1. Architecture and Evaluation of an Unplanned 802.11b Mesh Network
John Bicket, Daniel Aguayo, Sanjit Biswas, and Robert Morris MIT CSAIL
(MobiCom’05)
slides by Jong-Kwon Lee, presented by Zheng Zeng
January 20, 2006
Today, I’ll present a paper titled “…”.
This paper is from MIT and will be presented at MobiCom this year.

2. Community Wireless Networks
2 approaches to constructing community wireless networks
Carefully constructed multi-hop network with nodes in chosen locations + directional antennas for high-quality radio links
“Hot-spot” access points to which clients directly connect
Do not require much coordination
But, not much coverage per wired connection as multi-hop networks
 Combine the best characteristics of both approaches!
Unconstrained node placement (No planning)
Omni-directional antennas (No specifically engineered links)
Multi-hop routing (Improve coverage/performance)
Optimization of routing for throughput in a slowly changing networks (rather than for route repair in a mobile networks)

3. MIT Roofnet Roofnet Goal of this paper
Multi-hop b Internet access network
37 nodes spread over 4 km2 of a city.
Goal of this paper
Evaluation of the unplanned mesh architecture with a case study of Roofnet
End-to-end characteristics of Roofnet
c.f.) Previous work by the authors was on the physical-layer causes of packet loss (SIGCOMM’04)

4. Roofnet Design Urban and densely populated area
Mostly 3- or 4-story buildings
Each Roofnet node is hosted by a volunteer user.

5. Hardware
Roofnet node = PC b card + roof-mounted omni-directional antenna
The PC’s Ethernet port provides Internet service to the user.
802.11b wireless card in each node
Based on the Intersil prism 2.5 chip-set
The radios operate with RTS/CTS disabled.
All share the same b channel.
Use (Non-standard) pseudo-IBSS mode -> Nodes communicate directly without access points

6. Node Software
Node software = Linux + Routing S/W + DHCP server + web server
Roofnet Node (PC)
User
PC or
Laptop
From user’s perspective, the node acts like a cable/DSL modem.

7. Auto-Configuration: Addressing
Roofnet carries IP packets inside its own header format and routing protocol.
Each Roofnet node has an unique (internal) IP address of the form 10.x.x.x.  meaningful only inside Roofnet
The Roofnet S/W assigns itself addresses automatically, without requiring explicit configuration.
Low 24 bits = low 24 bits of Ethernet address
High 8 bits = unused class-A IP address block
A Roofnet node allocates IP addresses via DHCP to user hosts attached to the node’s Ethernet port.
From the reserved x IP address block
Uses NAT between Ethernet and Roofnet
[ x] -- (NAT) --> [10.x.x.x]

8. Auto-Configuration: Gateways and Internet Access
Assumption - A small fraction of Roofnet users share their wired Internet access links.
Identifying a Roofnet node as a gateway
On start-up, each Roofnet node checks to see if it can reach the Internet through its Ethernet port.
Succeed - advertise itself to Roofnet as an Internet gateway
Fail - acts as a DHCP server and default router for hosts on its Ethernet
Each gateway uses NAT between Roofnet and Internet.
When a node sends traffic through Roofnet to the Internet, the node selects the gateway to which it has the best route metric.
* Roofnet currently has 4 Internet gateways.

9. Routing Protocol Srcr: Roofnet’s routing protocol
Tries to find the highest-throughput route between any pair of Roofnet nodes.
Combination of link-state and DSR-style on demand querying
Srcr source-routes data packets, like DSR, in order to avoid routing loops when link metrics change.
Each Srcr node maintains a partial database of link metrics btw other pairs of nodes, and uses Dijkstra’s algorithm to find routes.
Each Roofnet gateway periodically floods a dummy query that allows all other nodes to learn about links on the way to that gateway. … (
Nodes learn fresh link metrics in 3 ways.
A node that forwards a packet over a link includes the link’s current metric in the packet’s source route, so that other nodes on the route can see the metric.
If a node needs to originate a packet but cannot find a route with its current database contents, it sends a DSR-style flooded query and adds the link metrics learned from any responses to its database.
Nodes that overhear queries and responses add the metric in those packets to their databases. )

