A resource scheduling method for an access network system

Abstract

The present invention provides a resource scheduling method and a prediction model for an access network system. The method includes: monitoring the number of enabled container replicas corresponding to the currently enabled wireless transceiver devices in the coverage area and the historical CPU usage time series of the corresponding control modules within a period of time; using the prediction model to predict the predicted value of the CPU usage required by the control module within a preset future time period according to the historical CPU usage sequence; and scheduling the operating resources required by the control modules corresponding to the wireless transceiver devices according to the predicted value of the CPU usage, the preset CPU usage warning value of the control module, and the number of enabled container replicas. The technical solution of the present invention predicts the load of the access network in advance and schedules the required operating resources for the control modules corresponding to the wireless transceiver devices enabled in all base stations in the access network in the cloud according to the predicted load amount, so that the access network can perform dynamic adjustment of resources according to real-time requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile communications, and more specifically, to a resource scheduling method for an access network system. Background Art

[0002] During the accelerated deployment of 5G, in addition to the challenges brought by the diversity of services to 5G, the tidal effect of communication services also causes the data volume and load corresponding to services to change over time and location.

[0003] The tidal effect is prevalent in areas such as office areas, schools, and residential areas. At different times and in different areas, the data services and load pressure in the cell are different. For example, the load of the base station in the office area during the day is higher than that in the residential area, and the load of the base station in the residential area late at night is higher than that in the office area, etc. As a functional entity of the access network, when facing the tidal effect, if the access network does not make corresponding adaptive adjustments, it will cause excessive load pressure at the peak and a large amount of resource waste at the trough. Therefore, in order to make reasonable use of the resources of the access network system, it is necessary to dynamically adjust the resources of the access network on demand.

[0004] Currently, the method for dealing with the tidal effect in the communication network is to put some base stations in the access network into sleep or shutdown according to the change of traffic load. However, this method is operated in a fixed time period or manually, and it cannot accurately perceive the actual situation of the tidal effect and cannot respond agilely to the change of traffic demand. Summary of the Invention

[0005] Therefore, the purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a resource scheduling method and a prediction model for an access network system.

[0006] The purpose of the present invention is achieved by the following technical solutions:

[0007] According to a first aspect of the present invention, there is provided a resource scheduling method for an access network system. The access network system includes a resource scheduling platform and a plurality of base stations. For the same coverage area, each base station is configured with multiple sets of wireless transceiver devices that can be enabled on demand. The wireless transceiver devices are controlled by a control module connected thereto and jointly establish a cell for serving users with the control module. The resource scheduling platform schedules the required operating resources for the control modules corresponding to the enabled wireless transceiver devices in all base stations in the cloud and deploys the control modules in the form of container replicas. The operating resources include CPU resources. The method includes scheduling the required operating resources for the control modules corresponding to the wireless transceiver devices configured for each coverage area in the following manner at a preset frequency:

[0008] Monitor the number of enabled container replicas corresponding to the currently enabled wireless transceiver devices in the coverage area and the historical CPU usage time series of their corresponding control modules over a period of time;

[0009] Use a pre-trained prediction model to predict the predicted CPU usage value required by the control module in a preset future time period based on the historical CPU usage sequence;

[0010] Obtain the expected number of container replicas to be enabled in the future based on the predicted CPU usage value, the preset CPU usage warning value of the control module, and the number of enabled container replicas, and schedule the operating resources required by the control module corresponding to the wireless transceiver device according to the expected number and the enabled number. Among them, when the expected number of container replicas is greater than the enabled number, increase the operating resources required by the corresponding control module by increasing the number of enabled container replicas; when the expected number of container replicas is less than the enabled number, reduce the operating resources required by the corresponding control module by reducing the number of enabled container replicas.

[0011] In some embodiments of the present invention, the method includes:

[0012] When the expected number of container replicas is greater than the enabled number, enable a new wireless transceiver device, establish a connection between the increased container replicas and the new wireless transceiver device, and generate a new cell for serving users;

[0013] When the expected number of container replicas is less than the enabled number, deactivate the cell controlled by the control module corresponding to the container replicas exceeding the expected number, recycle the operating resources of the corresponding container replicas, and turn off the corresponding wireless transceiver device.

[0014] In some embodiments of the present invention, the access network system includes base stations in at least one of the following architecture forms:

[0015] The base station is in an architecture form including a baseband processing unit and a radio frequency transceiver unit. The wireless transceiver device is a radio frequency transceiver unit, and the control module is a module that implements the functions corresponding to the baseband processing unit;

[0016] The base station is in an architecture form including a central unit, a distributed unit, and an active antenna unit. The wireless transceiver device is an active antenna unit, and the control module is a module that implements the functions corresponding to the central unit and the distributed unit. Moreover, the central unit and the distributed unit are deployed in the form of container replicas as different types of control modules; and

[0017] The base station has an architecture including a control plane central unit, a user plane central unit, a distributed unit, and an active antenna unit. The wireless transceiver device is set as the active antenna unit. The control module is a module that implements the functions corresponding to the control plane central unit, the user plane central unit, and the distributed unit. And the control plane central unit, the user plane central unit, and the distributed unit are deployed in the form of container replicas as different types of control modules.

