A traffic prediction controller for on-chip optical network and on-chip optical network

Abstract

The present invention provides a traffic prediction controller and an on-chip optical network applied to an on-chip optical network, which are used to receive the number of data packets at the current moment and the number of data packets that are about to leave from the current moment to the next moment, and output the prediction judgment result or RWA scheme and power control information of the next moment based on the traffic predictor in the traffic prediction controller; the RWA scheme and power control information of the next moment are output so that the electrical layer can perform routing path selection, wavelength allocation and optical power allocation, thereby configuring the optical layer to complete the communication transmission of optical signals. The traffic prediction controller designed by the present invention is applied to the on-chip optical network, and the overall traffic matrix is ​​predicted based on its own traffic predictor, and specific RWA decisions and optical power allocation are made according to the traffic prediction results. Therefore, the present invention takes into account the inherent correlation between source and destination flows, and has the advantages of high prediction accuracy, low network latency, and low optical power consumption.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication network architecture, and specifically relates to a traffic prediction controller applied to an on-chip optical network and an on-chip optical network. Background Art

[0002] Optical Network-on-Chip (ONoC) is a new on-chip transmission method with an optical-electrical double-layer structure. The electrical layer of the on-chip optical network is responsible for the processing of electrical information, routing path reservation, etc.; the optical layer is responsible for the communication transmission of on-chip optical signals. Compared with traditional on-chip electrical networks, on-chip optical networks have the advantages of high transmission rate and low latency. Before transmission, the communication pair needs to be routed and wavelength assigned (RWA). The on-chip optical network adopts optical circuit switching; when the source node needs to transmit a data packet, it first sends a request to establish a routing path at the electrical layer, and at the same time controls the working state of the microring resonator (MicroringResonator, MR) on the corresponding routing path to reserve the optical transmission path on the chip, and then removes the previously reserved path after the information transmission is completed.

[0003] However, establishing a routing path at the electrical layer will cause the data packet to wait too long at the source node, which cannot fully utilize the low latency advantage of ONoC. If the traffic that will be generated at the next moment can be predicted and the path can be reserved in advance, the waiting time of the communication pair can be reduced. In addition, the light source power required by the communication pair during transmission is the main source of ONoC power consumption. If each communication pair uses the same light source power or configures the light source power according to the worst case OSNR, it is inevitable that excess light source power will be generated at the receiving end.

[0004] In on-chip optical networks, ARIMA models or linear prediction methods are usually used, which have low prediction accuracy. In existing network traffic prediction work, most of the prediction objects are end-to-end traffic size, that is, the size of a single source-destination flow, which will cause the intrinsic correlation between source-destination flows to be ignored, and the prediction model training time and hardware overhead will also increase. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a traffic prediction controller and an on-chip optical network applied to an on-chip optical network. The technical problem to be solved by the present invention is achieved by the following technical solutions:

[0006] In a first aspect, the present invention provides a flow prediction controller applied to an on-chip optical network, wherein the on-chip optical network comprises an electrical layer and an optical layer connected by silicon through vias, wherein the flow prediction controller is located in the electrical layer of the on-chip optical network and is connected to each electrical node of the electrical layer through electrical wires on the electrical layer, wherein the flow prediction controller is used to

[0007] Receive the number of data packets at the current moment and the number of data packets that will leave from the current moment to the next moment, and output the prediction result or RWA solution and power control information at the next moment based on its own traffic predictor;

[0008] The number of data packets is the number of data packets between the source node and the destination node in the electrical node;

[0009] The RWA scheme and power control information of the next moment are outputted so that the electrical layer can perform routing path selection, wavelength allocation and optical power allocation, thereby configuring the optical layer to complete the communication transmission of optical signals.

[0010] In a second aspect, the present invention provides an on-chip optical network, comprising the traffic prediction controller for the on-chip optical network described in the first aspect.

[0011] Beneficial effects of the present invention:

[0012] The present invention provides a traffic prediction controller and an on-chip optical network applied to an on-chip optical network, which are used to receive the number of data packets at the current moment and the number of data packets that are about to leave from the current moment to the next moment, and output the prediction judgment result or RWA scheme and power control information of the next moment based on the traffic predictor; the RWA scheme and power control information of the next moment are output to enable the electrical layer to perform routing path selection, wavelength allocation and optical power allocation, thereby configuring the optical layer to complete the communication transmission of optical signals. The present invention is designed to be applied to a traffic prediction controller of an on-chip optical network, which predicts the overall traffic matrix based on its own traffic predictor, and makes specific RWA decisions and optical power allocation according to the traffic prediction results. Therefore, the present invention takes into account the inherent correlation between source and destination flows, and has the advantages of high prediction accuracy, low network delay, and low optical power consumption.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of a 3D network architecture design of ONoC under a mesh topology based on a traffic prediction controller provided in an embodiment of the present invention;

[0015] Figure 2It is a flow chart of a flow prediction controller applied to an on-chip optical network provided by an embodiment of the present invention;

[0016] Figure 3 It is a specific structural design diagram of the flow prediction controller provided by an embodiment of the present invention and a connection relationship diagram with an electrical router;

[0017] Figure 4 A generalized flow prediction block diagram provided for an embodiment of the present invention;

[0018] Figure 5 A neural network structure diagram of a traffic predictor based on LSTM provided in an embodiment of the present invention;

[0019] Figure 6 A specific working flow diagram of a traffic prediction controller in an on-chip optical network provided by an embodiment of the present invention;

[0020] Figure 7 A structural diagram of the connection between the traffic prediction controller and the on-chip electrical router in the on-chip optical network provided by an embodiment of the present invention;

[0021] Figure 8 It is a connection structure diagram of the optical layer and the electrical layer in the on-chip optical network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0023] refer to Figure 1 The present invention provides an on-chip optical network, comprising the traffic prediction controller for the on-chip optical network. The present invention provides a traffic prediction controller for the on-chip optical network, referring to FIG1, Figure 1 The figure shows a schematic diagram of the 3D network architecture design of ONoC under the mesh topology based on the traffic prediction controller provided by an embodiment of the present invention. The on-chip optical network includes an electrical layer and an optical layer connected by silicon through vias, characterized in that the traffic prediction controller is located in the electrical layer of the on-chip optical network, and is connected to each electrical node of the electrical layer through an electrical wire on the electrical layer. The optical layer includes a plurality of optical nodes, and the optical nodes are connected together by waveguides. Each optical node is composed of an optical router and an OE interface. The optical router is composed of a microring resonator MR and a waveguide, and the MRs are connected by waveguides; the electrical nodes in the electrical layer are connected together by electrical wires, and each electrical node is composed of an electrical router and a processing unit, and the electrical router is connected to the processing unit by electrical wires.