10. Routing Metric Estimated Transmission Time (ETT) metric.
ETT predicts the total amount of time it would take to send a data packet along a route.
Each Roofnet node sends periodic 1500-byte broadcasts at each available b bit-rate and periodic minimum-size (60-byte) broadcasts at 1 Mbps.
ETT for a given link = expected time to successfully send a 1500-byte packet at that link’s highest-throughput bit-rate, including the time for the number of retransmissions predicted by the measured delivery probabilities.
ETT for a route = sum of ETTs for each of the route’s links
highest-throughput bit-rate = bit-rate with highest (delivery prob. X bit-rate)
t : e2e throughput
ti : throughput of hops

11. Bit-Rate Selection
In b, a high bit-rate with up to 50% loss is preferable to the next-lowest bit-rate.  Higher throughput with relatively high loss rates
SampleRate – Roofnet’s own algorithm to choose among the b transmit bit-rates of 1, 2, 5.5, 11 Mbps
Adjusts the bit-rate as it sends data packets over a link.
Like ETT, it judges which bit-rate will provide the highest throughput based on delivery probabilities measured at the different bit-rates.
Unlike ETT, its decision is based on actual data transmissions rather than on periodic broadcast probes.  It can adjust its choice more quickly and accurately.

12. Evaluation Method 4 sets of measurements on Roofnet
Multi-hop TCP – 15 sec 1-way bulk TCP transfer btw each pair of Roofnet nodes
Single-hop TCP – TCP throughput measured on the direct radio link btw each pair of nodes
Loss matrix – loss rate measured btw each pair of nodes using 1500-byte broadcasts at each b bit-rate
Multi-hop density – multi-hop TCP throughput measured btw a fixed set of 4 nodes, while varying the number of Roofnet nodes participating in routing.
Some of the analyses involve simulated route throughput calculated from the single-hop TCP.

13. Evaluation: end-to-end performance of Roofnet
Basic Performance
Basic measurements of throughput and latency over the network.
Link Quality and Distance
Study of how Srcr makes use of links
Effect of Density
Effect of node density on connectivity and throughput
Mesh Robustness
How Roofnet takes advantage of a highly connected mesh
Architectural Alternatives
Comparison btw multi-hop routing and a single-hop architecture
Inter-hop Interference
Inter-hop collisions are a major limiting factor in multi-hop throughput.

14. Basic Performance (1)
Distribution of throughput among all pairs of Roofnet nodes
Largely related to hop-count (see Table 1)
Med = 400 kbps
Avg = 627 kbps
(Figure 2)
The median is 400 kbps, and the average is 627 kbps.
Table 1 -> arranged by hop-count
Table 2 -> theoretical max throughput
Table 1 and 2 suggest that Roofnet’s one-hop routes operate at an average speed consistent with the 5.5 Mbps transmission rate, but that longer routes are disproportionately slower. This is discussed later, and the reason is that multi-hop routes suffer from inter-hop collisions and have lower performance than predicted.

15. Basic Performance (2)
TCP throughput to each node from its chosen Gateway
No problem in
interactive sessions …
The average throughput for each hop-count is typically higher because Roofnet GWs are more centrally located than the average Roofnet node.
Even at 4 hops, the average of 379 kbps is comparable to many DSL links.
The average latency was 22 ms to the gateways, and it is hardly noticeable in interactive sessions.

16. Link Quality and Distance (1)
Throughput vs Distance of each available link (top)
Only the links that Srcr uses in some route (bottom)
Srcr uses almost all of the links faster than 2 Mbps.
(TOP)
Most of the available links are between 500 and 1300 meters long, and can transfer about 500 kbps at their best bit-rate.
There are a small number of links a few hundred meters long with throughputs of 2 Mbps or more, and a few longer high-throughput links.
(Bottom) …
This means that fast short hops are the best policy.
=> Fast short hops are the best policy.

17. Link Quality and Distance (2)
CDF of delivery probabilities for the links used by Srcr at the bit-rate chosen by SampleRate.
Median = 0.8
=> Srcr often uses links with relatively high link-level loss rates.
A 1-hop route with high loss can deliver more data than 2-hop route with perfect links. …
This is because
A 1-hop route with …

18. Effect of Node Density
The following results are for the effect of node density on connectivity and throughput.
[Simulation] These results are simulated ones.
For each subset size n, a random set of n nodes are selected.
An estimate of the multi-hop throughput btw every pair in the subset is computed using only members of the subset as potential forwarders. …
The reason for increase in hop-count is that a denser network offers a wider choice of short high-quality links, even though using them causes routes to have more hops.

