1. Fat Virtual Access Points
Srikanth Kandula
Kate Lin, Tural Badirkhanli and Dina Katabi

2. State-of-the-art (802.11):
Connect to the AP with the highest RSSI

3. In Homes, Hotspots… Uplink Bottleneck 2+2+2+…
Problem 1:
AP Uplink ~ 2Mbps
(DSL/Cable Modems)
Wireless Link ~ 54Mbps
(Theoretical Max)
2+2+2+…
Uplink Bottleneck
Can Aggregate Bandwidth from nearby APs!

4. At Work … Load Imbalance 4Mbps 7Mbps
Problem 2:
20Mbps
Unnecessary congestion; nearby APs are idle
Spread load
Individual changes help globally
15Mbps
4Mbps
7Mbps
Load Imbalance
Divide Total Bandwidth Among Users

5. 2 Problems, 1 Solution User can aggregate bandwidth from all APs
State-of-the-art:
Abstraction:
Join closest AP
Join a Virtual AP, that is the sum of nearby APs
User can aggregate bandwidth from all APs
Compete for total  balance load across APs

6. Realize a “fat virtual AP” with only client-side changes

7. Basic Operation
2Mbps
2Mbps
20Mbps
But, what about receive?

8. Basic Operation
2Mbps
2Mbps
Drop
Power-Save Q
20Mbps
Pretend in power-save, so AP buffers when disconnected (similar to Chandra et. al. VirtualWiFi)
Divide Time and Data Across APs to get “Fat Virtual AP”

9. Realizing a Fat Virtual AP is Hard
Sustain TCP flows through each AP
Cannot lose packets yet Switch quickly
Which APs to connect to and for how long?
Some APs are more valuable than others
How to divide traffic across the APs?
FatVAP, an driver design
divides time across APs to maximize throughput
is transparent to APs and remote ends

10. FatVAP Overview Scan for available APs
Channel 6
Channel 6
Channel 1
Channel 36
Scan for available APs
Compute a schedule to divide time across APs
Switch APs as per schedule
Spread traffic by pinning flows to APs

11. How much time to spend at an AP?
Achievable Bandwidths– end-to-end e, wireless w
Useful fraction of time
Subsumes Wireless Link Quality, Contention at AP, APs uplink capacity

12. How to Divide Time Across APs?
AP Bandwidth (Mbps)
AP1
AP2
AP3
End-to-end Achievable 5 4 3
Wireless Achievable 8
Usable Fraction
100%
50%
38%
5 Mbps, 100% busy
Optimal = 7 Mbps
Pick APs Greedily, on End-to-end rate
! more bang for the buck if wireless b/w is large

13. How to Divide Time Across APs?
AP Bandwidth (Mbps)
AP1
AP2
AP3
AP4
AP5
AP6
End-to-end Available 1
4.5
Wireless Available 5
Usable Fraction
20%
100%
5 Mbps, 100% busy
Pick APs Greedily, on End-to-end rate
Pick APs Greedily, on Wireless rate
! cost to switch is ≈ 5 ms
! can’t linger too long (100ms period)
Only 75% usable
No Greedy Solution!

14. Like Bin Packing, maximize value with bounded cost!
Say, fi is fraction of time at APi
Let s be switching time and D be the period
Value (Bandwidth)
Usefulness Constraint
Cost (Time)
Like Bin Packing, maximize value with bounded cost!
(pseudo)-polynomial solution

15. But, How to Estimate Bandwidths?
Wireless Achievable
Naively– send-rate of probe burst, APs report load
Idea: Use synchronous acks
Client TX Queue
AP Buffers
Time from head of tx queue to end of transmission (ack)

16. But, How to Estimate Bandwidths?
Wireless Achievable
End-to-end
Naively, send-rate of probe burst or APs report load t
Count bytes rcvd in a window
Idea: Use synchronous acks
Client TX Queue
AP Buffers
Time from head of tx queue to end of transmission (ack)

