Optics Communications 437 (2019) 276–284
Contents lists available at ScienceDirect
Optics Communications journal homepage: www.elsevier.com/locate/optcom
Adaptive dynamic wavelength and bandwidth allocation algorithm based on error-back-propagation neural network prediction Bo Liu a , Lijia Zhang b,c , Fu Wang b,c ,∗, Ming Liu b,c , Yaya Mao a , Lilong Zhao a , Tingting Sun a , Xiangjun Xin b,c a
Institute of Optoelectronics, Nanjing University of Information Science & Technology (NUIST), Nanjing, 210044, China School of Electronic Engineering, Beijing University of Posts and Telecommunications (BUPT), Xitucheng Road NO.10, Beijing, 100876, China c Beijing Key Laboratory of Space–ground Interconnection and Convergence, Beijing University of Posts and Telecommunications (BUPT), Xitucheng Road NO.10, Beijing, 100876, China b
ARTICLE
INFO
Keywords: Dynamic wavelength and bandwidth allocation Software defined networking WDM/TDM-PON Neural network
ABSTRACT With the emergence of new services such as high-definition video, cloud computing and virtual reality, Wavelength Division Multiplexing/Time Division Multiplexing-Passive Optical Network (WDM/TDM-PON) has become a key technology for future access network. The development of Software Defined Networking (SDN) and machine learning also brings new opportunities to the flexibility of WDM/TDM-PON. In this paper, an Adaptive Dynamic Bandwidth Allocation algorithm based on error-back-propagation Neural Network Prediction (ADBANNP) was proposed to reduce the Round-trip-time (RTT) delay in the bandwidth-scheduling period. A Dynamic Wavelength Allocation (DWA) algorithm and a novel architecture of WDM/TDM-PON was proposed to simple the multi-carrier source. The experimental results validated that the proposed algorithm could dynamically allocate wavelengths under multi-carrier at 12.5G intervals. Compared with the traditional algorithms, the proposed algorithm could reduce the transmission delay by 25%.
1. Introduction With the development of 5th Generation mobile communication (5G) and Fiber To The X (FTTX), Internet services have largely raised the requirements for the flexibility and capacity of optical access network. Optical line terminal (OLT) with capacity more than 1Tbps has been implemented, but the bandwidth flexibility needs to be improved [1]. WDM/TDM-PON technology has become a significant technology for the future access network. The Dynamic Bandwidth Allocation (DBA) in the access network is proposed to solve the bandwidth allocation problem of upstream traffic in PON, to avoid large delay and jitter caused by unbalanced load [2–5]. With the development of WDM/TDM-PON, it is necessary to allocate the wavelengths besides time slots. The flexibility of optical access network can be improved by solving the problem of joint assignment of wavelength and time slot [6]. SDN is an effective way to improve the flexibility of the access network. A remote controller is added to the WDM/TDM-PON to acquire the data of traffic load and the bandwidth resource in optical network unit (ONU) through a secure channel, so that it can execute the DWBA algorithm more efficiently and accurately. The DWBA includes the DBA and DWA. The DBA was proposed to allocate time slots to ONUs in EPON, which utilizes Time Division
Multiple Access (TDMA) to implement multiple access of multiple ONUs [2,7]. The interleaved polling with adaptive cycle time (IPACT) algorithm defines a range of services executed by OLT [8]. MultiPoint Control Protocol (MPCP) is used to manage upstream traffic and avoid the contending of time slots. The MPCP control message includes REPORT message and GRANT message. REPORT message is produced by ONUs to report the size of queues, whereas GRANT message is used to grant assigned wavelengths, time slots and start time from OLT to ONU. The queuing delay is produced during the authorization process from queuing the packets to receiving the GRANT message. The RTT delay is generated due to the interaction among control messages [9]. With the maturity of WDM/TDM-PON, the DWBA algorithm has been extensively explored for time slots and wavelengths [6]. Many researches worked on WDM/TDM-PON. An effective method to manage bandwidth resources and energy of WDM/TDM-PON was proposed to realize the joint allocation of wavelengths and time slots [10–12]. A novel scheme called Modified Stable Matching Algorithm (MSMA) was proposed to enhance the scheduling efficiency [10]. Lannoo designed a WDM/TDMPON architecture to improve the flexibility of access networks [11]. In addition to the allocation of the wavelengths and time slots, the threedimensional bandwidth allocation of the space division multiplexing
∗ Corresponding author at: School of Electronic Engineering, Beijing University of Posts and Telecommunications (BUPT), Xitucheng Road NO.10, Beijing, 100876, China. E-mail address:
[email protected] (F. Wang).
https://doi.org/10.1016/j.optcom.2018.12.064 Received 23 October 2018; Received in revised form 21 November 2018; Accepted 17 December 2018 Available online 21 December 2018 0030-4018/© 2018 Elsevier B.V. All rights reserved.