[0018] In some embodiments of the present invention, for the control modules corresponding to the baseband processing unit, the distributed unit, the central unit, or the user plane central unit, the expected number of container replicas corresponding to each type of control module is determined respectively. The expected number is obtained by taking the ceiling of the product of the number of enabled containers and the proportionality coefficient; wherein, the proportionality coefficient is the ratio of the predicted CPU usage rate value to the CPU usage rate warning value.

[0019] In some embodiments of the present invention, the expected number is determined in the following manner:

[0020]

[0021] Wherein, R′ expect represents the expected number of container replicas to be enabled; R cur represents the number of enabled container replicas; CPU pred represents the predicted CPU usage rate value, and CPU threshold represents the CPU usage rate warning value, and ceil represents the ceiling function.

[0022] In some embodiments of the present invention, when the expected number is obtained by using the ceiling function, compare the expected number with the expected number threshold set in advance for the container replicas corresponding to this control module, and when the expected number is greater than the expected number threshold, select the expected number threshold as the final expected number.

[0023] In some embodiments of the present invention, when reducing the number of enabled container replicas, the running resources of the corresponding container replicas are recycled based on the order of the priorities of the enabled container replicas from low to high.

[0024] In some embodiments of the present invention, the prediction model is an ARIMA-LSTM hybrid model, and the ARIMA-LSTM hybrid model is trained with the historical CPU usage rate sequence that has passed the stationarity test as the input and the CPU usage rate prediction value sequence as the output.

[0025] According to the second aspect of the present invention, there is provided a prediction model for predicting the required running resources of a control module in the method described in the first aspect, characterized in that the prediction model includes:

[0026] An ARIMA model is used to output a first CPU usage prediction value sequence and a residual sequence based on a historical CPU usage rate sequence that passes the stationarity test. Among them, the residual value in the residual sequence is the difference between the first CPU usage prediction value in the first CPU usage prediction value sequence and the corresponding CPU usage value in the input CPU usage rate sequence;

[0027] An LSTM model is used to output a second CPU usage prediction value sequence based on the residual sequence. An attention mechanism layer is added after the output gate of the LSTM model;

[0028] An output module is used to add the first CPU usage prediction value sequence and the second CPU usage prediction value sequence, and output a CPU usage prediction value sequence including CPU usage prediction values within a preset future time period.

[0029] According to the third aspect of the present invention, there is provided an electronic device, characterized in that it includes:

[0030] A memory, where the memory is used to store executable instructions;

[0031] The one or more processors are configured to implement the steps of the method described in the first aspect by executing the executable instructions.

[0032] Compared with the prior art, the advantages of the present invention are:

[0033] On the one hand, the base station is cloudified and multiple sets of wireless transceiver devices enabled on demand are configured in the same coverage area of the same base station, so that the elastic scaling technology in the cloud native mechanism can be used to flexibly adjust the operating resources required by the base station and the number of enabled wireless transceiver devices; on the other hand, the resources required by the base station are predicted in advance through a prediction model, so that the access network system can perform scaling scheduling in advance according to the predicted resources to ensure the timeliness of resource scheduling. Description of the Drawings

[0034] The following further describes the embodiments of the present invention with reference to the drawings, where:

[0035] Figure 1 It is a schematic diagram of a resource scheduling process of an access network system provided according to an embodiment of the present invention;

[0036] Figure 2 It is a schematic diagram of a prediction process of a prediction model provided according to an embodiment of the present invention;

[0037] Figure 3 It is a schematic diagram of a training process of an ARIMA model provided according to an embodiment of the present invention;

[0038] Figure 4 Schematic diagram of the training process of the LSTM model provided according to an embodiment of the present invention;

[0039] Figure 5 Another resource scheduling process schematic diagram of the access network system provided according to an embodiment of the present invention;

[0040] Figure 6 Schematic diagram of the access network system architecture provided according to an embodiment of the present invention. Detailed implementation manners

[0041] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below through specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0042] As mentioned in the background art section, the existing communication network cannot accurately sense the occurrence of the tidal effect, nor can it perform dynamic and flexible adjustment of resources according to real-time requirements. To solve the above technical problems, in the embodiments of the present application, the base stations in the access network system are cloudified (that is: the control modules corresponding to the radio transceiver devices in the base stations are deployed in the cloud in the form of container replicas, and at the same time, operation resources are scheduled for the control modules corresponding to the radio transceiver devices enabled in the base stations in the cloud), and multiple sets of radio transceiver devices enabled on demand are configured in the same coverage area of the same base station, so that the access network system can utilize the elastic scaling technology in the cloud native mechanism to flexibly adjust the operation resources required by the base station and the number of enabled radio transceiver devices. Generally speaking, within a predetermined frequency, the present invention uses a prediction model to sense the future resource requirements of each coverage area, dynamically scale and adjust the resources required for each coverage area, allocate resources in advance, and better meet the resources required by future service users in each coverage area. Among them, the present invention predicts the future required resources of each coverage area of the base station through a prediction model trained with the historical CPU usage rate sequence of each coverage area of the base station as input and the predicted CPU usage rate sequence as output, better senses the resource requirements brought by the load tidal effect, and faces the changes in resource requirements more agilely, so that the access network system can perform corresponding scaling operations when the predicted required resources exceed or are lower than the pre-set required resource threshold, so as to schedule the resources required by the coverage area by performing scaling operations in advance, thereby ensuring the timeliness of resource scheduling.