[0024] It is worth noting that ONoC is a 3D structure consisting of an optical layer and an electrical layer. The electrical layer and the optical layer are connected by through silicon vias (TSV). Each optical node in the optical layer consists of an optical router and an OE interface. The optical router consists of a microring resonator (MR) and a waveguide. The internal waveguide is responsible for the transmission of optical signals, and the MR is responsible for controlling the transmission direction of the optical signal. The optical routers are connected by waveguides to complete the high-speed transmission of signals. The OE interface is responsible for the conversion between optical and electrical signals. Corresponding to the optical layer, each electrical node in the electrical layer contains an electrical router and a processing element. The electrical router is responsible for reserving the routing path of the optical signal when performing circuit switching. The processing element is responsible for generating and receiving electrical signals. The electrical router and the processing element are connected by on-chip electrical wires. In the electrical layer, the traffic prediction controller is responsible for the traffic prediction, routing decision and power allocation of the entire ONoC. It is located in the center of the electrical layer and is connected to the electrical router of each electrical node through on-chip electrical wires.

[0025] like Figure 2 As shown, Figure 2 A flow chart of a flow prediction controller applied to an on-chip optical network provided by the present invention is provided. A flow prediction controller applied to an on-chip optical network provided by the present invention is used for:

[0026] Step 1: Receive the number of data packets at the current moment and the number of data packets that will leave from the current moment to the next moment, and output the prediction result or RWA solution and power control information for the next moment based on its own traffic predictor.

[0027] Step 2: Output the RWA scheme and power control information at the next moment so that the electrical layer can perform routing path selection, wavelength allocation, and optical power allocation, thereby configuring the optical layer to complete the communication transmission of optical signals.

[0028] The number of data packets is the number of data packets between the source node and the destination node in the electrical node.

[0029] The traffic prediction controller is used to send the prediction judgment result or RWA scheme and power control information of the next moment to the electrical router; the electrical router is used to make routing path reservation, wavelength allocation and power allocation of the optical signal according to the prediction judgment result or RWA scheme and power control information of the next moment, and send MR control information to the optical router to control the working state of the corresponding MR, so as to select the specified light source wavelength and current to complete the communication transmission.

[0030] like Figure 3 As shown, Figure 3 It is a specific structural design diagram of the flow prediction controller provided by an embodiment of the present invention and a connection relationship diagram with an electrical router. The flow prediction controller of the present invention includes: a flow matrix acquisition module, a flow memory module, a flow predictor, a judgment module, an RWA module, a power control module and a gating module.

[0031] It is worth noting that the input of the traffic prediction controller is the number of packets at time t. t (s, d) and the number of packets leaving between time t and time t+1. t (s,d); The output is the predicted judgment result at time t+1, false t+1 (s,d) or RWA solution RWA t+1 (s, d) and power information P t+1 (s, d), is sent to the on-chip electrical router corresponding to the source node s and the intermediate nodes.

[0032] The traffic matrix acquisition module is used to output the traffic matrix at the current moment according to the number of data packets at the current moment, and to output the departure traffic matrix at the current moment according to the number of data packets that are about to leave from the current moment to the next moment.

[0033] Specifically: (1) The input of the traffic matrix acquisition module is the number of packets at time t, traffic t (s, d) and the number of packets that will leave from time t to t+1 lrave t (s,d), the output is the traffic matrix Traffic at time t t and the leaving traffic matrix Leave at time t t Traffic t is the traffic matrix at time t, where the element traffic t (s,d) represents the total number of data packets between source node s and destination node d at time t. t is the leaving traffic matrix, where the element leave t (s, d) represents the number of packets that are about to leave between the source node s and the destination node d from time t to t+1.

[0034] The flow memory module is used to store the historical flow matrix before the current moment and the flow matrix at the current moment.

[0035] Specifically: (2) The traffic memory module is responsible for storing the historical traffic matrix. The input is the traffic matrix Traffic at time t. tThe output is the historical traffic matrix Traffic at time t and k-1 times before time t t-k+1 ,Traffic t -k+2 ,…,Traffic t , which is the historical traffic matrix of a total of k moments. k is also called the time expansion step, which determines how many moments the neural network algorithm in the LSTM-based traffic predictor is expanded.

[0036] The traffic predictor is used to input the historical traffic matrix and the traffic matrix at the current moment into the trained LSTM neural network so that the LSTM neural network outputs the predicted traffic matrix at the next moment, and predicts the predicted arrival traffic matrix at the next moment based on the predicted traffic matrix at the next moment, the departure traffic matrix at the current moment, and the traffic matrix at the current moment.

[0037] Specifically: (3) The input of the traffic predictor is the historical traffic matrix Traffic at k moments t-k+1 ,Traffic t -k+2 ,…,Traffic t , the output is the predicted traffic matrix Predict at time t+1 t+1 . Predict t+1 is the predicted traffic matrix, where the element predict t+1 (s,d) represents the predicted number of all data packets between the source node s and the destination node d at time t+1.