B. Liu, L. Zhang, F. Wang et al.
Optics Communications 437 (2019) 276–284
Fig. 1. Software-defined WDM/TDM-PON.
module, the bandwidth prediction module and the DWBA module. The computing power of controller allows some complex algorithms to run on these modules and facilitate the management of administrators. The controller can obtain the global bandwidth information and allocate the network bandwidth more efficiently. In the WDM/TDM-PON, a software-defined OLT is a key part. The OLT includes a Multi-wavelength Light Source (MLS) [17]. In the wavelength-scheduling module of OLT, multi-wavelength signals pass through an Arrayed Waveguide Grating (AWG) and each wavelength is allocated to different ports. All wavelengths are demultiplexed by AWG and then enter the corresponding intensity modulator (IM) to modulate the downstream signal. A 1:N optical switch is connected to each output of AWG. The controller can control optical switches. The 1*N wavelength switch realizes wavelength output from any port. The wavelength is selected and allocated to corresponding port through the optical switch array and linked to the ONU via a N:1 coupler. In ONUs, the downstream power is divided into two parts by a 1:2 splitter. One part enters a receiver to be demodulated, the other enters the RSOA to be remodulated as uplink data. The uplink signals of different ONUs converge to a fiber in a N:1 couple. The DWBA algorithm can adjust bandwidth through time slot and wavelength. This paper will conduct traffic prediction and dynamic bandwidth allocation under this framework to achieve the DWBA algorithm.
was also explored [13]. With the developments of machine learning and artificial intelligence, it brings new opportunities to the development of DWBA [14]. The traffic prediction based on neural network can be used to improve the performance of optical networks [15]. In this paper, a novel DWBA algorithm is proposed to reduce the delay of RTT and improve the bandwidth efficiency using a network traffic prediction module. In the WDM/TDM-PON, the controller can improve the flexibility of the bandwidth. The proposed algorithm can predict the network traffic based on the error-back-propagation neural network, so that the GRANT message can be sent before the beginning of next DBA period. The transmission of upstream traffic is more efficient than traditional algorithms. The rest of the paper is organized as follows. Section 2 introduces the software-defined WDM/TDM-PON architecture. Section 3 introduces the proposed DWBA algorithm. Section 4 gives the simulation and experimental results. The Section 5 is the conclusion. 2. Software-defined WDM/TDM-PON architecture In the traditional WDM/TDM-PON, the produce of the multi-carriers source often attracts the attention of researchers [16]. A low-cost and stable method to produce the multi-carriers source can be used for the WDM/TDM-PON [17]. In the proposed architecture, we use a cyclic frequency shifter to produce the multi carriers. The SD-WDM/TDMPON architecture is shown in Fig. 1. In this architecture, the control plane is separated from the infrastructure plane. The controller is responsible for the management of the upstream bandwidth in the OLT. The infrastructure plane is used to forward traffic packet without the bandwidth management module. The transmission of GRANT message and REPORT message will not be blocked by safety channel, which ensure the efficiency of the bandwidth scheduling [18]. Devices in the WDM/TDM-PON support the OpenFlow protocol, which communicate with the controller by the SouthBound Interface (SBI). The ONUs on each branch of the remote node consist of a sub PON. The similar architecture was discussed in EPON [2,7]. WDM/TDM-PON is divided into multi-sub PONs according to the physical location. The ONUs in a sub PON share a number of wavelengths. All ONUs participate in the dynamic bandwidth allocation in same sub PON. All wavelengths from sub PON are transmitted to each ONU through a power splitter. The Tunable Optical Filter (TOF) can be used to select the wavelength of downstream signal. The controller can receive the REPORT message and reply the GRANT message. It is composed of three parts: the request cache
3. Adaptive dynamic bandwidth allocation algorithm based on error-back-propagation neural network prediction The proposed software-defined WDM/TDM-PON architecture targets to acquire the load statistics of every ONU and then schedule bandwidth. However, the control message between ONU and OLT will lead to the RTT delay, which will decrease the efficiency of DWBA. Considering the similarity and self-correlation of Internet traffic, we can adopt the error-Back Propagation (BP) neural network algorithm to predict the network traffic for reducing the influence of the RTT delay. Based on the proposed architecture of WDM/TDM-PON and traffic prediction, we proposed a DWBA algorithm supporting service classification. The processing of DWBA includes DBA and DWA. The DBA can be executed in every bandwidth-scheduling period, for example 2 ms. The configuration of optical switches in DWA will introduce high latency, which cannot be executed in every DBA period. However, the DWA can be executed when the load condition is changed. 277