[0043] To better understand the present invention, the training of the prediction model will be introduced from several aspects such as model structure, training samples, and training methods below.

[0044] I. Model structure

[0045] Since the prediction model predicts the traffic load of the access network including the traffic load at time t2 / time period (t2 = t + t1, where t represents the current time / time period and t1 represents the operation time required for the scaling operation) based on the historical traffic load time series, to implement the above technical solution, a time series model needs to be selected. According to an embodiment of the present invention, the prediction model can be implemented using an ARIMA model (Autoregressive Integrated Moving Average model) or an LSTM model (Long Short-Term Memory, LSTM). Further, considering that in the access network system, the level of traffic load is mainly reflected by the CPU usage rate of the control module, and due to the tidal effect and its own fluctuations in the access network system, the CPU usage rate of the control module includes both a linear part and a nonlinear part, making the ARIMA model (the ARIMA model has good linear fitting ability and relatively poor prediction results for non-linear and non-stationary data) or the LSTM model (on the contrary to the ARIMA model, it has good non-linear fitting ability and relatively poor prediction results for linear and stationary data) unable to accurately predict the CPU usage rate sequence that includes both a linear part and a nonlinear part in this application. Therefore, to improve the accuracy of predicting the CPU usage rate sequence, according to an embodiment of the present invention, the ARIMA model with good fitting ability for the linear part and the LSTM / BiLSTM model with good fitting ability for the nonlinear part are combined to form an ARIMA-LSTM hybrid model or an ARIMA-BiLSTM hybrid model to predict the CPU usage rate sequence that includes both a linear part and a nonlinear part. Since the difference between the ARIMA-LSTM hybrid model and the ARIMA-BiLSTM hybrid model is only that the long short-term memory network in the ARIMA-LSTM hybrid model is a unidirectional long short-term memory network, and the long short-term memory network in the ARIMA-BiLSTM hybrid model is a bidirectional long short-term memory network, and their structures are similar and the training processes are the same, therefore, the following takes the ARIMA-LSTM hybrid model as an example to illustrate the prediction model:

[0046] The ARIMA-LSTM hybrid model provided by the embodiment of the present application, as Figure 2 shown, includes:

[0047] The ARIMA model is used to output the first CPU usage prediction value sequence M′(t) and the residual sequence N(t) according to the historical CPU usage rate sequence F(t) that passes the stationarity test. Among them, the residual value in the residual sequence N(t) is the difference between the first CPU usage prediction value in the first CPU usage prediction value sequence M′(t) and the corresponding CPU usage value in the input CPU usage rate sequence F(t).

[0048] The LSTM model is used to output the second CPU usage prediction value sequence N′(t) according to the residual sequence N(t).

[0049] The output module is used to add the first CPU usage prediction value sequence M′(t) and the second CPU usage prediction value sequence, and output the CPU usage prediction value sequence F′(t) including the CPU usage prediction values within a preset future time period, that is: F′(t) = M′(t) + N′(t).

[0050] Furthermore, considering that for time series data, the attention mechanism layer can make the model pay more attention to the time points that have a greater impact on the result (the time points with a greater impact refer to the time points with the highest attention scores, where the attention scores are calculated by the attention mechanism layer based on the hidden states of the previous n historical moments and the current cell state. The higher the attention score, the higher the importance of the hidden state of that historical moment to the input at the current moment, and the greater its impact on the result). Therefore, in order to further improve the prediction accuracy of the prediction model, according to an embodiment of the present invention, one or more attention mechanism layers are added after the output gate of the LSTM model. By combining the attention mechanism layer and the LSTM model, when calculating the output state at the current moment (T moment), the hidden states of the previous n moments (T-1, T-2,..., T-n moments) can be comprehensively considered, and different attention coefficients are assigned to the hidden states of the historical sequence based on the attention mechanism layer, so that the generation probability of each item in the output sequence of the LSTM model is affected by the hidden states of multiple input historical sequences, thereby improving the prediction accuracy.

[0051] II. Training Samples

[0052] After the network structure of the prediction model is set up, in order to be able to use the prediction model to predict the CPU usage prediction values within a preset future time period according to the historical CPU usage rate sequence that passes the stationarity test, it is also necessary to construct a training set and use the constructed training set to train the prediction model, so as to obtain the optimal parameters of the prediction model through continuous iterative training, so that the prediction model can output the CPU usage prediction values within a preset future time period according to the input historical CPU usage rate sequence.

