1. One-Hop Out-of-Band Control Planes for Low-Power Multi-Hop Wireless Networks
Chaojie Gu Nanyang Technological University
Rui Tan Nanyang Technological University
Xin Lou Advanced Digital Sciences Center, Illinois at Singapore
Dusit Niyato Nanyang Technological University
Today I will present our work on one-hop out-of-band control planes for low-power multi-hop wireless networks. This is a joint work with Rui, Xin and Dusit.

2. Low-Power Multi-Hop Wireless Networks
Ad hoc & easy deployment, good scalability
Centralized network control
WirelessHART (used in >8,000 industrial systems)
ISA100.11a
Low-power multi-hop wireless networks have been widely deployed in many systems such as smart buildings and manufacturing systems. The key advantage of the low-power multi-hop networks is that they allow ad hoc and easy deployment. They can also easily scale up to cover large geographic areas.
These systems often impose very high network performance requirement. To achieve these requirement, many these networks adopt the centralized network control scheme’\ to manage the routing and data transmission. Two examples are WirelessHART and ISA100.11a.

3. In-band vs Out-of-Band
In-band control plane
Out-of-band control plane X X X X
Undesirable coupling
Lose control in data-plane failures
No coupling
Simple and resilient
However, all these centralized network control systems adopt an in-band scheme, in which the control plane and data plane share the same data communication network. However, this scheme introduces undesirable coupling between the two planes.
For example, in the in-band control scheme, when a node lose connection with its neighbors in data plane, the controller cannot reach the node. Differently, in the out-of-band control scheme, the control and data planes use different networks. So, in the data plane network failure, the controller can still reach the disconnected node.
In this work, we propose to apply the out-of-band control plane for low-power multi-hop wireless networks, in which the control plane and data plane are two physically separated communication networks. In this way, the failures in the data plane network will not affect the control plane, and thus the controller recover the data plane network in time. Moreover, the controller can continuously monitor the data-plane network link qualities and adjust the routing and transmission scheduling in time to keep optimal network performance.

4. Related work Bandwidth aggregation
Homogeneous radio: FatVAP [NSDI’08], FastForward [MASS’13]
Heterogeneous radios: MultiNets [TECS’14], Multipath TCP [MobiCom’16]
Improve throughput rather than optimality, manageability
Out-of-band control plane
WASP [ANCS’14]
Wi-Fi Direct: data
LTE: control
Not for low-power networks
To implement the out-of-band scheme, each network node will need multiple network interfaces.
There are two categories of existing studies on exploiting multiple network interfaces.
The first category is bandwidth aggregation, it uses multiple network interfaces, either homogeneous or heterogeneous, to transmit/receive data simultaneously to increase data throughput. However, bandwidth aggregation only focuses on improving data throughput rather than optimality and manageability of the network.
The second category is designing out-of-band control plane for wireless networks. To the best of our knowledge, WASP is the only system that implements out-of-band control plane for multi-hop wireless networks. It uses Wi-Fi Direct and LTE of smartphones to build the data and control planes. However, WASP is not designed for low-power networks.

5. Outline Background and motivation Approach System design Evaluation
Conclusion
So far, I have introduced the background and motivation of this work, now I will present our approach.

6. Low-Power Wide Area Networks
Kilometers communication range
One-hop control plane
Good manageability
LoRaWAN
License-free ISM band
Open data link standard
Unmanaged network
Control plane
Data plane
In this work, we propose to use increasingly available low-power wide-area network technologies to build the control plane. This is because, with their long-range communication capability, the control plane can be a one-hop network that improves the network manageability and robustness.
If we use low-power short-range radios to build the control plane, it will be a multi-hop network as the data plane that faces the same network management issues.
There are several LPWAN solutions available at market, such as LoRaWAN, SIGFOX，NB-IOT, and Weightless-P. In this work, we choose LoRaWAN to build our system and gain insights.
This choice is due to its use of license-free ISM band, open data link standard and unmanaged network which means that we can set our own network, we don’t have to use the infrastructure provided by the ISP.

7. LoRaWAN Characteristics
Class-A Characteristics
Uplink-downlink asymmetry
Concurrent uplinks
Non-concurrent downlinks
Session initiated by end node
ALOHA MAC
Unacceptable collisions for control messages X
SF7
SF8
SF11
SF7
SF10
SF9
LoRaWAN defines three classes of hardware, class A, B, and C. We choose to design our system based on Class-A. The other classes are not widely supported by hardware devices. Although, LoRaWAN provides long-range communication capability, we need to manage some features of Class-A.
The first one is uplink-downlink asymmetry. LoRaWAN supports uplink concurrency if they use different spreading factors, but no downlink concurrency. Note that spreading factor is a parameter for modulation in LoRaWAN. Moveover, A LoRaWAN downlink frame from the controller to a network node must be in response to an uplink frame.
The second feature of LoRaWAN is its ALOHA MAC. If there are two nodes using the same spreading factor transmit at the same time, the two packets will collide. This is highly undesirable because the control plane packets are critical.

8. Outline Background and motivation Approach System design Evaluation
Conclusion
So far, I have presented our out-of-band approach. Now, I will present the detailed system design.