B. Liu, L. Zhang, F. Wang et al.
Optics Communications 437 (2019) 276–284
The real value of the output is R𝑚 . If the value of the output is different from real value, the error (E) is obtained: 1∑ 1∑ (R − O𝑚 )2 = (𝑅 − 𝑓 (𝑛𝑒𝑡𝑚 ))2 (5) E= 2 𝑚 𝑚 2 𝑚 𝑚 Then the algorithm performs the error backpropagation. The algorithm adjusts the weights of the neural network according to the difference between the predicted output and the real traffic size, so that the output of the neural network can be approximated to the expectation value. The correction weight and the threshold value are realized by the Gradient Descent algorithm. The error of the output node can be expressed with the weight of the output layer and the No. 2 hidden layer as follows:
Fig. 2. Neural network model.
𝜕E = 𝛿𝑚 ⋅ 𝑦𝑘 , 𝛿𝑚 = −(R𝑚 − O𝑚 )f ′ (𝑛𝑒𝑡𝑚 ). 𝜕𝑊𝑘𝑚 The threshold can be expressed as:
3.1. Traffic prediction based on error-back-propagation neural network
𝜕E = 𝛿𝑚 . 𝜕𝜃𝑚
Traffic prediction can improve the performance of the dynamic bandwidth allocation algorithm [15]. By accurately predicting the traffic, we can pre-allocate the bandwidth in the period of RTT delay of control messages. The back propagation (BP) neural network algorithm is a simple and stable nonlinear prediction method. The neural network adopts a structure composed of an input layer, two hidden layers and one output layer. There are 100 nodes in the input layer and No. 1 hidden layer. The input values are the traffic data recorded in previous DBA periods. The number of nodes in the No. 2 hidden layer is 30. There is only one nodes in output layer, as shown in Fig. 2. In the choices of the number of hidden layers and nodes, we need to consider the accuracy and speed of convergence. After verification of various parameters, the above parameters can achieve good results. x1 to xn are the input values of the neural network and a set of normalized real Internet traffic is used. wij and wjk are weights of neural networks and constantly adjusted in the training process to achieve the best prediction effect. y1 is the prediction value of neural network, representing the prediction results. After the inverse normalization, the result is the expected bandwidth request for the next DBA period. At the beginning of the network running, the prediction models have not been trained. The traffic prediction is not executed, but the controller records the traffic data. When data collection is completed, the controller train the neural network model and perform traffic prediction. During training, recoded data is divided into training part and validation part. The training of BP neural network includes two processes: information propagation and error backpropagation. The input is xi . The output of the No. 1 hidden layer is yj . The output of the No. 2 hidden layer is zk . The output of the output layer is 𝑂m . The weight of the input layer to the No. 1 hidden layer is wij and the threshold is 𝜃j . The weight of the No. 1 hidden layer to the No. 2 hidden layer is wjk and the threshold is 𝜃k . The weight of the No. 2 hidden layer to the output layer is wkm and the threshold is 𝜃m . The input data is processed from the input layer to the output layer and the transfer function is provided as follows: f (x) =
1 . 1 + 𝑒−𝑥
The output of the No. 1 hidden layer is: ( ) ∑ y𝑗 = 𝑓 𝑤𝑖𝑗 ∗ 𝑥𝑖 − 𝜃𝑗 = 𝑓 (𝑛𝑒𝑡𝑗 ).
(8)
W𝑘𝑚 (𝑛 + 1) = 𝑊𝑘𝑚 (𝑛) + 𝜂𝛿𝑚 𝑦𝑘 .
𝜂𝑚 is the learning rate between hidden layer 2 and output layer. Similarly, the threshold adjustment function is: 𝜃𝑚 (𝑛 + 1) = 𝜃𝑚 (𝑛) + 𝜂 ′ 𝛿𝑚 .
(9)
In the same way, we can introduce the adjustment function between the threshold and the weight of other layers: W𝑗𝑘 (𝑛 + 1) = 𝑊𝑗𝑘 (𝑛) + 𝜂𝑘 𝛿𝑘 𝑦𝑗 𝜃𝑘 (𝑛 + 1) = 𝜃𝑘 (𝑛) + 𝜂𝑘′ 𝛿𝑘 W𝑖𝑗 (𝑛 + 1) = 𝑊𝑖𝑗 (𝑛) + 𝜂𝑗 𝛿𝑗 𝑥𝑖 𝜃𝑗 (𝑛 + 1) = 𝜃𝑗 (𝑛) + 𝜂𝑗′ 𝛿𝑗
.