[0038] The judgment module is used to judge whether each element value in the predicted arrival traffic matrix at the next moment is less than 0, and output the prediction judgment result of the next moment according to the judgment result. If the prediction judgment result is a prediction failure, the prediction judgment result is output to the gating module; if the prediction judgment result is a prediction success, the predicted arrival traffic matrix at the next moment is sent to the RWA module.

[0039] Specifically: (4) The input of the judgment module is the predicted arrival traffic matrix Arrive at time t+1 t+1 , where the element arrive t+1 (s, d) is the predicted number of newly arrived packets between source node s and destination node d at time t+1. t+1 The positive and negative elements in the output matrix of the prediction result at time t+1 are False t+1 , where the element is false t+1(s, d) represents the prediction result of the predicted data packet between the source node s and the target node d, and the value is 1 or 0. Prediction result false t+1 (s, d) is 1, indicating that the prediction fails, and the prediction is no longer performed through the RWA module and the power control module, and the prediction result is directly false. t+1 (s, d) is sent to the selection module; the prediction result is false t+1 (s, d) being 0 indicates successful prediction, and the RWA scheme and power are predicted through the RWA module and the power control module.

[0040] The RWA module is used to predict the RWA solution matrix at the next moment according to the predicted arrival traffic matrix at the next moment, and send the matrix to the gating module.

[0041] Specifically: (5) The input of the RWA module is the predicted arrival traffic matrix Arrive at time t+1 t+1 (s, d). According to the source-destination node information, output the RWA solution matrix RWA at time t+1 t+1 , where the element RWA t+1 (s,d) is the RWA scheme for the predicted data packet between source node s and destination node d.

[0042] The power control module is used to predict the power control information matrix at the next moment according to the RWA scheme matrix at the next moment, and send it to the gating module.

[0043] Specifically: (6) The input of the power control module is the RWA scheme matrix RWA at time t+1 t+1 , the output is the power control information matrix P at time t+1 t+1 , where the element P t+1 (s, d) is the power control information of the predicted data packet between the source node s and the destination node d.

[0044] The gating module is used to determine the output of the traffic prediction controller to be the prediction result, or the RWA solution and power control information according to the prediction result.

[0045] Specifically: (7) The input of the gating module is the prediction result matrix False at time t+1 t+1 、 RWA Solution Matrix RWA t+1 and the power control information matrix P t+1 . According to the prediction result matrix, the element is false t+1 The value of (s,d) determines the output of the flow prediction controller. t+1When (s,d) is 1, the output of the flow prediction controller is false t+1 (s,d); when false t+1 When (s,d) is 0, the output of the traffic prediction controller is RWA solution t+1 (s, d) and power control information P t+1 (s,d).

[0046] like Figure 4 As shown, Figure 4 A generalized flow prediction block diagram provided for an embodiment of the present invention. First, the flow data set is normalized to remove the influence of outliers in the data on the prediction accuracy. The flow data set comes from the historical flow matrix of M moments under the same network scale and topology. In order to ensure the prediction accuracy, the "M moments" should contain as much time domain information as possible.

[0047] Determine the time expansion step k. The time expansion step k determines how many historical traffic matrices are used to predict the traffic matrix at the next moment in the training model and the prediction model. If the storage space allows, the longer the time expansion step k, the higher the prediction accuracy.

[0048] Train the traffic prediction model. The training process requires input values ​​and observation values. In the data set, Traffic 1 ,Traffic 2 ,…,Traffic k As the input value at the first moment, Traffic k+1 As the observation value at the first moment; Traffic 2 ,Traffic 3 ,…,Traffic k+1 As the input value at the second moment, Traffic k+2 As the observed value at the second moment, and so on. When the mean square error between the predicted value and the observed value is less than the threshold or the number of iterations reaches the upper limit, the training of the traffic prediction model is completed and it is used as a traffic predictor.

[0049] After the training of the traffic prediction model is completed, the data of the traffic matrix is ​​normalized. The traffic predictor is based on the historical traffic matrix Traffic at time t and the previous (k-1) times. t-k+1 ,Traffic t-k+2 ,…,Traffic t (a total of k traffic matrices at time), predict the traffic matrix at time (t+1), and output the predicted traffic matrix Predict at time t+1 after denormalization t+1 .

[0050] The flow predictor of the present invention can be constructed based on the LSTM neural network structure, and can also be implemented based on neural networks such as FNN, CNN, ANN, etc. The present invention is not limited to the flow predictor implemented based on the LSTM neural network. As an embodiment, refer to Figure 5 , Figure 5 A neural network structure diagram of a LSTM-based traffic predictor provided in an embodiment of the present invention.

[0051] The traffic predictor embodiment of the present invention includes a normalization layer, an input layer, an LSTM layer, a random dropout layer, a fully connected layer, a regression layer, and a denormalization layer.

[0052] The traffic predictor is used to input the historical traffic matrix and the traffic matrix at the current moment after the training is completed, and output the predicted traffic matrix at the next moment through the normalization layer, input layer, LSTM layer, random inactivation layer, fully connected layer, regression layer and denormalization layer.

[0053] Among them, the first layer is responsible for standardizing the data and reducing the impact of outliers in traffic data on prediction accuracy.

[0054] The second layer is the input layer, which inputs data to the entire LSTM network.

[0055] The third layer is the LSTM layer, which maps k input data to the hidden feature space. Figure 5 h is the detailed structure of the LSTM hidden unit from time t-k+1 to time t. t is the output of the hidden unit at time t, C t is the cell state of the hidden unit at time t. t ,i t and t They are forget gate, input gate and output gate respectively. σ and tanh are activation functions.

[0056] The fourth layer is a random dropout layer, which randomly removes some neural units during the training process to prevent overfitting of the model training.

[0057] The fifth layer is a fully connected layer, which realizes the dimension transformation of data.

[0058] The sixth layer is the regression layer. The input of the LSTM-based traffic predictor is the traffic matrix Traffic at time t and k-1 times before time t. t-k+1 ,Traffic t-k+2 ,…,Traffic t , that is, the historical traffic matrix of a total of k moments, and the output is the predicted traffic matrix Predict at moment t+1 t+1 .