[0053] According to an embodiment of the present invention, the training set includes a first training set and a second training set. Among them, the first training set is used to construct and train an ARIMA model, which includes a plurality of first training samples, and each first training sample includes a CPU usage rate sequence. According to an embodiment of the present invention, the CPU usage rate sequence in the first training set can be obtained by dividing the historical CPU usage rate sequence of the access network according to a specified sequence length. For example, a historical CPU usage rate sequence with a total sequence length of 50 is divided into 5 first training samples with a sequence length of 10, that is, the 5 first training samples are {F1, F2, …, F 10}, {F 11 , F 12 , …, F 20}, {F 21 , F 22 , …, F 30}, {F 31 , F 32 , …, F 40} and {F 41 , F 42 , …, F 40}; among them, in order to increase the number of first training samples as much as possible, preferably, a sliding window strategy with a window size of D and a sliding step of d can be adopted to divide the historical CPU usage rate sequence of the access network. For example, a historical CPU usage rate sequence with a total sequence length of 50 is divided into 41 first training samples with a window size of 10 and a sliding step of 1, then the 41 first training samples are {F1, F2, …, F 10}, {F2, F3, …, F 11}, {F3, F2, …, F 12}, …… {F 41 , F 42 , …, F 50}. It should be noted that the above settings of the sequence length are only for illustration to intuitively understand the technical solution of the present invention. In actual applications, it can be set as needed according to the requirements of the application site and the specific situation of the device.

[0054] In addition, it is also worth noting that when obtaining the CPU usage rate sequence, the sampling interval of the CPU usage rate can be set as needed. For example, the sampling interval can be set to 1s, 10s, 1min, 5min, 1h, 2h or others. According to an embodiment of the present invention, the inventor found through analyzing the historical CPU usage rate sequence of the access network that within one hour, the CPU usage rate fluctuates little and can be approximately considered to remain unchanged. Therefore, the sampling interval of the CPU usage rate is set to 1h.

[0055] The second training set is used to train the LSTM model, which includes a plurality of second training samples, and each second training sample includes a residual sequence; wherein, the residual sequence in each second training sample is the difference between the CPU usage rate sequence input to the trained ARIMA model and the corresponding predicted CPU usage rate sequence output. Among them, the CPU usage rate sequence input to the trained ARIMA model may be the same as or different from the CPU usage rate sequence in the first training set. To reduce the time for constructing the training set, preferably, the CPU usage rate sequence input to the trained ARIMA model is the same as the CPU usage rate sequence in the first training set.

[0056] III. Training Method

[0057] After the network structure and training samples of the prediction model are both constructed, in order to use the prediction model to accurately predict the CPU usage rate prediction value based on the historical CPU usage rate sequence, it is also necessary to use the training set to train the prediction model to obtain the optimal parameters of the prediction model. The following explains the training process of the prediction model:

[0058] The training process of the prediction model includes the training process of the ARIMA model and the training process of the LSTM model. Among them, the training process of the ARIMA model includes two parts: the construction of the ARIMA model and the training of the ARIMA model, as Figure 3 shown. The following separately explains the construction of the ARIMA model and the training process of the ARIMA model:

[0059] 1. Training of the ARIMA Model

[0060] 1.1 Construction of the ARIMA Model

[0061] Perform a stationarity test on the CPU usage rate sequence in the first training set to obtain the first test result. When the first test result is positive, determine the numerical range of the autoregressive term number and the numerical range of the moving average term number according to the autocorrelation plot and partial autocorrelation plot of the CPU usage rate data sequence respectively, and establish multiple ARIMA models according to each value in the numerical range of the autoregressive term number and each value in the numerical range of the moving average term number, and apply the Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) to determine the p, d, q parameters of the ARIMA model from multiple ARIMA models, where p represents the autoregressive term, d represents the difference term, and q represents the moving average term. When the first test result is negative, first perform a differencing process on the CPU usage rate data sequence to obtain the differenced CPU usage rate data sequence. Then perform a stationarity test on the currently differenced CPU usage rate data sequence to obtain the second test result. If the second test result is positive, return to the step of determining the numerical range of the autoregressive term number and the numerical range of the moving average term number according to the autocorrelation plot and partial autocorrelation plot of the CPU usage rate data sequence respectively. If the second test result is negative, then perform a differencing process on the differenced CPU usage rate data sequence again to obtain the multi-order differenced CPU usage rate data sequence, and set the multi-order differenced CPU usage rate data sequence as the currently differenced CPU usage rate data sequence, and return to perform a stationarity test on the currently differenced CPU usage rate data sequence until the second test result is positive. According to an embodiment of the present invention, determining the numerical range of the autoregressive term number and the numerical range of the moving average term number according to the autocorrelation plot and partial autocorrelation plot of the CPU usage rate data sequence respectively includes the following steps: First, calculate the autocorrelation coefficient and partial autocorrelation coefficient according to the autocorrelation plot and partial autocorrelation plot respectively, and then determine the numerical range of the autoregressive term number according to the lag order corresponding to when the autocorrelation coefficient first appears truncated, and determine the numerical range of the moving average term number according to the lag order corresponding to when the partial autocorrelation coefficient first appears a decaying trend.

[0062] 1.2. Training of the ARIMA model

[0063] After obtaining the p, d, q parameters of the ARIMA model, it is also necessary to use the first training set to train the ARIMA model to determine the coefficients of the ARIMA model. According to an embodiment of the present invention, when training the ARIMA model, first, a stationarity test needs to be performed on the CPU usage rate sequence in the first training set, and a differencing process is performed on the CPU usage rate sequence that does not meet the stationarity check to obtain a CPU usage rate sequence that passes the stationarity check. Then, the CPU usage rate sequence that passes the stationarity check is input into the ARIMA model with the determined p, d, q parameters for training, so as to obtain a trained ARIMA model, and further, the CPU usage rate prediction value sequence of the linear part can be predicted according to the input historical CPU usage rate sequence.