(10)
The above equations are the basis of the error-back-propagation neural network algorithm used in this paper. The performance of traffic prediction model is shown in Section 4. 3.2. Dynamic bandwidth allocation algorithm based on error-back -propagation neural network for traffic prediction In order to guarantee the quality of service (QoS) for different services, the services are divided into three levels: Expedited Forwarding (EF), Assured Forwarding (AF), and Best Effort (BE). EF service has a strict demand for time delay, which is the highest priority. AF service is insensitive to time delay and needs to keep the minimum guaranteed bandwidth, which is second highest priority service. The BE service has the lowest priority. In the WDM/TDM-PON, an ONU cannot occupy the time slots for a long time in order to ensure the fairness of QoS. According to the current average packet size, the bandwidth requests can be divided into light load request (LLR) and heavy load request (HLR). Light load request is the request whose size is smaller than the median of request size, whereas heavy load request is the request whose size is bigger than the median of request size. In the traditional dynamic bandwidth allocation algorithm, ONU first sends REPORT message to the controller. The controller sends GRANT to ONUs after receiving the REPORT message. ONU forwards upstream packets according to the GRANT message [9]. The bandwidth grant process is shown in Fig. 3. Q11, Q12, and Q13 respectively correspond to the EF, AF, and BE queues of ONU1. Q21, Q22, and Q23 respectively correspond to the EF, AF and BE queues of ONU2. Before next DBA period, the controller uses the trained error-back-propagation neural network to predict the size of traffic queues in advance according to the previous REPORT messages. The controller sends pre-GRANT message by traffic prediction. The order of upstream forwarding is EF-LLR, EF-HLR, AF-LLR, BE-LLR, AF-HLR, and BE-HLR. When the controller
(1)
(2)
(3)
𝑗
The output node: ( ) ∑ O𝑚 = 𝑓 𝑤𝑘𝑚 ⋅ 𝑧𝑘 − 𝜃𝑚 = 𝑓 (𝑛𝑒𝑡𝑚 ).
(7)
Then the weight adjustment function is expressed as follows:
𝑖
The output of the No. 2 hidden layer: ( ) ∑ Z𝑘 = 𝑓 𝑤𝑗𝑘 ∗ 𝑦𝑗 − 𝜃𝑘 = 𝑓 (𝑛𝑒𝑡𝑘 ).
(6)
(4)
𝑘
278
B. Liu, L. Zhang, F. Wang et al.
Optics Communications 437 (2019) 276–284
Fig. 3. Process of the DBA algorithm.
receives the real REPORT message in the next period, the GRANT message based on real traffic is calculated and sent out. At the beginning of the networking, it is necessary to initialize the network and basic parameters of the neural network model. The nonprediction ADBA algorithm is used to preheat the time of T0. The ADBA algorithm only forwards upstream packet by real REPORT message without prediction, as shown in Algorithm 1. Each service queue in ONU sends a REPORT message to the controller regularly. The controller collects these REPORT messages to obtain network traffic information 𝐴𝐹 and L𝐵𝐸 ) of each queue length in and calculate the median (L𝐸𝐹 𝑗 , L𝑗 𝑗 ONUs. Previous traffic data constitute a training set. A neural network model is generated for each queue in each ONU. The model is updated at every T𝑟𝑒 according to recent load statistics. After the time T𝑟𝑒 , the initialization stage of dynamic bandwidth allocation is completed. Then queues in ONUs have a trained neural network model, which can be used to predict the size of traffic in the next period. The bandwidth requests are divided into HLR and LLR according to the median of business request size obtained by the prediction results. The controller sends the pre-GRANT message to ONU before the beginning of the next period. In the period from the beginning of DBA period to receive the GRANT message, the ONU can send upstream packets according to the prediction results. After the controller receives the real traffic statistics, it evaluates the load condition and the median of queue size. Then controller predicts the queue sizes for the next cycle. We use two parameters to determine the load condition: light load coefficient (𝛼) and heavy load coefficient (𝛽) (𝛽 > 𝛼). W is the total bandwidth of the network. If the total bandwidth of requests is less than 𝛼 ⋅ W, the network is in light load condition. If the requested bandwidth is more than 𝛽⋅W, the network is in heavy load condition. The network is in a balanced load if the bandwidth of total requests is between the two values. After determining the load condition, the DBA algorithm is carried out and the GRANT message is sent. Then the controller immediately calculates the predicted bandwidth for the next cycle. Before completing the polling period, the controller sends the pre-GRANT message for the next cycle. Once the new polling cycle starts, ONU forwards the upstream packets according to the pre-GRANT message. The process of ADBA-NNP is shown in Algorithm 1. 𝑀𝑗𝑖 means the median of queue i (EF, AF, or BE) in the ONU(j). L𝑖𝑗 is the byte length of queue I in ONU(j).
