1Samir R. Das

Stony Brook Mesh Router:Architecting a Multi-Radio

Multihop Wireless LAN

Samir R. Das

(Joint work with Vishnu Navda, Mahesh Marina and Anand Kashyap)

Computer Science DepartmentSUNY at Stony Brooksamir@cs.sunysb.edu

http://www.cs.sunysb.edu/~samir

2Samir R. Das

A New Opportunity Has Arrived!

Linksys WRT54G access point/router runs Linux. User programmable. Decent processor and memory. Costs $70.

Several router platforms provide multiple PC/mini-PCI/PCI card interfaces. Decent processor and memory. Can run FreeBSD/Linux. Costs $250-$400.

What a systems researcher can do with all these?

3Samir R. Das

Stony Brook Wireless Router

Traditional Wireless LAN needs “wired” connectivity to access points.

Deployment slow and expensive, particularly for wide area.

Wired Backbone

Access Points

Clients

Ethernet

4Samir R. Das

Get rid of the wires!

Use a mesh routing backbone. Clients can associate with any access point/router. Complete transparency. Multiple radio interfaces on each router assigned to different bands/channels.

Wired Backbone

Access Points/ Mesh Routers

Clients

Ethernet

5Samir R. Das

Architectural Choices Clients run on infrastructure mode.

Associate to a nearby AP. Unaware of the wireless backbone.

Use WDS (wireless distribution system) for inter-AP communication.

Use a routing protocol for inter-AP routing. Link state-based routing. Choice of link cost metric?

Multiple radios on each AP Channel assignment problem.

6Samir R. Das

Routing

Layer 2 handoff triggers routing updates.

Mesh network cloud of APs

7Samir R. Das

Routing

Handoff delay with Prism2-based cards and HostAP driver = 240ms at L2 + 28ms per hop at L3.

Mesh network cloud of APs

8Samir R. Das

Multihop Relaying Performance with Multiple Channels

Setup: 802.11b prism2-based cards. HostAP driver. Relaying on WDS links.

Gains over single channel not always spectacular.

Suspect radio leakage.

0

0.5

1

1.5

2

2.5

3

3.5

2 Hops, 1 Channel

2 Hops, 2

Channels

3 Hops, 1 Channel

3 Hops, 2

Channels

3 Hops, 3

Channels

Mbps

Base case: 1 hop throughput 5.5 Mbps

TCP throughput

9Samir R. Das

Channel Assignment Problem: Observations and Approaches

Channel switching takes time (~100ms) in COTS hardware

Rule out dynamic approaches. Statically? Semi-dynamically?

Channel assignment is a topology control problem. Two neighboring node can talk only when they have a radio

on a common channel. Ideally, one should jointly solve channel assignment and

routing. Our approach: Assign channels to radios to minimize

interference (objective), but preserve original topology (constraint).

10Samir R. Das

Conflict Graph-based Greedy Algorithm

Visits nodes in a certain order and assigns channels to radios such that all links from this node gets a channel.

Channel selection based on a greedy heuristic.

Maintain a conflict graph on the side to model interference. Compute the heuristic on this graph.

Centralized; but can be distributed.

3 nodes2 radios/node3 non-overlappingchannels

11Samir R. Das

Visits nodes in a certain order and assigns channels to radios such that all links from this node gets a channel.

Channel selection based on a greedy heuristic.

Maintain a conflict graph on the side to model interference. Compute the heuristic on this graph.

Centralized; but can be distributed.

3 nodes2 radios/node3 non-overlappingchannels

Conflict Graph-based Greedy Algorithm

12Samir R. Das

3 nodes2 radios/node3 non-overlappingchannels

Conflict Graph-based Greedy Algorithm Visits nodes in a certain

order and assigns channels to radios such that all links from this node gets a channel.

Channel selection based on a greedy heuristic.

Maintain a conflict graph on the side to model interference. Compute the heuristic on this graph.

Centralized; but can be distributed.

13Samir R. Das

Conflict Graph-based Greedy Algorithm

3 nodes2 radios/node3 non-overlappingchannels

Visits nodes in a certain order and assigns channels to radios such that all links from this node gets a channel.

Channel selection based on a greedy heuristic.

Maintain a conflict graph on the side to model interference. Compute the heuristic on this graph.

Centralized; but can be distributed.

14Samir R. Das

The Devil is in the Model Interference model (used in objective)

Current model: Two links on the same channel with a common node interferes. Nothing else interferes.

Future: Model overlapping channels and radio leakage. Model interference beyond one hop. Factor in load?

What to optimize? Minimize max interference. Maximize no. of concurrent transmissions.

Topology (used as a constraint) Current model: Preserve the original topology. Future: Use the sub-topology actually used by

routing.

15Samir R. Das

Can iterative approaches helpin lieu of joint optimization?

Convergence? Practicality?

Routing

Channel Assignment

Influencesinterference

Influences topology

16Samir R. Das

Random Graph-based Simulations

50 nodes. Dense network. 12 independent channels.

17Samir R. Das

NS-2 Simulations

50 node. Dense network. MAC layer capacity with Poisson traffic on each link.

9.5 xSeveral orders of magnitude

18Samir R. Das

Summary

Extend infrastructure-mode WLAN to a mesh network.

Complete client transparency. Handoff driven routing update. Multiple radio on each router. Channel

assignment problem.