[0064] 2. Training of the LSTM model

[0065] According to an embodiment of the present invention, as Figure 4 shown, when training the LSTM model, the LSTM model outputs a predicted value sequence of the CPU usage rate of the non - linear part according to the residual sequence in the second training set, and uses the self - supervised sequence obtained by shifting the residual sequence forward by one time point and the loss value calculated from the predicted value sequence of the CPU usage rate of the non - linear part to update the parameters of the LSTM model until the loss function converges or the number of iterations reaches the iteration number threshold, thereby obtaining the optimal LSTM model. According to an embodiment of the present invention, the Mean Squared Error (MSE) function is used as the loss function for training, and the Adam algorithm is used as the optimizer, and the training target is until the loss function converges or the number of iterations reaches the iteration number threshold.

[0066] The prediction model trained through the above - mentioned embodiment can predict the predicted value of the CPU usage rate required within a preset future time period based on the historical CPU usage rate sequence.

[0067] As mentioned before, in order to solve the technical problems mentioned in the background art part, the present invention proposes a technical solution of cloudifying the base station and introducing a prediction model, so as to utilize the elastic scaling technology in the cloud native mechanism to flexibly adjust the resources required for each coverage area of the base station.

[0068] According to an embodiment of the present invention, the present invention provides a resource scheduling method for an access network system. Among them, the access network system includes a resource scheduling platform and multiple base stations. For the same base station, for the same coverage area (the same coverage area can be the area covered by all sectors of the base station (coarse - grained), or the area covered by a certain sector of the base station (fine - grained). In specific implementation, it can be selected as needed, and the embodiments of the present application do not impose any restrictions on it), multiple sets of wireless transceiver devices enabled on demand are configured. The wireless transceiver devices are controlled by the control modules connected to them and jointly establish a cell for serving users with the control modules. The resource scheduling platform schedules the required operating resources for the control modules corresponding to the enabled wireless transceiver devices in all base stations in the cloud and deploys the control modules in the form of container replicas. The operating resources include CPU resources. The method includes scheduling the required operating resources and / or deploying the control modules for the control modules corresponding to the wireless transceiver devices configured for each coverage area at a preset frequency by the following steps, as Figure 1 shown, including:

[0069] S1: Monitor the number of enabled container replicas corresponding to the currently enabled wireless transceiver devices in the coverage area and the historical CPU usage time series of their corresponding control modules over a period of time. Since this solution designs the base station in a cloudified manner, and resource monitoring by the cloud service platform (resource management platform) is an existing technology, the specific implementation process of this solution will not be described herein. For example, in the kubernetes system, cAdvisor is used to collect the usage of resources such as CPU, memory, and network in container replicas (Pods). In addition, considering the existence of a load balancing mechanism in the cloud service platform, in one embodiment of the present invention, for an enabled wireless transceiver device, it is only necessary to monitor the historical CPU usage time series of its corresponding control module over a period of time.

[0070] S2: Use the prediction model provided in the foregoing embodiment to predict the predicted value of the CPU usage required by the control module in a future preset time period based on the historical CPU usage sequence.

[0071] S3: Obtain the expected number of container replicas to be enabled in the future based on the predicted value of CPU usage, the preset CPU usage warning value of the control module, and the number of enabled container replicas, and schedule the operating resources required by the control module corresponding to the wireless transceiver device according to the expected number and the enabled number.

[0072] According to an embodiment of the present invention, when scheduling the operating resources required by the control module corresponding to the wireless transceiver device according to the expected number and the enabled number, if the expected number of container replicas is greater than the enabled number, increase the operating resources required by the corresponding control module by increasing the number of enabled container replicas; if the expected number of container replicas is less than the enabled number, reduce the operating resources required by the corresponding control module by reducing the number of enabled container replicas. At the same time, to ensure the communication quality of the access network system, preferably, when reducing the number of enabled container replicas, recycle the operating resources of the corresponding container replicas in the order of increasing priority of the enabled container replicas. Among them, the priority of the enabled container replicas can be implemented based on the scoring mechanism of the cloud platform itself (such as the scoring mechanism in K8s). The higher the score of the container replica, the higher its priority, and the lower the score of the container replica, the lower its priority. After obtaining the scores of each container replica, recycle the operating resources of the corresponding container replica by deleting the container replicas that exceed the expected number and have the lowest scores.

[0073] According to an embodiment of the present invention, schedule the operating resources required by the control module corresponding to the wireless transceiver device in advance by a predetermined time according to the expected number and the enabled number and deploy the control module. The advance predetermined time can be set according to the needs of the implementer, and the present invention does not impose any restrictions thereon. For example, it is 1 minute, 2 minutes, 200 seconds, etc. in advance.