large delay, so we should reduce the frequency of wavelength scheduling when the ONU bandwidth is sufficient. The sub PON is the unit of wavelength scheduling. The sum of the request bandwidth of each sub PON is used as a parameter to determine whether the wavelengths need to be scheduled. We assume that wavelength scheduling is performed at every 𝑇re time. When the neural network model of the queue is updated, the controller determines whether the wavelength needs to be reallocated. In the proposed PON, the wavelengths pool is divided into fixed wavelengths and shared wavelengths. Each sub PON has a fixed wavelength and other wavelengths are shared wavelengths. The fixed wavelengths cannot be scheduled and they are allocated to a fixed sub PON. The shared wavelengths can be allocated according to the DWA algorithm. After the controller obtains the wavelength bandwidth of each sub PON, the allocation of DWA is performed according to the total bandwidth of the sub PON. The Pseudo code of DWA is shown in Algorithm 2. Through the cooperation of DWA and DBA, the controller can realize dynamic scheduling in the two dimensions of wavelength and time slot. 4. Experimental setup and results In order to verify the performance of proposed DBA algorithm, we made a simulation to validate the performance. The ADBA-NNP algorithm was compared with the traditional IPACT algorithm and ADBA algorithm. The ADBA only used the real traffic REPORT to allocate the bandwidth without traffic prediction, shown in Algorithm 1. Table 1 provides the parameters in the simulation. The bandwidth of the sub PON was set to 10 Gbps and the sub PON included 16 ONUs. The traffic source was the same for each ONU. The upstream traffic was divided into three queues of EF, AF, and BE according to the physical requirement. The proportions of the three queues were 20%, 40%, and 40%, respectively. The traffic of EF service with a high priority was generated with the Poisson distribution model and had a fixed packet size of 70 bytes. The AF and BE traffic flows were self-similar traffic (H = 0.8) generated based on the ON/OFF model. The packet size distribution was between 64 and 1518 bytes and conformed to the threepeak distribution. The three peaks were at 64, 594 and 1518 bytes, which were closer to the real network situation. The BP neural network algorithm contained an input layer, two hidden layers and one output layer, as shown in Fig. 2. The number of iterations for the training was 20,000. The learning rate was 0.05 and the target was 10−10 . At the training phase, 1000 datasets of 100 traffic numbers were entered in each model. After two hidden layers, the output was the result of the prediction. The corresponding trained models were generated for the
In addition to the implementation of the DBA, DWA plays a vital role in WDM/TDM-PON. Wavelength scheduling depends on the configuration rate of the optical switches. Wavelength scheduling will cause a 279
B. Liu, L. Zhang, F. Wang et al.
Optics Communications 437 (2019) 276–284
Table 1 Simulation parameters. Parameters
Values
Total bandwidth Maximum polling cycle (T𝑟𝑒 ) ONU Cache size for each ONU Business types Protection interval between packets Communication delay of ONU and controller Minimum guaranteed bandwidth for AF queues Minimum guaranteed bandwidth for BE queues
10 Gbps 2 ms 32 10 M bytes 3 512 ns 200 us 122,500 bytes 122,500 bytes
three services to be predicted. When the network load was 1, the total requested bandwidth was equal to the capacity of the network. The 𝛽 is 0.75 and the 𝛼 is 0.6 in the simulation. At first, we analyzed the performance of traffic prediction models. ON/OFF model was used to generate the upstream data traffic. Then, the BP neural network was used to training and validation. We validated the BP neural network model in simulated traffic. In the training phase, recoded data was divided into two part: training part and validation part. Only the training part could be used to train the model, and the other data was used for validation to avoid the over fitting. In the process of training, the Mean Square Error (MSE) got to the target value after
Fig. 4. MSE vs. epochs.
twelve epochs of training, as shown in Fig. 4. The model had better fitting effect and faster convergence speed, and could meet real-time requirements in traffic prediction. 280
B. Liu, L. Zhang, F. Wang et al.
Optics Communications 437 (2019) 276–284
Fig. 5. (a) Fitting of training data (b) Validation of new data.