[0059] The traffic predictor is used to input the historical traffic matrix and the traffic matrix at the current moment into the trained LSTM neural network, so that the LSTM neural network outputs the predicted traffic matrix at the next moment, and input the predicted traffic matrix at the next moment, the departure traffic matrix at the current moment, and the traffic matrix at the current moment into the following formula to obtain the predicted arrival traffic matrix at the next moment;

[0060] arrive t+1 (s,d)=predict t+1 (s,d)-(traffic t (s,d)-leave t (s,d))

[0061] Among them, Arrive t+1 Represents the predicted arrival traffic matrix at time t+1, with the element arrive t+1 (s,d) is the predicted number of newly arrived packets between source node s and destination node d at time t+1. t+1 Represents the predicted traffic matrix at time t+1, where the element predict t+1 (s,d) is the predicted number of all packets between source node s and destination node d at time t+1. t Represents the traffic matrix at time t, where the element traffic t (s,d) is the number of all packets between source node s and destination node d at time t. t represents the departure flow matrix at time t, where the element leave t (s,d) is the number of packets that are about to leave between source node s and destination node d between time t and t+1.

[0062] The judgment module is used to judge the predicted arrival traffic matrix Arrive at the next moment. t+1 Each element arrives t+1 Is the (s, d) value less than 0? If so, the element arrives at time t+1. t+1 If the prediction result of the data packet corresponding to (s, d) is a prediction failure, the prediction result will be False t+1 Set to 1; if the element arrives t+1 If the (s,d) value is not less than 0, it is determined that the element arrives at time t+1. t+1 If the prediction result of the data packet corresponding to (s, d) is successful, the prediction result will be False. t+1Set to 0, and after the prediction and judgment are completed, the prediction and judgment results are sent to the selection module.

[0063] The gating module is used to t+1 When the prediction result is False, the RWA scheme and power control information at the next moment are used as the output of the traffic prediction controller; t+1 When it is 1, the prediction result is False t+1 As the output of the flow prediction controller.

[0064] Among them, the RWA solution matrix RWA at the next moment t+1 Includes multiple elements, each element RWA t+1 (s, d) is the RWA scheme for the predicted data packet between the source node s and the target node d; the power control information matrix P at the next moment t+1 Includes multiple elements, each element P t+1 (s, d) is the power control information of the predicted data packet between the source node s and the destination node d.

[0065] like Figure 6 As shown, Figure 6 The specific working flow diagram of the traffic prediction controller in the on-chip optical network provided by the embodiment of the present invention, the specific process is as follows:

[0066] 1. Train the traffic predictor. The data set is the historical traffic matrix of M moments under the same network scale and structure. In order to improve the prediction accuracy, M moments should contain as much time domain information as possible. When the curve of the mean square error between the predicted value and the observed value tends to converge with the number of iterations, the training ends.

[0067] 2. At time t, the on-chip routers of all nodes in the network extract the number of packets in the local data flow through the input buffer of the local input port. t (s, d) and the number of packets leaving between time t and time t+1. t (s, d) and sent to the traffic matrix acquisition module of the traffic prediction controller. The traffic matrix acquisition module constructs the traffic matrix Traffic at time t t , and stored in the traffic memory module. t Elements in traffic t (s, d) represents the number of all packets between the source node s and the destination node d at time t. At the same time, the traffic matrix acquisition module constructs the leaving traffic matrix Leave at time t t Leave t The element leavet (s, d) represents the number of packets that are about to leave between the source node s and the destination node d from time t to t+1.

[0068] 3. Based on the trained traffic predictor, according to the historical traffic matrix Traffic at time t and k-1 times before time t t-k+1 ,Traffic t-k+2 ,…,Traffic t (i.e., a total of k traffic matrices), predict the predicted traffic matrix Predict at the next moment (i.e., moment t+1) t+1 . Predict t+1 The elements in predict t+1 (s, d) represents the predicted number of all data packets between the source node s and the target node d at time t+1.

[0069] 4. Calculate the predicted arrival traffic matrix Arrive at time t+1 t+1 , where the element arrives t+1 (s,d) is the predicted number of newly arrived packets between source node s and destination node d at time t+1. The calculation formula is:

[0070] arrive t+1 (s,d)=predict t+1 (s,d)-(traffic t (s,d)-leave t (s,d))

[0071] When arrive t+1 When (s,d) is positive, it means that new data packets have arrived at time t+1, and the number and source-destination node information are known. t+1 When (s,d) is 0, it means that no data packet arrives at time t+1. t+1 When (s,d) is negative, it means the prediction failed.

[0072] 5. According to Arrive t+1 The positive and negative elements in the output matrix of the prediction result at time t+1 are False t+1 , where the element is false t+1 (s,d) represents the prediction result of the predicted data packet between the source node s and the target node d. The prediction result is false t+1 (s,d) is 1, indicating that the prediction fails, and the RWA module and the power control module are no longer used for prediction, and the prediction result is directly false. t+1(s,d) Send the strobe module; predict the result false t+1 (s,d) being 0 indicates successful prediction, and the RWA scheme and power are predicted through the RWA module and the power control module.

[0073] 6. For arrive with positive value t+1 (s, d), the RWA module makes an RWA solution at time t+1 for each predicted data packet at time t+1 t+1 (s, d), forming the RWA solution matrix RWA t+1 .

[0074] 7. When the routing path and wavelength of the data packet at time t+1 are determined, the power control module calculates the insertion loss Loss during the transmission of the data packet at time t+1 t+1 (s, d), and according to the receiver's receiving sensitivity P receiver , calculate the optical power P required to transmit this data packet at time t+1 t+1 (s, d) (unit: dBm), forming the power control information matrix P t+1 . Loss t+1 (s, d) represents the predicted insertion loss of the data packet between the source node s and the destination node d at time t+1, in dB. receiver Indicates the receiving sensitivity of the destination node's receiver, in dBm.

