Integrated Routing & MAC Scheduling for Wireless Mesh Networks
Zhibin Wu
Dipankar Raychaudhuri
WINLAB, ECE Dept. Rutgers University
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
Emerging Broadband Radios and Throughput-Thirsty Devices
Short-range, high-speed radio technologies emerging:Technology Name Link Speed (in Mbps) BW (MHz) Frequency Band
802.11n 74-248 20 or40 2.4GHz and 5GHz
WiMedia 53-480 500+ 3.1 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
New Forms of Wireless Networks
“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
BTS •RNC •PDSN
InternetInternet
Wi-Fi
•WiMAX
•CDMA2000
•BS •ASN-GW
•Wireless Mesh Network
•4 | IRMA fro Wireless Mesh Networks| December 2007
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 802.11+ ad hoc routing
High-Speed Wireless Mesh Network
•5 | IRMA fro Wireless Mesh Networks| December 2007
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
•Self Interference
•Inter-flow Interference
•Hidden Node
•Exposed Node
•6 | IRMA fro Wireless Mesh Networks| December 2007
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
Joint Routing/Scheduling: Example
a 4x performance
gain can be achieved
by carefully selecting
routing paths and
MAC schedules
JOINITLY.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
(b)
12
3
1516
12
1314
ACK
ACK
ACK
ACK
t
Data
ACK
ACK
DataData
DataData
Data
DataData
ACK
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
(a)
RTS12
3
567
139
RTS
RTSCTS ACK
CTS ACK
RTS RTS
t
Data
RTSRTS
Data
•802.11 MAC with designated Min-hop paths
•Synchronized TDMA MAC with better paths
•8 | IRMA fro Wireless Mesh Networks| December 2007
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
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
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
Numerical Results
Gaps between the throughput bounds obtained by LP optimizations and a 802.11 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
IRMA System Model
Global control plane and data planeAll control signaling on a separate plane
Parameters of IRMA component in data plane is determined by control algorithms
Control Algorithms
Global Control Plane (GCP)
Data Packet
Control Message
•Protocol stacks in IRMA system
•Application
•CSMA MAC
•Control Plane •Data Plane
PHY
TDMA MAC
•Controlled Routing
PHY
• Routing
IRMA Control Agent
•IRMA Algorithms
IRMA Component
•13 | IRMA fro Wireless Mesh Networks| December 2007
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.
•Traffic Variation
Detection
IRMA
• Optimization
Path/Schedule Adjust
Working
Bootstrap
Topology
DiscoveryLoad default
• IRMA parameters
•14 | IRMA fro Wireless Mesh Networks| December 2007
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
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 Fi , 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
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 Fi , i= 1,2 …, M
a) Generate link Metric based on available “free” TDMA Slots
b) finding shortest path for flow Fi with the “bandwidth” metric. and assign conflict-free TDMA slots for this flow
(a) (b) A C
D B
A C
B D
•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
DIRMA-MH
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
A B FD ECSREQ
SAPPSUPDSUPD
SREQ
SAPPSREQ
SREJ
•Slot State Transitions
•Solve Hidden Node and Exposed Problems
Free RX-OKTX-OK
Prohibited
Update Update Update
Updat e
Send SREQ, Receive SAPP
SREQ+SR EJ
SCAN
Update
SREQ+SR EJ SREQ+SR
EJ
Reserved
Receive SREQ, Send SAPP
•18 | IRMA fro Wireless Mesh Networks| December 2007
DIRMA-BR
Dynamic hop-by-hop routing/scheduling
1. Put “forwarders” in a candidate pool
2. Candidates are sorted based on the combination of two metrics:
o M1:Distance to d
o M2: Available BW for this hop and next hop
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
Min-hop direction
Interference-aware direction
Interfered Area
C
BA
D
•19 | IRMA fro Wireless Mesh Networks| December 2007
Performance Evaluation
Implement the GCP and IRMA algorithms in ns-2.28
A separate control radio and channel in GCP
Compare the performance
IRMA algorithms
Baseline approacheso DSDV+802.11o AODV+802.11
•Ns-2 Simulation Parameters
•20 | IRMA fro Wireless Mesh Networks| December 2007
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
Evaluation of the Control Overhead
IRMA approach reduce signaling overhead because per-flow control is much more efficient than per-packet control
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
•22 | IRMA fro Wireless Mesh Networks| December 2007
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
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
Please visit http://www.winlab.rutgers.edu/~zhibinwu
•25 | IRMA fro Wireless Mesh Networks| December 2007
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 2006.
2. Z. Wu, S. Ganu, I. Seskar and D. Raychaudhuri, ''Experimental investigation of PHY layer rate control and frequency selection in 802.11-based ad-hoc networks'', in Proceedings of ACM SIGCOMM Workshops, August, 2005.
3. 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 2005.
