Wireless data flow control
Data flow control over the internet is one of the most widespread control applications on the planet. TCP is e.g. based on transmit packet window manipulation at the transmitting source, based on ACK/NACK feedback. Hence this control problem is a networked one, with the controller located in another node than the controlled process which resides at the other side of TCP.
Wireless data flow control will become increasingly important with the introduction of 5G wireless technology in the coming years, mainly for two reasons. First new use cases appear, like critical machine type communication (C-MTC). C-MTC aims to introduce wireless control of advanced high bandwidth processes in robotics, automotive safety, medicine and power distribution. In these applications quantities like latency and bit error rate will need to be reduced significantly as compared to what can be guaranteed today. Secondly, new wireless spectrum above 10 GHz will be exploited. At these frequencies the rate of variation of the radio channel and the associated wireless data rate both increase due to an increased radio shadowing in more challenging indoor environments.
Since the wireless interface is normally the bottle-neck, it needs to be fully utilized. The quickly varying wireless rate, together with the backhaul delays to the node that feeds the wireless transmit node with data, makes it necessary to use transmit data queues. These queues store data for coming wireless transmissions at the wireless interface, thereby avoiding data starvation in case of a sudden increase of the wireless data rate. Since these queues add to the round trip delays, their queuing delays need to be balanced against the likelihood of an empty queue, in order to keep the latencies down. It is the task of flow control to solve this networked control problem. Here the increasing number of users adds to the difficulty. This follows since flow control operates per data stream or bearer. Since there will be thousands of users served by each base station, and since each user may use many bearers, there will be a large amount of flow controller instances operating in parallel. This is why computational complexity needs to be a part of the overall design equation.
Three main requirements therefore needs to be met by new wireless flow controllers. They need to
- control the data stream so as to guarantee a set delay.
- be designed to be globally stable, to guarantee correct operation at all times.
- have a low and constant computational complexity.
The above requirements cannot be met by conventional internet traffic control proptocols like TCP. To improve, so called active queue management (AQM) has been introduced. AQM surveys the transmit data queues and discards packets in case the dwell times of packets in the queues grow too much. Such early packet discards means that NACK messages reach the internet data source early as well, a fact thatreduces the round trip delay. This is however not enough for the new 5G wireless use cases. As described below there are a number of principles that can be used to meet the new requirements. An overview of delay control and alignment for wireless feedback control appears in the paper 2.
Queue dwell time control
Since the above discussion indicates that a main part of the control problem is associated with the queue, a natural objective is to control the dwell time of the transmit data queue. As shown by Fig. 1, new packets enter the queue from a data source in another (logical) node, closer to the internet data source than the node connected to the wireless interface. The flow controller is located in this node. The two nodes are connected with a backhaul interface that is an additional source of delay. The queue is emptied by the wireless rate determined by the scheduler. The scheduler in hence in control of the wireless interface, with flow control providing a service to the scheduler by control of the transmit data queue dwell time. Since it is determined by the scheduler, the wireless rate is a measurable disturbance that directly affects the controlled process (the queue). It can therefore be used for feedforward control, as shown by Fig. 1. The leaky integrator is created by an overlaid AQM process that is assumed to be running in parallell.
Figure 1. The block diagram of the networked data flow control system of 12. The objective is to control the dwell time of data packets in the transmit data queue, which is represented by the leaky integrator to the right of the block diagram.
The main features of the process model include the backhaul downlink and uplink delays that affect the commanded bit rate in the downlink, as well the feedback and feedforward information in the uplink. The queue is modelled as a leaky integrator, due to the existence of overlaid AQM. Since the flow of data is one-directional, there is also a saturation affecting the loop.
The paper 12 proposes the use of a linear lead-lag feedback controller designed for L_2 stability by means of the Popov criterion. The paper also proposes a second order predictive feedforward, from the delayed measured wireless rate. As depicted in Fig. 1, embedding of a time invariant system is obtained by two nonlinear transformations. The controller can then be designed for transmit data volume control, rather than dwell time control which would require time varying techniques to be applied. The price is that the resulting dwell time is left free running. However, since also the queue dwell time reference value is transformed with the momentary wireless rate, the resulting reference value for the data volume of the queue, will be representative for reaching also the dwell time setpoint.