[0075] P t+1 (s,d)=Loss t+1 (s,d)+P receiver (dBm)

[0076] 8. The gating module determines the false elements in the matrix according to the prediction result. t+1 The value of (s,d) determines the output of the flow prediction controller. t+1 When (s,d) is 1, the output of the flow prediction controller is false t+1 (s,d); when false t+1 When (s,d) is 0, the output of the traffic prediction controller is RWA solution t+1 (s, d) and power control information P t+1 (s,d).

[0077] 9. The traffic prediction controller will predict the prediction result at time t+1 as false t+1 (s,d) or RWA solution RWA t+1 (s, d) and power control information P t+1(s, d) is returned to the source node s and the on-chip router corresponding to the intermediate node. t+1 (s, d) and power control information P t+1 (s,d), path reservation, wavelength allocation, and power allocation are performed. When the on-chip electrical router receives the prediction result false t+1 When (s, d), path reservation, wavelength allocation, and power allocation are not performed.

[0078] At time t, the error detector receives the number of packets from the local input port traffic t (s,d) and the RWA solution at time t in the storage unit of the routing control module t (s,d), RWA solution t The predicted number of packets of (s, d) is compared with the path information in the solution. If the number of packets is wrong, the source and destination node information is wrong, or the RWA solution is not received, the error is immediately reported. t (s, d) is set to 1 and sent to the routing control module. The routing control module receives an error with a value of 1. t After (s, d), XY routing is immediately performed for the arriving data packet, the first wavelength among the idle wavelengths is selected, and the worst-case optical source power is adopted. The worst-case optical source power is the optical source power of the path corresponding to the lowest OSNR in the network.

[0079] like Figure 7 As shown, Figure 7 A connection structure diagram of a flow prediction controller and an on-chip electrical router in an on-chip optical network provided by an embodiment of the present invention. The electrical router of the present invention comprises a routing control module, an electrical cross-switch module, an MR and light source control module, and an error detector, wherein the output of the flow prediction controller is connected to the input of the routing control module, the input of the MR and light source control module, and the input of the flow prediction controller is connected to the input port of the electrical cross-switch module.

[0080] The electrical crossbar switch module has six input ports and six output ports. The port information of the six input ports and six output ports of the electrical crossbar switch module is cached in a buffer corresponding to the port.

[0081] Buffer, used to extract the number of all data packets in the local data flow at the current moment and the number of data packets that are about to leave from the current moment to the next moment, and send the number of all data packets and the number of data packets that are about to leave to the traffic prediction controller and the error detector.

[0082] The error detector is used to compare the number of all data packets at the current moment with the predicted number of data packets and predicted path information in the RWA solution at the current moment, and output the detection result to the routing control module.

[0083] The error detector is used to calculate the number of traffic packets at the current moment. t (s, d) and the current RWA solution RWA t The predicted number of packets in (s, d) is compared with the predicted path information. If the number of packets is inconsistent, the source and destination node information is inconsistent, or the RWA solution RWA is not received t (s, d), it is determined that the prediction result at the current moment is wrong, and the error is set to error t (s, d) is set to 1, otherwise, the error bit is set to error t (s,d) is set to 0; the error bit is error t The result of (s, d) being 0 or 1 is taken as the detection result, and the detection result is sent to the routing control module.

[0084] Among them, error t (s, d) indicates whether the prediction of the data packet between the source node s and the target node d at the current time t is wrong. A value of 1 indicates an incorrect prediction, and a value of 0 indicates a correct prediction.

[0085] The routing control module is used for sending the routing decision of XY at the current moment to the electrical cross-connect switch module when the detection result is 1, selecting the first wavelength among the idle wavelengths as the communication wavelength, and configuring the light source power under the worst case, where the light source power under the worst case is the transmitted optical power of the path corresponding to the lowest OSNR in the network; when the detection result is 0, sending the RWA solution RWA at the current moment to the electrical cross-connect switch module t (s, d) and the corresponding power control information P t The routing decision of (s,d).

[0086] Specifically, the routing control module has a storage unit for storing the prediction and judgment results or the RWA scheme and power control information output by the traffic prediction controller, so that the routing control module executes the RWA scheme of the current moment predicted at the previous moment at the current moment, and stores the prediction and judgment results of the next moment predicted by the traffic prediction controller at the current moment, the RWA scheme of the next moment and the corresponding power control information in the storage unit.

[0087] The traffic prediction controller is used to predict the prediction result of the next moment, or the RWA scheme and corresponding power control information of the next moment based on the traffic predictor according to the number of all data packets at the current moment and the number of data packets that are about to leave from the current moment to the next moment, and send the prediction result of the next moment, or the RWA scheme and corresponding power control information of the next moment to the routing control module and the MR and light source control module.

[0088] The MR and light source control module is used to output the current of the specified light source according to the prediction result of the next moment or the RWA solution and the corresponding power control information, and send the current of the specified light source to the OE interface.

[0089] The routing control module is used to send the routing decision at the current moment to the electrical crossbar switch module according to the detection result of the error detector and the prediction result at the next moment or the RWA scheme at the next moment and the corresponding power control information.

[0090] The electrical crossbar switch module is used to configure the connection between the corresponding input port and the output port according to the routing decision at the current moment, so as to perform routing path reservation, wavelength allocation and power allocation, and cooperate with the generated MR control information to configure the optical layer to complete the communication transmission of the optical signal.

[0091] like Figure 7 As shown, the six input ports and six output ports of the electrical crossbar switch module are respectively input and output ports in four directions of east, south, west and north, local input ports and local output ports, vertical input ports and vertical output ports.

[0092] Among them, the input and output ports in the four directions of east, south, west and north are responsible for communicating with other on-chip electrical routers at the same layer. The local input port and local output port are responsible for connecting and communicating with the processing unit of the local node. The vertical input port is connected to the receiver of the OE interface through TSV; the vertical output port is connected to the light source driver of the OE interface through TSV.