17. But, How to Estimate Bandwidths?
Wireless Achievable
End-to-end
Naively, send-rate of probe burst or APs report load t
Count bytes rcvd in a window
Idea: Use synchronous acks
May not receive data always
Idea: only count back-to-back large packets!
Client TX Queue
AP Buffers
Time from head of tx queue to end of transmission (ack)

18. How to Spread Traffic Across APs?

19. How to Spread Traffic Across APs?
Put flows through all APs
virtualize state
an IP for each interface
toggle APs (and channels)
By default, kernel sends all traffic to one AP
AP1
AP2
MIT /24
T-Mobile /24
Toggler
Hardware (Wireless Card)
State
AP1 State
AP2 State
Two Interfaces

20. How to Spread Traffic Across APs?
Put flows through all APs
virtualize state
an IP for each interface
toggle APs (and channels)
By default, kernel sends all traffic to one AP
Spread flows to APs
Fast header re-writing
AP1
AP2
Toggler
Hardware (Wireless Card)
AP1 State
AP2 State
Two Interfaces
Spreader
Distribute load w/o changing APs and applications

21. Switching Quickly Without Drops
A Driver For Each Interface?
warm-up cost on switch
one instance + soft-switch
AP1
AP2
One Driver

22. Switching Quickly Without Drops
A Driver For Each Interface?
warm-up cost on switch
one instance + soft-switch
Control Packets
Isolate Transitions
AP1
AP2
Hardware (Wireless Card)
AP1 State
AP2 State
Re-send AUTH
Send AUTH

23. Switching Quickly Without Drops
A Driver For Each Interface?
warm-up cost on switch
one instance + soft-switch
Control Packets
Isolate Transitions
Pkts stuck in driver at switch
Private Queues
AP1
AP2
Delay Switch till pkts drain
Drop Packets (madwifi)

24. Switching Quickly Without Drops
A Driver For Each Interface?
warm-up cost on switch
one instance + soft-switch
Control Packets
Isolate Transitions
Pkts stuck in driver at switch
Private Queues
AP1
AP2
Attach/Detach Queue = Pointer Swap
Enables high-rate TCPs through multiple APs

25. FatVAP Realizes a Fat Virtual AP
Which APs to connect to and for how long?
Estimate Bandwidths, Solve Optimization
How to divide traffic across the APs?
Virtualize, Pin Flows to APs, rewrite headers
Switch quickly but without losing packets
In-driver, Private Queues, Isolation
And, with only client-side changes

26. Related Work VirtualWiFi (Microsoft Research)
AP Selection (Intel Research, U Michigan)
SyncScan (UCSD)
MadWifi (open-source)
Divide Time across APs to maximize throughput
Sustain TCP flows through multiple APs
Transparently spread traffic across APs

27. Results

28. Experimental Setup Compare FatVAP driver with unmodified MadWifi
Scenarios
Testbed built from Cisco, NetGear and MadWifi APs
Residential deployments
Commercial hotspots
Traffic
Long-lived TCP flows
BitTorrent (Azureus client, Planetlab peers)
Mimic Web Browsing (modified WebStone)

29. Can FatVAP Aggregate Bandwidth?
6Mbps
Throughput (Mb/s)
Number of APs
~22 Mbps
Aggregates end-to-end up to the wireless bottleneck

30. Can FatVAP Balance Load?
2Mbps
12Mbps C1 C2 C3 C4 C5

31. Can FatVAP Balance Load?
2Mbps
12Mbps
Unmodified MadWifi
FatVAP 5 4 3 2 1
4.4
3.8
3.5
3.3
3.1
2.9
2.8
2.7
Throughput (Mb/s) C1 C2 C3 C4 C5 .9 .9
Need not worry about putting more APs where there may be more users, or worry about assigning different channel widths to users…
C1 C2 C3 C4 C5
C1 C2 C3 C4 C5
Simplifies Network Deployment!

32. Can FatVAP Adapt to Changes?
5Mbps
15Mbps
Throughput (Mb/s)
Re-adjusts as necessary

33. Contributions A new model for managed 802.11 LANs
Aggregate uplink, Balance load
First to realize a fat virtual AP
Divide time and traffic across APs
Transparent to APs, applications, servers