19. Mesh Robustness: # of Neighbors
The measurements of robustness investigate the benefits of the routing choices afforded by a mesh architecture and omni-directional antennas.
The first measure of a mesh network’s robustness is the # of potentially useful neighbors each node has.
(Figure 6)
Most nodes have many neighbors.
(Figure 7) …
Most nodes have many neighbors.
The majority of nodes use many neighbors in routing. (not through only 1 or 2 neighbors)
=> Roofnet makes good use of the mesh architecture.

20. Mesh Robustness : Vulnerability to Loss of Links
Another aspect of robustness is the extent to which the network is vulnerable to the loss of its most valuable links.
(Figure 8)
[caption]…
This figure shows that the best few links contribute …
But, dozens of …
The best few links contribute noticeably to average throughput.
But, dozens of the best links must be eliminated before throughput is reduced by half.

21. Mesh Robustness (cont’d) : Vulnerability to Loss of Links
Effect of cumulatively eliminating the best-connected nodes: Results from multi-hop density measurement
Best-connected nodes: those appearing in the most all-pair routes
The best two nodes are important for performance.
(Figure 9) shows the effect of …
This result is from multi-hop density measurement, while the previous figure is from the simulation.
Here, the best-connected nodes are those that appear in the most all-pair routes.
This figure shows that the best two nodes are important for performance.

22. Architectural Alternatives (1)
Comparison against communication over a direct radio link to a gateway (Access-point Network)
With Optimal Choice of Gateways
For single-hop – each successive GW is the node that maximizes the # of additional nodes (Conn) with non- zero throughput to some GW.
For multi-hop – similar, but with multi-hop connectivity and multi- hop throughput
=> Multi-hop has a throughput advantage because it can often use a sequence of short high- quality links rather than a single long low-quality link.
For evaluation of Roofnet’s architecture, comparison is performed against …
(Table 4) shows the comparison … …
In single-hop architecture, 5 GWs are needed to cover all Roofnet nodes.
[the latter of caption]
=> …

23. Architectural Alternatives (2)
With Random Choice of Gateways
=> As with optimal choice, multi-hop routing improves connectivity and throughput.
Careful gateway choice increases throughput for both multi-hop and single-hop routing.
(Table 5) shows the result with random choice of GWs.
In a single-hop architecture, 25 GWs are needed to cover all the nodes with random choice of GWs.
=> …
Comparison of the two tables shows that careful …

24. Architectural Alternatives (3)
For five or fewer gateways, randomly chosen multi-hop gateways provide better performance than carefully chosen single-hop gateways

25. Inter-hop Interference
Reason of degradation with multi-hop routes?
Concurrent transmissions on different hops of a route collide and cause packet loss.
Can RTS/CTS improve performance?
node pair
From Table 1, the results for basic performance, we can see that the throughput performance with each additional hop degrades fast.
(Figure 10) shows this phenomenon more clearly.
[caption]…
The reason can be explained as follows:
That is, concurrent transmissions …
So, we may consider using RTS/CTS to improve performance, since RTS/CTS mechanism is intended to prevent these collisions.
Table 6 shows the the effect of RTS/CTS on throughput for a few randomly-chosen node pairs.

26. Network Use
One of the four Roofnet gateway monitored the packets forwarded between Roofnet and the Internet.
The gateway’s radio was busy for about 70% of the monitoring period.
Mostly TCP traffic; only 1% UDP traffic
Traffic
Connections
Web requests to the Internet 7%
68%
BitTorrent peer-to-peer file sharing
30% 3%

27. Conclusions Ease of deployment
Omni-directional antennas
Self-configuring S/W
Link-quality-aware multi-hop routing
37 nodes with little administrative or installation effort on the part of researchers
Unplanned mesh architecture of Roofnet works well.
Average throughput of 627 kbps between nodes
The entire network is served by just a few gateways.