[0093] The information of the six input ports and output ports is first cached in the input buffer. When the electrical crossbar switch receives the routing control information sent by the routing control module, the electrical crossbar switch configures the connection between its own corresponding input ports and output ports according to the routing control information.

[0094] At time t, the input buffer of the local input port extracts the number of all packets in the local data flow traffic t (s, d) and the number of packets leaving from time t to time t+1 leave t(s,d), and traffic t (s,d) and leave t (s,d) is sent to the traffic prediction controller.

[0095] traffic t (s,d) represents the number of packets between source node s and destination node d at time t. t (s, d) represents the number of packets that are about to leave between source node s and destination node d from time t to t+1. The traffic prediction controller returns the prediction result false at time t+1 through the prediction operation. t+1 (s,d) or RWA solution RWA t+1 (s, d) and power control information P t+1 (s,d).false t+1 (s, d) represents the prediction result of the traffic prediction controller on the predicted data packet between the source node s and the target node d at time t+1. Prediction result false t+1 (s,d) is 1, indicating that the prediction failed, and the prediction result is false t+1 (s,d) will be sent to the electrical router by the traffic prediction controller; the prediction result is false t+1 (s,d) is 0, indicating a successful prediction, and the traffic prediction controller no longer sends the prediction result false to the electrical router. t+1 (s,d), and send RWA solution RWA to the electrical router t+1 (s, d) and power control information P t+1 (s,d). RWA t+1 (s, d) represents the RWA solution for the predicted data packet between the source node s and the target node d at time t+1. t+1 (s, d) represents the power control information of the predicted data packet between the source node s and the destination node d at time t+1.

[0096] refer to Figure 8 , Figure 8 It is a connection structure diagram of the optical layer and the electrical layer in the on-chip optical network provided by an embodiment of the present invention. Figure 8 The OE interface includes a light source driver and a receiver.

[0097] The MR and light source control module is used to output the current of the specified light source according to the prediction result of the next moment or the RWA scheme and the corresponding power control information, and send the current of the specified light source to the light source driver of the OE interface.

[0098] The OE interface and the optical router are used to use the specified light source wavelength and current according to the current control of the specified light source, and to control the working state of the MR according to the MR control information, and to control the transmission direction of the optical signal in the optical router.

[0099] The receiver is used to convert the optical signal and transmit the optical signal to the electrical router through the vertical input port of the electrical crossbar switch module.

[0100] The electrical router communicates between on-chip electrical routers according to the link selection information in the RWA solution, and transmits the MR control information to the optical router through TSV so as to configure and reserve optical layer resources.

[0101] Specifically: The routing control module receives the RWA solution RWA t+1 (s, d) and power control information P t+1 (s,s), the electrical crossbar switch will be configured to perform routing path reservation, wavelength allocation, and power allocation. Figure 8 It is the connection structure between the optical layer and the electrical layer in the on-chip optical network. Among them, the microring resonator (MR) only allows optical signals with the same working wavelength to pass through the download port when it is in working state. The receiver (Receiver) can receive the optical signal with the same wavelength as it, and send it to the electrical router through the vertical input port to complete the conversion of optical signal to electrical signal. Different light sources (Laser) can emit light signals of different wavelengths. The light source driver (Driver) is responsible for driving the light source, by controlling the current I Laser , thereby controlling the light source transmission power. MR control information can control the working state of MR in the optical router. By setting the working state of MR in each part of the optical router, the transmission direction of the optical signal inside the optical router can be controlled. When making a path reservation, the electrical router follows the RWA solution RWA t+1 The link selection information in (s, d) enables communication between the source node s and the destination node d via the on-chip electrical router.

[0102] Each on-chip electrical router on the path then sends MR control information to the optical router through TSV to control the working state of the MR in the corresponding on-chip optical router. When power control is performed, the MR in the electrical router and the light source control module are connected according to the power control information P t+1 (s,d) to obtain the current I Laser , sent to the light source driver in the OE interface through TSV, controlling the use of the specified light source wavelength and current I Laser .

[0103] If the routing control module receives a prediction result of 1, false t+1(s,s), routing path reservation, wavelength allocation, and power allocation are no longer required in advance.

[0104] At time t, the error detector receives the number of packets from the local input port traffic t (s,d) and the RWA solution at time t in the storage unit of the routing control module t (s,d), RWA solution t The predicted number of packets in (s, d) is compared with the predicted path information. If the number of packets is inconsistent, the source and destination node information is inconsistent, or the RWA solution RWA is not received t (s,d), indicating that the prediction result is wrong, and the error detector immediately sets the error bit error t (s, d) is set to 1 and sent to the routing control module; otherwise, the error detector immediately sets the error bit error t (s,d) is set to 0. error t (s, d) indicates whether the prediction of the data packet between the source node s and the destination node d at time t is wrong. A value of 1 indicates an error in the prediction, and a value of 0 indicates a correct prediction. After receiving the error bit with a value of 1, the routing control module is responsible for making XY routing decisions for the arriving data packets, selecting the first wavelength among the idle wavelengths, and configuring the worst-case optical source power. The worst-case optical source power is the transmitted optical power of the path corresponding to the lowest OSNR in the network.

[0105] In summary, combined with Figure 7 as well as Figure 8 The input of the on-chip electrical router includes: the prediction result false at time t+1 from the traffic prediction controller t+1 (s,d) or RWA solution RWA t+1 (s, d) and power control information P t+1 (s, d), the received signal from the OE interface receiver, the data stream from the processing unit and the data from other on-chip electrical routers. The output of the on-chip electrical router includes: the number of packets output to the traffic prediction controller at time t t (s, d) and the number of packets leaving between time t and time t+1. t (s, d), the current I output to the OE interface light source driver VCSEL , MR control information output to the optical router, data stream output to the OE interface light source driver, data stream output to the processing unit, and data output to other on-chip electrical routers.