[0074] According to an embodiment of the present invention, the expected quantity is obtained by taking the ceiling of the product of the number of enabled containers and the proportionality coefficient; wherein, the proportionality coefficient is the ratio of the CPU usage prediction value to the CPU usage warning value. According to an embodiment of the present invention, the expected quantity is determined in the following manner:

[0075]

[0076] wherein, R′ expect represents the expected quantity of container replicas to be enabled; R cur represents the number of enabled container replicas; CPU pred represents the CPU usage prediction value, and CPU threshold represents the CPU usage warning value, and ceil represents the ceiling function.

[0077] Furthermore, considering that although the above-mentioned embodiment adjusts the corresponding operating resources according to the traffic load of the base station in the cloud, since the hardware device (wireless transceiver device) of the base station has not been correspondingly adjusted, there is still a situation that does not match the traffic load of the base station. Based on this, in order to better cope with the tidal effect of the communication system, according to an embodiment of the present invention, when the expected quantity of container replicas is greater than the enabled quantity, a new wireless transceiver device is enabled, and a connection is established between the increased container replicas and the new wireless transceiver device to generate a new cell for serving users; when the expected quantity of container replicas is less than the enabled quantity, the cell controlled by the control module corresponding to the container replicas exceeding the expected quantity is deactivated, the operating resources of the corresponding container replicas are recycled, and the corresponding wireless transceiver device is turned off.

[0078] Furthermore, considering that there are currently three different architecture forms of base stations, when performing cloudification design of the base station, different architecture forms of base stations have different deployment methods and resource allocation methods. The following respectively describes the deployment methods and resource scheduling processes of different architecture base stations during cloudification design:

[0079] Architecture 1: The deployment method of the base station with the architecture form of the baseband processing unit (BBU) and the radio frequency transceiver unit (RRU) is as follows:

[0080] The wireless transceiver device is set as the radio frequency transceiver unit, and the control module is set as the module that realizes the functions corresponding to the baseband processing unit and is deployed in the resource scheduling platform in the form of container replicas;

[0081] The resource scheduling process is as follows:

[0082] When the desired number of container replicas is greater than the enabled number, on the one hand, the enabled number of container replicas is increased to increase the operating resources required by the corresponding control module. On the other hand, a new wireless transceiver is enabled, and a connection is established between the increased container replicas and the new wireless transceiver to generate a new cell for serving users. When the desired number of container replicas is less than the enabled number, on the one hand, the number of enabled container replicas is reduced to reduce the operating resources required by the corresponding control module. On the other hand, the cell controlled by the control module corresponding to the container replicas exceeding the desired number is deactivated, the operating resources of the corresponding container replicas are recycled, and the corresponding wireless transceiver is turned off.

[0083] Architecture 2: The architecture form of a Central Unit (CU), a Distributed Unit (DU), and an Active Antenna Unit (AAU)

[0084] The deployment method of the base station is as follows:

[0085] The wireless transceiver is set as an active antenna unit, the control module is set as a module that realizes the functions corresponding to the central unit and the distributed unit, and the central unit and the distributed unit are deployed in the form of container replicas as different types of control modules. Among them, one central unit can be connected to multiple distributed units, and each distributed unit corresponds to a wireless transceiver.

[0086] The resource scheduling process is as follows:

[0087] When the desired number of container replicas of the control modules corresponding to the central unit and the distributed unit is greater than the enabled number, on the one hand, the enabled number of container replicas of this type of control module is increased to increase the operating resources required by the corresponding control module. On the other hand, if the increased container replicas are of the control module corresponding to the distributed unit, a new wireless transceiver is also enabled, and connections are established between the distributed unit and the increased container replicas and between the increased container replicas and the newly enabled wireless transceiver to generate a new cell for serving users.

[0088] When the desired number of container replicas of the control modules corresponding to the central unit and the distributed unit is less than the enabled number, on the one hand, the number of enabled container replicas of this type of control module is reduced to reduce the operating resources required by the corresponding control module. On the other hand, if the reduced container replicas are of the control module corresponding to the distributed unit, the operating resources of the container replicas of the control module corresponding to the distributed unit exceeding the desired number are also recycled, and the wireless transceiver and the distributed unit connected to them are turned off to deactivate the cell controlled by the control module corresponding to the distributed unit exceeding the desired number.

[0089] Architecture 3: The architecture form of the control plane central unit (CUCP), user plane central unit (CUUP), distributed unit (DU), and active antenna unit (AAU)

[0090] The deployment method of the base station is as follows:

[0091] The wireless transceiver device is set as the active antenna unit, and the control module is set as the module that realizes the functions corresponding to the control plane central unit, user plane central unit, and distributed unit. Moreover, the control plane central unit, user plane central unit, and distributed unit are deployed in the form of container replicas as different types of control modules; among them, one control plane central unit can be connected to multiple user plane central units, each user plane central unit is connected to one distributed unit, and one distributed unit corresponds to one wireless transceiver device.

[0092] The resource scheduling process is as follows:

[0093] When the desired number of container replicas of the control module corresponding to the user plane central unit or distributed unit is greater than the enabled number, on the one hand, the enabled number of container replicas of the control module of this type is increased to increase the operating resources required by the corresponding control module; on the other hand, if the increased container replicas are those of the control module corresponding to the distributed unit, new wireless transceiver devices are also enabled, and connections are established between the distributed unit and the increased container replicas, as well as between the increased container replicas and the newly enabled wireless transceiver devices, to generate new cells for serving users.