One hundred groups of data were randomly selected from the training set to input into the trained model, and the inputs were compared with the predicted values, as shown in Fig. 5(a). It can be seen that the blue curve (prediction values) almost coincides with the red curve (expectation values), which shows that the model has a good memory effect for the training set. Then, 100 groups of data outside the training set were input to the BP neural network model for prediction. The prediction results are compared with the real traffic, as shown in Fig. 5(b). The prediction results using the blue dotted line are not completely consistent with the actual traffic using the red solid line, but the fitting degree is high. Therefore, the prediction results can be used to guide the bandwidth scheduling in the access network. Then the traffic prediction model was used to the DBA algorithm. Fig. 6 shows the average delay of three algorithms. The ADBA-NNP algorithm was compared with IPACT algorithm and ADBA algorithm without traffic prediction. IPACT is a classical DBA algorithm. As shown in Fig. 6, the average delay of ADBA-NNP algorithm is the lowest under
the same load condition, followed by the ADBA algorithm without traffic prediction. Especially when the load is more than 0.7, the advantage of ADBA-NNP is obvious. When the load is 1, the ADBA-NNP algorithm is 91% lower than the IPACT algorithm in terms of average packet delay and 84% lower than the ADBA algorithm. The controller can send the pre-GRANT message to the ONUs before receiving the REPORT message. In this way, the proposed algorithm eliminates the RTT delay of the control message between the ONU and the controller. Fig. 7 shows the performance of different queues in the average packet delay. The packet delay of the EF service is always lowest. The bandwidth request of AF queues has also been satisfied to a certain extent. With the increase in the traffic load, the BE queues with the lowest priority are sacrificed and the packet delay increases faster. Fig. 8 shows the throughput for the three queues when the network load is increasing. When the load is light, the bandwidth can satisfy all kind of queues. Nevertheless, when the network load is heavier than 0.8, the BE queue is sacrificed to meet services of EF and AF. 281
B. Liu, L. Zhang, F. Wang et al.
Optics Communications 437 (2019) 276–284
Fig. 9. Resource utilization rates of ADBA-NNP, ADBA-NNP without traffic prediction and IPACT.
Fig. 6. Average delay of ADBA-NNP, ADBA-NNP without traffic prediction, and IPACT.
were calculated through the proposed DWA algorithm. The experimental device is shown in Fig. 10. The system is divided into three parts: the multi-carrier light source, the wavelength-scheduling unit, and the FPGA-based controller. In the multi-carrier light source, a continuous wavelength laser was used as the seed source. A fiber loop was used to generate stable multi-carriers. Firstly, the input light was put into a 50:50 coupler. Moreover, a 12.5G cyclic frequency shifter generated the frequency shift. The ONU generated virtual services, allocated them to three queues of EF, AF, and BE, and sent REPORT messages to the controller regularly. The controller recorded the REPORT message generated by the sub PON and sent the wavelength GRANT messages according to the proposed DWA algorithm. We monitored the dynamic wavelength assignment by virtual ONUs. In this experiment, the assigned wavelength was not modulated and it was only used to verify the proposed DWA algorithm. Wavelength scheduling is completed by the interaction between ONU and the controller. As shown in Fig. 10, a multi-carrier source was generated. Firstly, the RF source generated two radio frequency signals with the phase difference of 90◦ , which were loaded to the IQ modulator. The optical signal then passed through a band pass filter and an Erbium Doped Fiber Amplifier (EDFA). After passing through the polarization controller, the signal was put into the 2 ∗ 2 coupler. The other input was the Continuous Wave (CW) source at 1549.5 nm. The output signal of the 2 ∗ 2 coupler was a multi-wavelength signal. The output of the optical frequency shifter linked a Band-Pass Filter (BPF), which was used to filter out the signal outside the upstream multi-carrier bandwidth. Then the signal was amplified by an EDFA. The radio frequency signals were sinusoidal signals of 12.5 GHz. The passband of filter was from 1546.5 nm to 1549.5 nm. The amplified power of EDFA was 30 dB. The output power of the CW light source is -10 dBm. The output of Point A is shown in Fig. 11. The 29 wavelengths can be used for the wavelength scheduling. The controller is simulated with a FPGA. The controller is connected to optical switches. The bandwidth requests of ONUs are generated by virtual machine simulation. The wavelength-scheduling unit controls the output of the wavelengths. In this experiment, we assumed that there were 3 sub PONs. Each sub PON had 3 ONUs. The total number of available wavelengths was six. Three wavelengths were fixed wavelengths and each sub PON had a fixed wavelength. The other three wavelengths were shared wavelengths. The normalized bandwidth was used in the experiment. A wavelength could carry bandwidth requests of 6 units. The shared wavelengths should be allocated according to the network load and bandwidth requests. It was assumed that the heavy load index (𝛽) is 0.75. Table 2 shows the different load cases. The output wavelengths of the wavelength-scheduling node in Case 1 is shown in Fig. 12. The output of sub PON-1 includes one wavelength and the output of sub PON-2 includes two wavelengths. The output of sub PON-3 is one wavelength. Case 1 is in balance load condition. DWA algorithm allocates
Fig. 7. Average delay of ADBA-NNP algorithm with multi-priority queues.
Fig. 8. Throughput of different queues with different loads in ADBA-NNP algorithm.