[0106] The present invention provides a traffic prediction controller and an on-chip optical network applied to an on-chip optical network, which are used to receive the number of data packets at the current moment and the number of data packets that are about to leave from the current moment to the next moment, and output the prediction judgment result or RWA scheme and power control information of the next moment based on its own traffic predictor; the RWA scheme and power control information of the next moment are output so that the electrical layer can perform routing path selection, wavelength allocation and optical power allocation, thereby configuring the optical layer to complete the communication transmission of optical signals. The present invention is designed to be applied to a traffic prediction controller of an on-chip optical network, and the overall traffic matrix is ​​predicted based on the traffic predictor in the traffic prediction controller, and specific RWA decisions and optical power allocation are made according to the traffic prediction results. Therefore, the present invention takes into account the inherent correlation between source and destination flows, and has the advantages of high prediction accuracy, low network delay, and low optical power consumption.

[0107] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0108] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality of components or steps.

[0109] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A traffic prediction controller for an on-chip optical network, the on-chip optical network comprising an electrical layer and an optical layer connected by through silicon vias, characterized in that: The traffic prediction controller is located at the electrical layer of the on-chip optical network and is connected to each electrical node of the electrical layer through electrical wires on the electrical layer. The traffic prediction controller is used to Receive the number of data packets at the current moment and the number of data packets that will leave between the current moment and the next moment, and output the prediction result or RWA solution and power control information at the next moment based on its own traffic predictor; The number of data packets is the number of data packets between the source node and the destination node in the electrical node; Output the RWA scheme and power control information of the next moment, so that the electrical layer can select the routing path, allocate the wavelength and allocate the optical power, thereby configuring the optical layer to complete the communication transmission of the optical signal; The flow prediction controller includes: a flow matrix acquisition module, a flow memory module, a flow predictor, a judgment module, an RWA module, a power control module and a gating module; The traffic matrix acquisition module is used to output the traffic matrix at the current moment according to the number of data packets at the current moment, and to output the departure traffic matrix at the current moment according to the number of data packets that will leave from the current moment to the next moment; The flow memory module is used to store the historical flow matrix before the current moment and the flow matrix at the current moment; The traffic predictor is used to input the historical traffic matrix and the traffic matrix at the current moment into the trained LSTM neural network, so that the LSTM neural network outputs the predicted traffic matrix at the next moment, and predicts the predicted arrival traffic matrix at the next moment based on the predicted traffic matrix at the next moment, the departure traffic matrix at the current moment, and the traffic matrix at the current moment; The judgment module is used to judge whether each element value in the predicted arrival flow matrix at the next moment is less than 0, and output the prediction judgment result at the next moment according to the judgment result. If the prediction judgment result is a prediction failure, the prediction judgment result is output to the gating module; if the prediction judgment result is a prediction success, the predicted arrival flow matrix at the next moment is sent to the RWA module; The RWA module is used to predict the RWA solution matrix at the next moment according to the predicted arrival traffic matrix at the next moment, and send it to the gating module; The power control module is used to predict the power control information matrix at the next moment according to the RWA scheme matrix at the next moment, and send it to the gating module; The gating module is used to determine the output of the traffic prediction controller to be the prediction result, or the RWA solution and power control information according to the prediction result.

2. The traffic prediction controller for on-chip optical network according to claim 1, characterized in that: The traffic predictor is used to input the historical traffic matrix and the traffic matrix at the current moment into the trained LSTM neural network, so that the LSTM neural network outputs the predicted traffic matrix at the next moment, and inputs the predicted traffic matrix at the next moment, the departure traffic matrix at the current moment, and the traffic matrix at the current moment into the following formula to obtain the predicted arrival traffic matrix at the next moment; arrive t+1 (s,d)=predict t+1 (s,d)-(traffic t (s,d)-leave t (s,d)) Among them, Arrive t+1 Represents the predicted arrival traffic matrix at time t+1, with the element arrive t+1 (s,d) is the predicted number of newly arrived packets between source node s and destination node d at time t+1; predict t+1 Represents the predicted traffic matrix at time t+1, where the element predict t+1 (s, d) is the predicted number of all packets between source node s and target node d at time t+1; Traffic t Represents the traffic matrix at time t, where the element traffic t (s, d) is the number of packets between source node s and destination node d at time t; Leave t represents the departure flow matrix at time t, where the element leave t (s,d) is the number of packets that are about to leave between source node s and destination node d between time t and t+1; The judgment module is used to judge the predicted arrival traffic matrix Arrive at the next moment. t+1 Each element arrives t+1 Is the (s, d) value less than 0? If so, the element arrives at time t+1. t+1 If the prediction result of the data packet corresponding to (s, d) is a prediction failure, the prediction result will be False t+1 Set to 1; if the element arrives t+1 If the (s,d) value is not less than 0, it is determined that the element arrives at time t+1. t+1 If the prediction result of the data packet corresponding to (s, d) is successful, the prediction result will be False. t+1 Set to 0; and after the prediction and judgment are completed, send the prediction and judgment results to the gating module; The gating module is used to t+1 When the prediction result is False, the RWA scheme and power control information at the next moment are used as the output of the traffic prediction controller; t+1 When it is 1, the prediction result is False t+1 as the output of the flow prediction controller; Among them, the RWA solution matrix RWA at the next moment t+1 Includes multiple elements, each element RWA t+1 (s, d) is the RWA scheme for the predicted data packet between the source node s and the target node d; the power control information matrix P at the next moment t+1 Includes multiple elements, each element P t+1 (s, d) is the power control information of the predicted data packet between the source node s and the destination node d.