[0094] When the desired number of container replicas of the control module corresponding to the user plane central unit or distributed unit is less than the enabled number, on the one hand, the number of enabled container replicas of the control module of this type is reduced to reduce the operating resources required by the corresponding control module; on the other hand, if the reduced container replicas are those of the control module corresponding to the distributed unit, the operating resources of the container replicas of the control module corresponding to the distributed unit that exceed the desired number are also released, and the wireless transceiver devices and distributed units connected to them are shut down to deactivate the cells controlled by the container replicas of the control module corresponding to the distributed unit that exceed the desired number.

[0095] Furthermore, considering that when the base station of the access network system performs an expansion operation, it is necessary to pull the corresponding base station image from the image repository of the resource scheduling platform. Since the base station image contains all the software environments on which the base station depends and occupies a large amount of CPU and memory during operation, in order to avoid the base station expansion from occupying too much system resources, according to an embodiment of the present invention, the access network system also sets a desired number threshold for each type of container replica number of the base station. When the desired number obtained by using the ceiling function is greater than the desired number threshold set for the corresponding container replica, the desired number threshold is selected as the final desired number, and subsequent expansion operations are performed, such asFigure 5 as shown

[0096] It should be noted that when the base station conducts cloudification design, it can be implemented based on different cloud servers. During specific implementation, it can be selected according to actual requirements, and this embodiment does not impose any restrictions on it. For example, according to an embodiment of the present invention, the resource scheduling platform in the access network system can be implemented based on the Kubernetes container cluster management platform, as Figure 6 shown, which includes a master node and multiple access nodes; among them, the master node is deployed with a resource monitoring module, a load prediction module, an elastic scaling module, a replication controller (Replication Controller), and an access control interface (APIServer); the access nodes are deployed with a management and control module and a control module corresponding to enabling the radio transceiver device in the base station. According to an embodiment of the present invention, the control module in the access node is a module for implementing the functions corresponding to the control plane central unit, the user plane central unit, and the distributed unit, and the control plane central unit, the user plane central unit, and the distributed unit are deployed in the form of Pods (containing one or more container replicas) as different types of control modules, and one control plane central unit can be connected to multiple user plane central units, each user plane central unit is connected to a distributed unit, and one distributed unit corresponds to an active antenna. Further, to ensure the processing efficiency of the communication data by the hardware devices on the base station side, preferably, an FPGA acceleration card can also be configured, and the FPGA acceleration card is called by a corresponding control module (such as DU Pod) to process the communication data to further improve the communication quality.

[0097] When the resource scheduling platform is working, it obtains the CPU usage rate sequence of the Pods of the control modules of the corresponding types in the centralized unit or the distributed unit through the resource monitoring module. Secondly, it transmits the obtained CPU usage rate sequence to the load prediction module. The load prediction module uses the prediction model to perform load prediction based on the CPU usage rate sequence provided by the resource monitoring module, and obtains the predicted CPU usage rate prediction value sequence. Then, the elastic scaling module calculates the desired number of replicas according to the CPU usage rate prediction value sequence predicted by the load prediction module, the pre-set CPU usage rate warning value of the Pods of the control modules of the corresponding types in the centralized unit or the distributed unit, and the number of replicas of the Pods of the control modules of the corresponding types in the centralized unit or the distributed unit enabled by the access node. Finally, it sends the desired number of replicas to the replica controller. The replica controller generates corresponding control instructions (scaling-up instructions or scaling-down instructions) according to the quantitative relationship between the desired number of replicas and the enabled number of replicas, and controls the access node to perform corresponding Pod scaling-up operations or Pod scaling-down operations through the access control interface. At the same time, since there is a management and control module in the access node, it can issue the configuration parameters required by the base station for the newly scaled-up Pod replicas, and establish or delete cells and other operations by establishing or deleting the connections between the newly scaled-up Pod replicas or the deleted enabled Pod replicas and the wireless transceiver devices.

[0098] Among them, it should be noted that the function of the management and control module is a prior art, and the embodiments of the present application do not involve improvements thereto, so the specific working principle thereof will not be elaborated in detail. According to an embodiment of the present invention, the functions implemented by the management and control module can also be integrated into the master node, and the corresponding base station startup, cell establishment or deletion and other operations are completed through the control of the master node.

[0099] It should be noted that although the above steps are described in a specific order, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even the order can be changed, as long as the required functions can be achieved.

[0100] The present invention can be a system, a method and / or a computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for enabling a processor to implement various aspects of the present invention are uploaded.

[0101] A computer-readable storage medium can be a tangible device that retains and stores instructions for use by an instruction execution device. A computer-readable storage medium may include, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanically encoded device, such as a punched card or raised structures in grooves having instructions stored thereon, and any suitable combination of the foregoing.

[0102] The embodiments of the present invention have been described above. The above description is exemplary and not exhaustive, and is also not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of the technology in the market, or to enable other ordinary skill in the art in the technical field to understand the embodiments disclosed herein.