An example of the performance of the controller appears in Fig. 2. It is concluded in 12 that the controller meets the three requirements outlined above.
Figure 2. The performance of the data flow controller of 12. The dwell time reference value is 0.125 s.
The controller of 12 is robustly stable to delays up to 0.50 s. It does not account for any precise prior knowledge of the delay, except indirectly in the selection of controller bandwidth. The paper 13 generalizes the controller design to include such delay compensation. To do so, an LQG based design is performed, based on a state space model with an embedded rational delay approximation. The result is an increased delay robustness, up to about 0.65 s. The performance is also improved. A key design step in 12 and 13 is a pre-computation of the L_2 stability region, followed by a selection of controller parameters giving well defined delay robustness properties. Further improvements of robustness is obtained by the combined feedback and feed-forward H_2 controller of 10. There stability is maintained for delays exceeding 2 s. This is a consequence of the applied robust design.
Inter-node round trip delay control
A problem with the dwell time controllers is that they are stable only up to a maximum delay. To mitigate this problem, another control objective is studied in the paper 7. There, the dwell time feedback control objective is abandoned in favor of an objective to control the round trip time of packets in flight. A block diagram representing this objective and the resulting networked control loop appears in Fig. 3.
Figure 3. Block diagram of the wireless networked data flow control model. Dashed arrows indicate parts of the NCS outside the analyzed embedded time invariant control loop. Dash-dotted lines indicate borders between nodes and interfaces, while the dotted line indicates the intersection between time varying and time-invariant parts of the NCS.
The main parts of the networked control loop model again include the saturation, the leaky queue dynamics due to AQM, as well as the downlink and uplink backhaul delays. Also in this case embedding is applied in order to be able to work with time invariant controller design. The embedding differs from the one of 12 and 13 though, in that the dwell time of the queue appears explicitely in the model. The round trip time reference value is transformed with the same techniques as in 12 and 13, to give a data volume reference value, this time representing the desired data volume in flight. The controller now also contains a leaky integrator that keeps track of the sequence number of the most recently sent data item bar(v(s)).The feedback information v(s) is the sequence number of the latest data acknowledged to be correctly received by the end user. This way, the feedback signal y(s) can be formed to represent the number of data items (bits/packets) in flight. The control signal u(s) remains to be a commanded bit rate for transmission over the downlink backhaul interface. A detailed derivation of the model appears in the paper 7.
The key contribution of the paper 7 is the stability analysis. The following result holds when the controller C(s) is selected to be a constant C, i.e. when proportional control is applied:
Theorem: Consider the embedded time invariant networked control system defined by Fig. 3 in case proportional control is used. Assume that the assumptions A1-A5 of 7 hold. Then, when the AQM leakage rate approaches 0, the networked control systems obeys the Popov inequality for any set of delays of Fig. 3, and any controller gain $C>0$.
The result hence proves that the control loop is globally L_2 stable, irrespective of the delays that affect the control loop. Furthermore, this holds for any positive proportional controller gain. This is a very strong result, showing that the controller is probably ideally suited for the intended application. The theorem is illustrated by Fig. 4. As reported in 7, the performance in practice is also very good.
Figure 4. Popov plot for the controller of Fig. 3, using C=100 and a round trip delay of 0.125 s.
Delay skew control
With the introduction of the new fifth generation (5G) wireless networks, the flow control problem becomes subject to an additional complication. The reason is the new frequency bands in the millimeter-wave region. At these frequency bands radio shadowing becomes much more pronounced. The consequence is a need to use simultaneous transmission from multiple base stations to each single mobile (UE) in order to obtain a sufficient coverage. The problem is then that the delay from the data packet split point to the mobile may be very different and varying, between different transmission paths. This is so since some network interfaces may be congested, and the wireless interfaces may be subject to very different channel fading and shadowing that change rapidly. The consequence is that delay skew control needs to be introduced.
Figure 5. The architecture of the URLLC round trip time skew control system.