3. The traffic prediction controller for on-chip optical network according to claim 1, characterized in that: The traffic predictor is constructed based on the LSTM neural network structure, and the traffic predictor includes a normalization layer, an input layer, an LSTM layer, a random inactivation layer, a fully connected layer, a regression layer, and a denormalization layer; The traffic predictor is used to input the historical traffic matrix and the traffic matrix at the current moment after the training is completed, and output the predicted traffic matrix at the next moment through the normalization layer, input layer, LSTM layer, random inactivation layer, fully connected layer, regression layer and denormalization layer.

4. The traffic prediction controller for on-chip optical network according to claim 1, characterized in that: The optical layer includes a plurality of optical nodes, the optical nodes are connected together through waveguides, each optical node is composed of an optical router and an OE interface, the optical router is composed of a microring resonator MR and a waveguide, and the MRs are connected through the waveguide; the electrical nodes in the electrical layer are connected together through electrical wires, each electrical node is composed of an electrical router and a processing unit, and the electrical router and the processing unit are connected together through electrical wires; The traffic prediction controller is used to send the prediction judgment result or RWA solution and power control information at the next moment to the electrical router; The electrical router is used to make routing path reservation, wavelength allocation and power allocation of the optical signal according to the prediction result of the next moment or the RWA scheme and power control information, and send MR control information to the optical router to control the working state of the corresponding MR, so as to select the specified light source wavelength, current and optical path to complete the communication transmission of the optical layer.

5. The traffic prediction controller for on-chip optical network according to claim 4, characterized in that: The electrical router includes a routing control module, an electrical cross-switch module, an MR and light source control module, and an error detector, the output of the flow prediction controller is connected to the input of the routing control module, the input of the MR and light source control module, and the input of the flow prediction controller is connected to the input port of the electrical cross-switch module; The electrical crossbar switch module has six input ports and six output ports, and the port information of the six input ports and six output ports of the electrical crossbar switch module is cached in a buffer corresponding to the port; Buffer, used to extract the number of all data packets in the local data flow at the current moment and the number of data packets that are about to leave from the current moment to the next moment, and send the number of all data packets and the number of data packets that are about to leave to the traffic prediction controller and the error detector; The error detector is used to compare the number of all data packets at the current moment with the predicted number of data packets and predicted path information in the RWA solution at the current moment, and output the detection result to the routing control module; The traffic prediction controller is used to predict the prediction result of the next moment, or the RWA scheme and corresponding power control information of the next moment based on the traffic predictor according to the number of all data packets at the current moment and the number of data packets that are about to leave from the current moment to the next moment, and send the prediction result of the next moment, or the RWA scheme and corresponding power control information of the next moment to the routing control module and the MR and light source control module; The MR and light source control module is used to output the current of the specified light source according to the prediction result of the next moment or the RWA solution and the corresponding power control information, and send the current of the specified light source to the OE interface; The routing control module is used to send the routing decision at the current moment to the electrical crossbar switch module according to the detection result of the error detector and the prediction judgment result at the next moment or the RWA scheme at the next moment and the corresponding power control information; The electrical crossbar switch module is used to configure the connection between the corresponding input port and the output port according to the routing decision at the current moment, so as to perform routing path reservation, wavelength allocation and power allocation, and cooperate with the generated MR control information to configure the optical layer to complete the communication transmission of the optical signal.

6. The traffic prediction controller for on-chip optical network according to claim 5, characterized in that: The OE interface includes a light source driver and a receiver; The MR and light source control module is used to output the current of the specified light source according to the prediction result of the next moment or the RWA scheme and the corresponding power control information, and send the current of the specified light source to the light source driver of the OE interface; The OE interface and the optical router are used to control the use of the specified light source wavelength and current according to the current control of the specified light source, and to control the working state of the MR according to the MR control information, and to control the transmission direction of the optical signal in the optical router; The receiver is used to convert the optical signal and transmit the optical signal to the electrical router through the vertical input port of the electrical crossbar switch module; The electrical routers communicate with each other on-chip electrical routers according to the link selection information in the RWA solution, and transmit the MR control information to the optical routers through TSV so as to configure and reserve optical layer resources.

7. The traffic prediction controller for on-chip optical network according to claim 5, characterized in that: The error detector is used to The number of all packets at the current moment traffic t (s, d) and the current RWA solution RWA t The predicted number of packets in (s, d) is compared with the predicted path information. If the number of packets is inconsistent, the source and destination node information is inconsistent, or the RWA solution RWA is not received t (s, d), it is determined that the prediction result at the current moment is wrong, and the error is set to error t (s, d) is set to 1, otherwise, the error bit is set to error t (s, d) is set to 0; the result of the error bit being 0 or 1 is used as the detection result, and the detection result is sent to the routing control module; Among them, error t (s, d) indicates whether the prediction of the data packet between the source node s and the target node d is wrong at the current time t. A value of 1 indicates an incorrect prediction, and a value of 0 indicates a correct prediction. The routing control module is used to When the detection result is 1, the routing decision of XY at the current moment is sent to the electrical crossbar switch module, the first wavelength among the idle wavelengths is selected as the communication wavelength, and the light source power under the worst case is configured. The light source power under the worst case is the transmitted optical power of the path corresponding to the lowest OSNR in the network; When the detection result is 0, the current RWA solution RWA is sent to the electrical crossbar switch module. t (s, d) and the corresponding power control information P t The routing decision of (s,d).

8. The traffic prediction controller for on-chip optical network according to claim 5, characterized in that: The six input ports and six output ports of the electrical crossbar switch module are respectively input and output ports in four directions of east, south, west and north, local input ports and local output ports, vertical input ports and vertical output ports; Among them, the input and output ports in the four directions of east, south, west and north are responsible for communicating with other on-chip electrical routers at the same layer. The local input port and local output port are responsible for connecting and communicating with the processing unit of the local node. The vertical input port is connected to the receiver of the OE interface through TSV; the vertical output port is connected to the light source driver of the OE interface through TSV.

9. An on-chip optical network, characterized in that: A traffic prediction controller for an on-chip optical network comprising any one of claims 1 to 8.

Images