As shown in Fig. 9, the bandwidth efficiency of the three algorithms varies with the traffic load. The three algorithms have the similar performance in bandwidth utilization when the network load is light. The bandwidth requirement of services is far less than the bandwidth capacity, so the traffic can be completely forwarded. When the traffic load is heavy, the performance of traditional algorithm is affected by the RTT delay between ONU and controller. The proposed ADBA-NNP algorithm can make full use of the bandwidth resources. When the network load is nearly 1, the bandwidth utilization of the ADBA-NNP algorithm is 29% higher than that of the IPACT and 25% higher than that of the ADBA algorithm without traffic prediction. The DWA algorithm was experimentally demonstrated. With the input parameters of the controller, the wavelengths to be authorized 282
B. Liu, L. Zhang, F. Wang et al.
Optics Communications 437 (2019) 276–284
Fig. 10. Schematic diagram of the experimental setup.
Fig. 11. Multi-wavelength spectrum.
Fig. 12. Wavelength assignment in Case 1.
Table 2 Four load cases. All bandwidth requests (Normalization unit) Case 1
BEF gro(1) 1
BAF gro(1) 1
BBE gro(1) 0.8
BEF gro(2) 3.2
BAF gro(2) 2.8
BBE gro(2) 3.1
BEF gro(3) 1.6
BAF gro(3) 1.3
BBE gro(3) 1
Case 2
BEF gro(1) 0.7
BAF gro(1) 0.6
BBE gro(1) 0.7
BEF gro(2) 0.6
BAF gro(2) 0.6
BBE gro(2) 1.1
BEF gro(3) 0.6
BAF gro(3) 0.6
BBE gro(3) 0.9
Case 3
BEF gro(1) 3.5
BAF gro(1) 2.4
BBE gro(1) 4
BEF gro(2) 2.3
BAF gro(2) 3.4
BBE gro(2) 3
BEF gro(3) 3
BAF gro(3) 4
BBE gro(3) 2.6
Case 4
BEF gro(1) 4.5
BAF gro(1) 2.4
BBE gro(1) 3.4
BEF gro(2) 1.8
BAF gro(2) 3.9
BBE gro(2) 3
BEF gro(3) 1.6
BAF gro(3) 3.3
BBE gro(3) 3.6
the wavelength according to the real bandwidth requirements of each sub PON. The requested bandwidth of sub PON-1 and sub PON-3 is smaller than bandwidth of the fixed wavelength. The bandwidth request of sub PON-2 is more than the fixed wavelength, so a shared wavelength are required for transmission. In the process of optical switching, the instability of the multicarrier and imperfect characteristics of AWG cause crosstalk between adjacent wavelengths. When the load traffic of Case 2 was input into the ONUs. The wavelength output is shown in Fig. 13. The total bandwidth of the case is 6.4, which belongs to the light load condition. The wavelengths were allocated according to the bandwidth requirement. Because the load of sub PONs is light and does not need to use the shared wavelengths, it is not necessary to perform the DWA algorithm. The fixed wavelength can fully meet the requirement of the upstream bandwidth. The output for Case 3 is shown in Fig. 14. The situation is in heavy load case. The wavelength was allocated according to the bandwidth requirement of the high-priority queue to ensure the transmission of the high-priority traffic. In this wavelength-scheduling period, DWA algorithm should meet the bandwidth requirement of the heavy load queue and ensure the efficiency of the high-priority service.
Fig. 13. Wavelength assignment in Case 2.
In a heavy load case, shared wavelengths are allocated according to the requirements of the EF queue. Although the bandwidth requests of different sub PONs are similar, the wavelength assignment varies with the length of EF queue. The results of Case 4 are shown in Fig. 15. The load condition is roughly the same to that in Case 3, but the requested bandwidth of the EF queues is different. Therefore, the result of the wavelength assignment in Case 4 is different from Case 3. The wavelengths is not only allocated by the size of the requested bandwidth, but also considering the size of high-priority queues. 283
