1. Horizon: Balancing TCP over multiple paths in wireless mesh networks
Bozidar Radunovic, Christos Gkantsidis, Dinan Gunawardena, Peter Key
Microsoft Research
Cambridge, UK

2. Wireless Mesh Networks

3. Goals Efficient use of resources “Fair” allocation of resources
Use multiple paths
“Fair” allocation of resources
Many users, long-distance vs. short-distance flows
Good application performance
TCP in particular
Deployable on existing network
Minor modifications of the existing stack.

4. Outline Controlling the network: Backpressure Dealing with TCP
Experimental results

5. Controlling the network: Backpressure
Algorithms to achieve utilization, fairness, good experience?
Protocol:
Which node transmits?
Which user/flow takes priority?
Where to forward the packets?
Our answer: Backpressure-based algorithms
Packet Queue
Flow A
Flow B
Essence of backpressure: give priority to
Nodes that have smaller queues
Difficult to implement
Flows that have smaller number of packets
Provably optimal
Very long theoretical support

6. Backpressure challenges
Flow A
Flow B
Our goal:
First to implement backpressure for wireless meshes
Lots of theory, no practice so far
Lots of challenges arise from practical constraints
Use multiple paths
Provide good performance for current applications
Applications
TCP Control Loop
Backpressure control
Wireless interactions
Examples:
Queuing by pressure triggers TCP congestion control
Back-pressure increases delay
Limited TCP window size penalizes path estimation
Out-of-order packet arrivals
... And many others
Complex Interactions

7. Our Multi-Path Routing
Input:
queuei(f): packet from flow f queued at node i
Output:
Ci(f): cost from node i to destination of flow f
bestFlowi: the flow to select for transmission at i
bNHi(f): the best next hop at i for flow f
Algorithm at node i:
Select where to transmit a packet: bNH(f)i = bestNextHopi(f) = argminj (queuei(f) / rate(i,j) + Cj(f))
Select the flow from which to transmit a packet: bestFlowi = argmaxf (queuei(f) / rate(i,bNHi(f)))
Update costs: Ci(f) = maxf(queuei(f) / rate(i,bNHi(f))) + CbNHi(f)
Propagate costs

8. Simple Example Main advantages:
Comparison:
Our scheme : 4 packets
Back-pressure: 13 packets
C1(1) = 4 S1
C2(2) = 6
C2(1) = 6
100
100
C1(1) = 2
C1(1) = 2
C3(1)= 4
Main advantages:
Minimal queuing: queue sizes do not grow with network
Estimates path quality with realistic TCP window size
Fast convergence S2 2 4 D2
C5(1) = 0
C3(1) = 2
C2(1) = 2
C4(1) = 0
100
100
C5(1) = 0
C6(1) = 0
C4(1) = 0 D1

9. Outline Controlling the network: Backpressure Dealing with TCP
Experimental results

10. Challenges dealing with TCP (1)
Our system performs congestion control ...
... so does TCP
... need to make sure that they are compatible
Idea: signal congestion to TCP
ECN-like approach
in some cases we communicate congestion by generating duplicate ACKs

11. Challenges dealing with TCP (2)
Recall: We use multiple paths ...
... TCP gets confused (path delay estimation, out of order delivery, etc.)
Solution: Use reassemble queue:
Minimize packet reordering
Avoid time-outs at all costs
TCP

12. Outline Controlling the network: Backpressure Dealing with TCP
Experimental results

13. Load balancing across two flows
Performance decreased: need more channels or better MAC
Both total rate and fairness improved
No significant difference
Flow S2-D2 stops
~ double the rate for both flows
2 disjoint areas
Only 6 nodes are dual-homed

14. Multi-homed networks Different access points/base stations
Same or different radios 2
1.5 1
0.5
Single flow
Relative improvement

15. More flows
More flows → all resources used → cannot increase total rate
Instead, we improve fairness (e.g. smallest rate)
8 flows

16. Goals Efficient use of resources “Fair” allocation of resources
Use multiple paths
“Fair” allocation of resources
Many users, long-distance vs. short-distance flows
Good application performance
TCP in particular
Deployable on existing network
No modifications of the existing stack.
How to select paths?
Other types of traffic?
How to deal with UDP?
What can we do with small changes?

17. Thank You

18. Optimizing utility of the network
Assume set of flows F, with fF
Flow conservation constraint:
For each node i and flow f:
Total flow into i for f = Total flow out of i for f
Wireless constraints:
Feasible node transmission scheduling and rates
Goal: max U(f) f
where U(f) is the utility function,
e.g. U(f) = log(xf), xf is the rate of f

19. Dual optimization problem
Primal problem
Dual problem
Goal: Maximize rate
Link constraints
i.e. rates & scheduling
Goal: Minimize cost
Link costs (price)
Solution of dual problem
Backlog

20. Challenges implementing dual problem
Goal: Minimize cost
Link prices
Difficulties:
Scheduling is NP-hard
Scheduling difficult to decentralize
Backpressure assumes queuing in the network
Use scheduling
Schedule only flows
⇒Approximate
Does NOT work well with TCP