4. 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 1670-1675
•26 | IRMA fro Wireless Mesh Networks| November 2007
Single-Path Routing vs. Multi-path Routing
Single Path Routing Multi-path Routing
Capacity/Throughput Less throughput bounds compare to multi-path routing
Load balancing over more routes, utilizing more capacity
Traffic Split No Yes. Arbitrary splitting, difficult for wireless implementation
Robust Resilient to link failures
Overhead More Route-discovery overhead, More soft states
Algorithm Dijkstra, Bellman-ford, OSPF (BGP)
Route-discovery by flooding
Packet Ordering Out-of-sequence problem
Compatibility Good. Bad•We prefer single path routing from system design perspective..
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) dij ≤
Ri
•2) R’k ≤
dkj
•k
•dk
j
•dij
•R ’k
•R’i•R
i •i •j
•k
•Unsuccessful transmission
•1) dij ≤
Ri
•2) R’k ≥
dkj
•k•R ’k
•Ri•i •j•d
ij
•dk
j
•R’i
Conflict Graph
Coloring Conflict Graph
Connectivity Graph G Conflict Graph G’
1 2
3 4
1
3
2
4
1
2
4
3
Each edge in G is a vertex in G’
Two adjacent vertexes in G’ cannot have same color
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
Slot No AfterSchedulingflow 15-14for 1st slotdemand
Afterschedulingflow 5-10for 1st slotdemand
After Schedulingflow 11-1for 1st slotdemand
After Schedulingflow 15-4 for 2nd slot
(fail!)
1 15-3, 7-14 15-3, 7-14 15-3, 7-14 15-3, 7-14
2 3-4 3-4 3-4 3-4
3 4-5 4-5 4-5 4-5
4 5-7 5-7 5-7 5-7
5 5-6 5-6 5-6
6 6-9 6-9 6-9
7 9-10 9-10, 4-1 9-10, 4-1
8 11-9 11-9, 15-3
9 9-6 9-6
10 6-4 6-4
•8
•2
•3
•5•1
•6
•7
•9
•4
•10
•11
•12
•13
•14
•15
•16
•17
•G (V,E)
Scheduling results:
Each flow only obtains a 1-slot/per frame bandwidth (0.1M throughput)
Impossible to accommodate those flows with 0.2Mbps throughput.
Increase to 20slots/frame does not help, because each flow will require 4 slots/frame.
No slot to assign
edge 3-4
•30
LP Formulation of Integrated Routing and MAC Scheduling (2)
Find the analytical throughput bounds Min-hop routing + link SchedulingoThe path for each <s, d> is known as the min-hop path.Joint routing/schedulingoSingle path routing, but path is uncertain.oMixed integer programming problem
Observations and conclusions from previous and our LP analysisIt’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 pathsOptimal solution needs global knowledge
Our contribution: Offline Optimization + Online algorithms
Overview of Wireless Networking Technologies
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:
oUMA and Femto Cell•Range
GSM(2G)
WiMAX
Wi-Fi
UMTS/CDMA2000
LTE/UMB
Global MobilityLocal Mobility
Femto Cell
MANET
100+ Mbps, low-power,
short range, minimal mobility
UMA
Bluetooth
•32 | IRMA fro Wireless Mesh Networks| November 2007
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
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
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
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 802.11+ 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
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
Interference-free TDMA Scheduling
Scheduling requires perfect interference awareness
to know exactly whose interference affect which node
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.
•Unsuccessful transmission
kR’k
Rii •jdij
dkj
R’i
Protocol Model of Interference
ittingnot transm is such that k, nodeany )2
)1'kkj
iij
R d
Rd
≤
≤
•Transmission range: R
•Interference range: R’
•Conditions for a successful transmission:
•1) dij ≤
Ri
•2) R’k ≥
dkj
•38 | IRMA fro Wireless Mesh Networks| November 2007
System Approach
Joint FDMA/TDMA/Routing Time/Frequency assignment
Traffic Map
Joint FDMA/TDMA/Routing Time/Frequency assignment
Traffic Map
•39 | IRMA fro Wireless Mesh Networks| November 2007
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’
1 2
3 4
1
3
2
4
1
2
4
3
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
A Typical Simulation Topology
•41 | IRMA fro Wireless Mesh Networks| November 2007
Challenges for Online Approach
Real-time characterization:
Information gathering beyond neighborhood.
Interference detection
o K-hop approximation
o A 2nd 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
LP Formulation of Integrated Routing and MAC Scheduling
∑=
M
1i ifMaximize
10 ≤≤≤≤ qrfqr iii
LeeBWfefi
i ∈≤=∑ each for ),()(
•8
•2
•3
•5•1
•6
•7
•9
•4
•10
•11
•12
•13
•14
•15
•16
•17
•G (V,E)
•M concurrent flows from s to d
•L: link set selected as paths
•r: offer-load/demands for each flow
•Subject to:
•+
1) Flow conservation>< iii dsf ,for path thealong same keeps
Constraints for link conflict based on “conflict graph” in interference model
•3) Fairness tradeoff
•2) Link capacity
•s1
•d1
•s2 •d2
•s3
•d3
•43 | IRMA fro Wireless Mesh Networks| December 2007