B. Liu, L. Zhang, F. Wang et al.
Optics Communications 437 (2019) 276–284
Acknowledgments The financial supports from National NSFC of China (No. 61425022/ 61522501/ 61675004/ 61307086/ 61475024/ 61672290/ 61475094/ 61605013/ 61378061/ 61675030/ 61575026), the National High Technology 863 Program of China (No. 2015AA015501, 2015AA015502, 2015AA015504, 2015AA016904, 2015AA016901), Beijing Nova Program, China (No. Z141101001814048), the Fund of State Key Laboratory of IPOC (BUPT), China. Supported by BUPT Excellent Ph.D. Students Foundation, China (CX2018305). References [1] M.V. Dolama, A.G. Rahbar, Modified smallest available report first: New dynamic bandwidth allocation schemes in QoS-capable EPONs, Opt. Fiber Technol. 17 (2010) 7–16. [2] J. Zheng, H.T. Mouftah, A survey of dynamic bandwidth allocation algorithms for ethernet passive optical networks, Opt. Switch. Netw. 6 (2009) 151–162. [3] J. Lai, H. Huang, W. Chen, L. Wang, M. Cho, Design and analytical analysis of a novel DBA algorithm with dual-polling tables in EPON, Math. Probl. Eng. 2015 (2015) 919278. [4] C. Lei, H. Chen, M. Chen, Y. Yu, Q. Guo, S. Yang, S. Xie, Dynamic and balanced capacity allocation scheme with uniform bandwidth for OFDM-PON systems, Opt. Commun. 338 (2015) 106–109. [5] K. Sone, G. Nakagawa, S. Oda, M. Takizawa, Y. Hirose, T. Hoshida, Remote Management and Control of WDM-PON System for Fronthaul in Cloud-Radio Access Networks, in: proc. Eur. Conf. on opt. Commun. ECOC, 2017, pp. 1–3. [6] A.H. Helmy, H.A. Fathallah, Taking turns with adaptive cycle time a decentralized media access scheme for LR-PON, J. Lightwave Technol. 29 (21) (2011) 3340–3349. [7] C.A. Chan, M. Attygalle, A. Nirmalathas, Local-Traffic-Redirection-Based dynamic bandwidth assignment scheme for EPON with active forwarding remote repeater node, J. Opt. Commun. Netw. 3 (3) (2011) 245–253. [8] G. Kramer, B. Mukherjee, G. Pesavento, Interleaved polling with adaptive cycle time (IPACT): A dynamic bandwidth distribution scheme in an optical access network, Photonic Netw. Commun. 4 (1) (2002) 89–107. [9] A. Buttaboni, M.D. Andrade, M. Tornatore, A multi-threaded dynamic bandwidth and wavelength allocation scheme with void filling for long reach WDM/TDM PONs, J. Lightwave Technol. 31 (8) (2013) 1149–1157. [10] S. Basu, G. Das, Scheduling hybrid WDM/TDM ethernet passive optical networks using modified stable matching algorithm, J. Lightwave Technol. 32 (15) (2014) 2613–2622. [11] B. Lannoo, G. Das, A. Dixit, D. Colle, M. Pickavet, P. Demeester, Novel hybrid WDM/TDM PON architectures to manage flexibility in optical access networks, Telecommun. Syst. 54 (2) (2013) 147–165. [12] W. Xia, C. Gan, W. Xie, C. Ni, Priority-rotating DBA with adaptive load balance for reconfigurable WDM/TDM PON, Opt. Fiber Technol. 26 (2015) 142–149. [13] C. Bock, J. Prat, S.D. Walker, Hybrid WDM/TDM PON using the AWG FSR and featuring centralized light generation and dynamic bandwidth allocation, J. Lightwave Technol. 23 (12) (2005) 3981–3988. [14] N. Merayo, D. Juárez, Juan C. Aguado, I. de Miguel, R.J. Durán, P. Fernández, R.M. Lorenzo, E.J. Abril, PID controller based on a self-adaptive neural network to ensure qos bandwidth requirements in passive optical networks, J. Opt. Commun. Netw. 9 (5) (2017) 433–445. [15] F. Morales, M. Ruiz, L. Gifre, Luis M. Contreras, V. López, L. Velasco, Virtual network topology adaptability based on data analytics for traffic prediction, J. Opt. Commun. Netw. 9 (1) (2017) A35–A45. [16] X. Xue, W. Ji, K. Huang, X. Li, S. Zhang, Tunable multiwavelength optical comb enabled WDM-OFDM-PON with source-free ONUs, IEEE Photonics J. 10 (3) (2018) 7202008. [17] F. Tian, X. Zhang, J. Li, L. Xi, Generation of 50 stable frequency-locked optical carriers for Tb/s multicarrier optical transmission using a recirculating frequency shifter, J. Lightwave Technol. 29 (8) (2011) 1085–1091. [18] H. Khalili, S. Sallent, J.R. Piney, et al., A proposal for an SDN-based SIEPON architecture, Opt. Commun. 403 (2017) 9–21.
Fig. 14. Wavelength assignment in Case 3.
Fig. 15. Wavelength assignment in Case 4.
5. Conclusion In this paper, we studied the dynamic wavelength and bandwidth allocation of the software-defined WDM/TDM-PON. A novel WDM/TDMPON was proposed. A multi-carrier source schemes and wavelengthscheduling unit were adopted to simplify architecture of WDM/TDMPON. We proposed a dynamic wavelength and bandwidth allocation algorithm based on back propagation neural networks. The simulation results validated performance of the proposed ADBA-NNP algorithm. In addition, we implemented the scheduling of the wavelength of the 12.5 G interval wavelengths. The proposed algorithm effectively reduced the delay and improved the throughput. At present, the experimental system has defects in the dynamic wavelength assignment due to the instability of multiple carriers and the imperfect experimental system. A more stable multicarrier source can be used to reduce interchannel interference.
284