The block diagram of Fig. 5 illustrates such a delay skew control architecture. Here ultra-reliable-low-latency-communications (URLLC) is a general acronym for networked communication systems fulfilling the indicated requirements and it is almost identical to C-MTC. To describe Fig. 5, there is again a controller node where the application controllers are located. The controller node is connected to the wireless transmission nodes over multiple network interfaces that may consist of optical fiber, copper wire or wireless backhaul. The transmission nodes serve separate wireless 5G interfaces connecting to the plant node interface, which is the radio interface of a 5G mobile. The application layer is marked dashed, and it is here that plant controller commands are sent from the plant controller to the plant and where feedback signals are sent back. This transmission is not aware of the lower data bearer layer details.
The 5G wireless packet data networks have delay properties that are less guaranteed than in wired circuit switched networks. The objective of the delay skew controller is therefore to measure the delay over each involved transmission node and regulate these delays so that they do not differ. At the same time they all need to remain at the loop delay reference selected for the particular feedback control application. Disturbances working against this include varying network interface delays and varying wireless data rates due to shadowing and fading. In addition, the loop delay itself acts to make the round trip delay skew problem hard. The round trip delay skew controllers described here use the transmit data queues of Fig. 5 as actuators to achieve the above objectives. Finally, note that there needs to be one instance of the delay skew controller for each data flow.
Downlink delay skew control
The delay skew control objective varies with the application. In case of conventional mobile broadband services, the downlink data needs to arrive at the mobile within a pre-specified time window to avoid buffer induced restarts of the transmission. In this case the control objective is to control the delay skew at the downlink wireless interfaces. Such delay skew controllers are discussed in the papers 4, 5, 6 and 8.
The papers 6 and 8 treat the downlink delay skew controller depicted in Fig. 6. The controller employs cascade control with a MIMO outer delay skew control loop that controls n SISO inner data flow controllers, by controlling the reference dwell times of the transmit data queues of each transmission node. As stated above, the transmit data queues are the actuators of the control system. Suitable inner loop controllers include the ones of 10, 12 and 13. The use of cascade control is further enabled by the static decoupling that is used by the outer delay skew controller.
Figure 6. Block diagram of an n-node downlink delay skew control system.
Regarding the design of the interfaces, the paper 8 concludes that it is advantageous to have as much symmetry as possible between the data paths since that results in superior disturbance rejection properties. To illustrate the achievable performance a 5 node case was simulated. The results of Fig. 7 show that the controller performs as expected.
Figure 7. Performance of the delay skew controller of Fig. 6, when used in a five node setup. The signals of the different wireless transmission nodes are color coded.
The paper 6 is focused on the stability of the control system. There, the constraint that the dwell time reference commands from the outer loop needs to be positive is introduced. It is then shown that the integral quadratic control (IQC) stability results of Kao and Rantzer can be applied. The stability bounds obtained by this methodology appear to be about as sharp as those obtained in the paper 5, where a two node case is studied. The stability analysis of 5 does not need to use IQC analysis. Instead a loop transformation reduces the problem to SISO form, after which a combined use of the infinite dimensional versions of the Nyquist- and Popov-critera are used to asses global stability of the loop. The stability analysis of 4 is also based on the Popov-criterion despite the fact that an arbitrary number of transmission nodes are treated. The explanation for this is the simplifcation of the outer loop, to a reference value adjustment mechanism.
Round trip delay skew control
Whenever the application is a networked feedback control loop, the control objective needs to be selected as the round trip delay skews. This objective ensures that the loop delays are kept equal, low and with a minimal amount of jitter over all data paths. Such delay skew controllers are described in the papers 1, 3 and 11.
The paper 11 treats the dual connectivity round trip time skew controller defined by Fig. 8. This two node 5G architecture may be the first to be deployed. As compared to the block diagram of Fig. 6, it can be noted that the feedback signals consist of measurements of the round trip time of each of the two data paths. These are easily obtained from the ACK/NACK messages corresponding to the downlink data packages.
Figure 8. Block diagram of a dual connectivity round trip delay skew controller, suitable for support of C-MTC networked control.