9. Energy Profiling LoRaWAN ≈ 2.94 ZigBee
Simplifies the control-plane network
Light control plane’s traffic
We conduct several experiments to profile the energy consumption of LoRaWAN. We use an example to help us get better understanding. In this example, a node will transmit a packet to another node that is one kilometer away. The ZigBee network needs at least 10 nodes to build a 9-hops network to finish this task while LoRaWAN network only needs a single hop.
We measure the power consumption of each network for transmitting same size packet. Based on our calculation, LoRaWAN’s total energy consumption is about 3 times of ZigBee’s.
Although LoRaWAN consumes more energy than ZigBee, the LoRaWAN network simplifies the control-plane network design due to its one-hop nature. Moreover, the concern of LoRaWAN’s higher energy consumption can be mitigated by the fact that the control plane’s traffic volume is much lower than the data plane’s.

10. TDMA X X Carrier sense is not available Communication session
t0, t1: Start and end transmission
t0’, t1’: Start and end reception
t0’: Unavailable
t1: Inaccurate
One-way time synchronization
Record t0, t1’ , Δ = t1’ - t0
Δ depends on SF and frame size X
Error 2.9 ms
To improve the reliability of control plane, we proposed to use TDMA to replace ALOHA. Note that the LoRa hardware does not have carrier sense capability, so we cannot implement CSMA.
Clock synchronization is a basis of TDMA. To synchronize the node and the controller, we should know the transmission time of a synchronization packet. This figure shows our synchronization approach. T0 and t1 represent the start and end time of transmission. T0’ and t1’ represent the start and end time of reception. However, t0’ is not available due to some hardware limitations. T1 is inaccurate because of OS delay. So in our approach, the communication latency is defined as t1’ subtract t0.
This figure shows the measured uplink latency under different spreading factors and frame sizes. The Y axis is the uplink latency and the X axis is the frame size. Different colors represent different spreading factors. We can see an interesting step patterns. With those measurement results as our prior knowledge, at run time, we can estimate the clock offset between the end node and the controller.
From our experiments, our approach can achieve a synchronization error of 2.9ms.

11. Heartbeat Time Slots Time … 1 1 1 2 3 1 1 2 3 1 1 1 2 3 1 1 2 3
In this work, we also design a heartbeat time slot mechanism.
We use this example to illustrate, In this example, there are 3 nodes. We evenly divided time into slots and allocate them in a round-robin fashion to the nodes. In these orange time slots, node 1 can transmit when it has pending control plane data.
Remember that each communication session must be initiated by the end node. However, sometimes the controller wants to actively control the network, but it has to wait for the next uplink transmission when the end node has pending control plane data.
To mitigate this issue, we use heartbeat time slots represented by the red slots. In these red slots, node 1 must initiate a communication session. In this way, the controller can periodically get a chance to send active control commands.

12. Outline Background and motivation Approach System design Evaluation
Conclusion
So far, I have presented our system design. Now, I will present the evaluation results.

13. System Prototype Node Controller Software Architecture
We use Raspberry Pi, a ZigBee module and a LoRaWAN module to prototype our system.
The nodes form a multi-hop data plane using ZigBee and a single-hop control plane using LoRaWAN.
The controller uses a LoRaWAN concentrator board that can receive packets over many channels at the same time.
The software architecture of our system is shown in the right figure but I will not discuss the details due to time limit.

14. Experiment Setting Testbed 1 controller 15 nodes Application
CTP [SenSys’09]: A collection protocol, maintain a minimum-cost routing tree
Interference
Source: A laptop
Wi-Fi: Channel 6
ZigBee: Channel 18
We conduct experiments on a testbed of a controller and 15 nodes. The nodes are placed at the grid points of a lab space.
We apply our system to physically separate the control plane and data plane of CTP, which is a distributed data collection protocol that aims to maintain a minimum-cost routing tree.
To create data-plane link quality variations, we placed a laptop close to the testbed to generate Wi-Fi traffic to interfere with the ZigBee. And we set the Wi-Fi and ZigBee to use same frequency.

15. Control Plane Performance
The first experiment is to evaluate the control plane performance in presence of interference.
We first estimate the per node energy consumption of the control plane. In the left figure, the Y axis represents control plane energy consumption and X axis represents heartbeat period. Different colors represent different intensities of interference. We can see that, the energy consumption decreases with the heartbeat period. This is because with longer heartbeat period, the nodes must transmit less frequently. Moreover, the energy consumption increases with the interference intensity due to the increased control-plane traffic.
Then, we measure the control-plane downlink frame delivery ratio. The results are shown in right figure. The Y axis represents control plane downlink frame delivery ratio and X axis represents heartbeat period. We can see that, the frame delivery ratio increases with the heartbeat period. This is because with a shorter heartbeat period, the controller will send more control commands over the downlinks. However, the capacity of the downlink is limited. This will lead to more downlink frame drops due to timeout. Moreover, even if the data-plane network experiences intensive interference, the frame delivery ratio is generally above 90%.

16. CTP vs Ours 2.97 mW additional power consumption
We also conduct a side by side experiment to compare CTP and our system.
The left figure shows the result of data-plane packet delivery ratio under different levels of Wi-Fi interference. Different colors represents different systems. When the Wi-Fi interference is low, CTP and our system achieve similar packet delivery ratios. When the interference is high, our system outperforms CTP by 15%.
The right figure shows the extra power consumption versus Wi-Fi interference. We can see that, the power consumption increases with interference, because with stronger interference, more uplink frames will be transmitted to report the link status.
However, our system only consumes about 3mW more power

17. Conclusion Physically separating control plane is desirable
Increase CTP’s PDR by 15% in strong interference
2.97 mW extra power consumption per node
To summarize, physically separating control plane is desirable in low-power multi-hop wireless networks.
Our experiment results show that we can increase CTP’s packet delivery ratio by 15% in strong interference with only 3 milliwatts extra power consumption per node.
Thank you. Now I will take questions.