capacity Traffic Split No Yes. Arbitrary splitting, difficult for wireless implementation Robust Overhead Algorithm Packet Ordering Dijkstra, Bellman-ford, OSPF (BGP) Resilient to link failures More Route-discovery overhead, More soft states Route-discovery by flooding Out-of-sequence problem Compatibility Good. Bad We prefer single path routing from system design perspective..

28 About interference models A node cannot derive the distance of a hidden node in interference neighborhood because it cannot decode the packet and correlate this signal with the node identifier. Successful transmission 1) d ij R i 2) R k d kj R R i i d j i ij R i Unsuccessful transmission 1) d ij R i 2) R k d kj R i i d ij j k d k j k R k d k j k R k

29 Conflict Graph Coloring Conflict Graph Connectivity Graph G Conflict Graph G Each edge in G is a vertex in G Two adjacent vertexes in G cannot have same color

30 IRMA-MH Example Link bandwidth 1Mbps, 10 slots in each TDMA frame 3 flows: 15 14, 5 10,11 1 Each flow has offer load 0.2Mbps, demanding 2 slots per frame 2 G (V,E) Slot No After Scheduling flow for 1 st slot demand After scheduling flow 5-10 for 1 st slot demand After Scheduling flow 11-1 for 1 st slot demand After Scheduling flow 15-4 for 2nd slot (fail!) , , , , , , , Scheduling results: Each flow only obtains a 1-slot/per frame bandwidth (0.1M throughput) No slot to assign edge 3-4 Impossible to accommodate those flows with 0.2Mbps throughput. Increase to 20slots/frame does not help, because each flow will require 4 slots/frame. 30

31 LP Formulation of Integrated Routing and MAC Scheduling (2) Find the analytical throughput bounds Min-hop routing + link Scheduling othe path for each <s, d> is known as the min-hop path. Joint routing/scheduling o Single path routing, but path is uncertain. omixed integer programming problem Observations and conclusions from previous and our LP analysis It s NP-hard to find all link conflict constraints in LP formulation. Possible optimal routing paths can be found to yield better throughput than min-hop paths Optimal solution needs global knowledge Our contribution: Offline Optimization + Online algorithms

32 Overview of Wireless Networking Technologies 100+ Mbps, low-power, short range, minimal mobility Local Mobility Global Mobility MANET Wi-Fi Femto Cell LTE/UMB UMA WiMAX Bluetooth UMTS/CDMA2000 GSM(2G) Range Low-rate cellular networks vs. High-speed LAN/PAN. Centralized Cellular Networks support global mobility, but with low link speed and complex architecture Wi-Fi based technologies provide high-speed link access, but limited mobility support. Inter-technology convergence: o UMA and Femto Cell 32 IRMA fro Wireless Mesh Networks November 2007

33 High-Speed Wireless Mesh Network (1) Scope of the HS-WMN cannot be large: [Gupta 00] :Given n nodes randomly located in a unit disk, the uniform per node throughput capacity with interference-free protocols is A small or mid-size network with less than 100 nodes. 33 IRMA fro Wireless Mesh Networks November 2007

34 High Speed Wireless Mesh Network (2) Mobility cause significant throughput reduction Even slight random movement cause route failure and packet loss. Uncertainty of routing discovery disrupts traffic and reduce throughput We focus on a multi-hop network with small number of nodes and minimal mobility. Each node is equipped with one or more short range, high-speed radios. 34 IRMA fro Wireless Mesh Networks December 2007

35 Application Scenarios for HS-WMN Wireless links can be used to replace Ethernet cables, phone lines, USB/Firewire Wireless connectivity among consumer electronics devices, PCs. Multimedia-content sharing in home among HDTV, DVD Player, ipod, XBOX Integrated voice, video and data services 35 IRMA fro Wireless Mesh Networks December 2007

36 Problem Definition & Current Approaches Requirement: High Throughput Bulk data transfer over wireless mesh architecture. Goal: Protocol and algorithm design to achieve maximal throughput for concurrent end to end multihop flows. Problems: What s the achievable end-to-end capacity given a certain small number of nodes in the topology? Is there a feasible design to approximate the capacity? What s the control overhead and how it scales with the traffic demands? Conventional Approach: IEEE ad hoc routing CSMA/CA is bad in multi-hop wireless networks. Poor interaction between two independent distributed L3 and L2 protocols 36 IRMA fro Wireless Mesh Networks November 2007

37 Cross-Layer Routing/MAC Optimizing the performance of individual layers is not like to work beyond a certain point [Barrett et al. 02]. A routing protocol needs to interact with the MAC layer in order to improve its performance. Adopting multiple performance metrics from layer-2 into routing protocols is an example. However, interaction between MAC and routing layers is so close that merely exchanging parameters between them is not adequate. Merging certain functions of MAC and routing protocols is a promising approach. It is particularly meaningful for multi-radio or multi-channel routing, because the channel/radio selection in the MAC layer can help the path selection in the routing layer. 37 IRMA fro Wireless Mesh Networks November 2007

38 Interference-free TDMA Scheduling Protocol Model of Interference Transmission range: R 1) d ij R i Interference range: R 2) R k d kj Conditions for a successful transmission: 1) d ij R i 2) any node k, such that d is not transmitting Scheduling requires perfect interference awareness to know exactly whose interference affect which node kj R ' k R i R ii Unsuccessful transmission d ij j d kj k R k Implication of R >R: A node cannot discover a hidden node Solutions come with overhead: Obtain location information from GPS and exchange Use a dedicate control radio to communicate in interference range Deliver scheduling information to k-hop neighborhood (k>1). Collision-free link scheduling not good as a per-packet mechanism. 38 IRMA fro Wireless Mesh Networks November 2007

39 System Approach Traffic Map Joint FDMA/TDMA/Routing Time/Frequency assignment 39 IRMA fro Wireless Mesh Networks November 2007

40 Theoretical background for Joint Scheduling/Routing Maximize end-to-end throughput : multi-commodity network flow problem (linear programming) Interference-free scheduling: coloring problem (graph theory) Finding maximal independent set in the conflict graph Connectivity Graph G Conflict Graph G Each edge in G is a vertex in G Two adjacent vertexes in G cannot have same color 40 IRMA fro Wireless Mesh Networks November 2007

41 A Typical Simulation Topology 41 IRMA fro Wireless Mesh Networks November 2007

42 Challenges for Online Approach Real-time characterization: Information gathering beyond neighborhood. Interference detection o K-hop approximation o A 2 nd radio monitors as far as up to interference range Run-time Optimization LP optimization algorithms are NP-hard Control overhead and robustness Concise, timely and accurate signaling. Solution: A dedicated control plane Same radio with a dedicated BW slice in time/frequency Another radio use a dedicated control channel 42 IRMA fro Wireless Mesh Networks December 2007

43 LP Formulation of Integrated Routing and MAC Scheduling s d s d s 3 7 G (V,E) 14 d 1 M concurrent flows from s to d L: link set selected as paths r: offer-load/demands for each flow Maximize Subject to: qr i f i M 1) Flow conservation i r i = 1 f keeps same along the path for, 2) Link capacity i f i f ( e) = f BW ( e), for each e L i i + 0 q 1 < si di Constraints for link conflict based on conflict graph in interference model 3) Fairness tradeoff > 43 IRMA fro Wireless Mesh Networks December 2007