Again a cascade controller with decoupling is used. The outer loop controls the reference values for the round trip delays of one SISO inner loop controller for each data path. These inner loop controllers are selected to be instances of the globally stable round trip delay controller of 7.
The performance of this round trip delay skew controller is good, as shown in Fig. 9. The delay skew controller is also proved to be globally stable in 10, using the same method of analysis as in 5.
Figure 9. Performance of the dual connectivity round trip delay skew controller.
The stability analysis of the n-node architecture of 1 is based on the Popov-criterion despite the fact that an arbitrary number of transmission nodes are treated. The explanation for this is the simplifcation of the outer loop, to a reference value adjustment mechanism. The results on stability of this paper are very strong with robust global L_2 stability, irrespective of the delays of the n data paths and the n proportional controller gains. This makes the delay skew controller ideally suited for C-MTC applications running over 5G communication networks, also at mmW frequencies.
1. T. Wigren, K. Lau, R. Delgado and R. H. Middleton, "Globally stable delay alignment for feedback control over NR multi-point connections", submitted to IEEE Trans. Contr. Network Systems (CONES), September, 2018.
2. R. H. Middleton, T. Wigren, L. Boström, R. A. Delgado, K. Lau, R. Karlsson, L. Brus and E. Corbett, "Feedback control applications in new radio - exploring delay control and alignment", IEEE Vehicular Technology Magazine, vol. 14, no. 2, pp. 70-77, June, 2019. DOI: 10.1109/MVT.2019.2892495.
3. R. A. Delgado, K. Lau, R. H. Middleton and T. Wigren, "Networked delay control for 5G wireless machine type communications using multi-connectivity" IEEE Trans. Contr. Systems Tech., vol. 27, no. 4, pp. 1510-1525, 2019.
4. K. Lau, T. Wigren, R. H. Middleton and R. Delgado, "Delay alignment control for 5G multi connectivity", Proc. IEEE CDC, Miami, FL, USA, pp. 7064-7070, December 17-19, 2018.
5. T. Wigren, K. Lau, R. Delgado and R. H. Middleton, "Delay skew packet flow control in 5G wireless systems with dual connectivity", IEEE Trans. Vehicular Tech., vol. 67, no. 6, pp. 5357-5371, 2018.
6. R. A. Delgado, T. Wigren, K. Lau and R. H. Middleton, "Stability properties of a MIMO data flow controller", ACC 2018, Milwaukee, USA, pp. 2638-2643, June 27-29, 2018.
7. T. Wigren and R. Karaki, "Globally stable wireless data flow control", IEEE Trans. Contr. Network Systems, vol. 5, no. 1, pp. 469-478, 2018. DOI: 10.1109/TCNS.2016.2619906 .
8. K. Lau, T. Wigren, R. Delgado and R. H. Middleton, "Disturbance rejection properties for a 5G networked data flow delay controller", IEEE CDC 2017, Melbourne, Australia, pp. 1681-1687, Dec. 12-15, 2017.
9. R. A. Delgado, T. Wigren, K. Lau and R. H. Middleton, "Stability properties of a MIMO data flow controller", Technical Report, University of Newcastle, August, 2017. Available: http://hdl.handle.net/1959.13/1344681 .
10. T. Wigren, "Loop-shaping feedback and feedforward control for networked systems with saturation and delay", Asian J. Contr., vol. 19, no. 4, pp. 1329-1349, 2017.
11. R. H. Middleton, T. Wigren, K. Lau and R. A. Delgado, "Data flow delay equalization for feedback control applications using 5G wireless dual connectivity", Proc. VTC 2017 Spring, Sydney, Australia, June 4-7, 2017.
12. T. Wigren, "Robust L_2 stable networked control of wireless packet queues in delayed internet connections", IEEE Trans. Contr. Sys. Tech., vol. 24, no. 2, pp. 502-513, 2016. DOI: 10.1109/TCST.2015.2455933.
13. T. Wigren, "Wireless feedback and feedforward data flow control subject to rate saturation and uncertain delay", IET Control Theory & Applications, vol. 10, no. 3, pp. 346-353, 2016. DOI: 10.1049/iet-cta.2015.0580.