Claims

1. A resource scheduling method for an access network system, characterized in that The access network system includes a resource scheduling platform and multiple base stations. The same base station is configured with multiple sets of wireless transceiver devices enabled on demand for the same coverage area. The wireless transceiver devices are controlled by a control module connected to them and jointly establish a cell for serving users with the control module. The resource scheduling platform schedules the required operating resources for the control modules corresponding to the enabled wireless transceiver devices in all base stations in the cloud and deploys the control modules in the form of container replicas. The operating resources include CPU resources. The method flexibly adjusts the required operating resources of the base stations and the number of enabled wireless transceiver devices by using the elastic scaling technology in the cloud native mechanism. The method includes scheduling the required operating resources for the control modules corresponding to the wireless transceiver devices configured for each coverage area in the following manner at a preset frequency: Monitoring the number of enabled container replicas corresponding to the currently enabled wireless transceiver devices in the coverage area and the historical CPU usage time series of the corresponding control module over a period of time; Using a pre-trained prediction model to predict the predicted CPU usage value required by the control module in a preset future time period based on the historical CPU usage sequence; Obtaining the expected number of container replicas to be enabled in the future based on the predicted CPU usage value, the preset CPU usage warning value of the control module, and the number of enabled container replicas, and scheduling the required operating resources for the control modules corresponding to the wireless transceiver devices according to the expected number and the enabled number. Among them, for the control modules corresponding to the baseband processing unit, distribution unit, central unit, or user plane central unit in the base station, the expected number of container replicas corresponding to each type of control module is determined respectively. The expected number is obtained by rounding up the product of the number of enabled containers and the proportionality coefficient. The proportionality coefficient is the ratio of the predicted CPU usage value to the CPU usage warning value. When the expected number of container replicas is greater than the enabled number, the required operating resources of the corresponding control module are increased by increasing the number of enabled container replicas. When the expected number of container replicas is less than the enabled number, the required operating resources of the corresponding control module are reduced by reducing the number of enabled container replicas.

2. The method according to claim 1, wherein The method includes: When the expected number of container replicas is greater than the enabled number, enabling new wireless transceiver devices, establishing a connection between the increased container replicas and the new wireless transceiver devices, and generating a new cell for serving users; When the expected number of container replicas is less than the enabled number, deactivating the cells controlled by the control modules of the container replicas exceeding the expected number, recycling the operating resources of the corresponding container replicas, and shutting down the corresponding wireless transceiver devices.

3. The method according to claim 2, wherein The access network system includes base stations in at least one of the following architecture forms: The base station is in an architecture form including a baseband processing unit and a radio frequency transceiver unit. The wireless transceiver device is the radio frequency transceiver unit, and the control module is a module that implements the functions corresponding to the baseband processing unit; The base station is an architecture comprising a centralized unit, a distributed unit, and an active antenna unit. The wireless transceiver is an active antenna unit. The control module is a module that implements the corresponding functions of the centralized unit and the distributed unit. The centralized unit and the distributed unit are deployed as different types of control modules in the form of container copies. and The base station is an architecture that includes a control plane centralized unit, a user plane centralized unit, a distribution unit, and an active antenna unit. The wireless transceiver is configured as an active antenna unit. The control module is a module that implements the functions corresponding to the control plane centralized unit, the user plane centralized unit, and the distribution unit. The control plane centralized unit, the user plane centralized unit, and the distribution unit are deployed as different types of control modules in the form of container copies.

4. The method according to claim 3, wherein The expected quantity is determined as follows: Among them, represents the expected number of container replicas to be enabled; represents the number of enabled container replicas; represents the predicted value of CPU usage rate, represents the warning value of CPU usage rate, and ceil represents the ceiling function.

5. The method according to claim 3, characterized in that, When the expected number is obtained using the rounding-up function, the expected number is compared with an expected number threshold pre-set for the container copy corresponding to the control module, and when the expected number is greater than the expected number threshold, the expected number threshold is selected as the final expected number.

6. The method according to claim 1, wherein When the number of enabled container replicas is reduced, the running resources of the corresponding container replicas are reclaimed based on the priority of the enabled container replicas from low to high.

7. The method according to any one of claims 1-6, characterized in that The prediction model is an ARIMA-LSTM hybrid model, and the ARIMA-LSTM hybrid model is obtained through training with a historical CPU usage sequence that passes a stationarity test as input and a CPU usage prediction value sequence as output.

8. A system for predicting the operating resources required by a predictive control module in any one of the methods recited in claims 1-7, the system comprising a prediction model, characterized in that, The prediction model includes: An ARIMA model is used to output a first CPU usage prediction value sequence and a residual sequence based on a historical CPU usage sequence that passes a stationarity test, wherein the residual value in the residual sequence is a difference between a first CPU usage prediction value in the first CPU usage prediction value sequence and a corresponding CPU usage value in the input CPU usage sequence; An LSTM model is configured to output a second CPU usage prediction value sequence based on the residual sequence, wherein an attention mechanism layer is added after an output gate of the LSTM model; An output module is used to add the first CPU usage prediction value sequence and the second CPU usage prediction value sequence, and output a CPU usage prediction value sequence including CPU usage prediction values within a future preset time period.

9. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, include: one or more processors; as well as a memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method according to any one of claims 1 to 7 by executing the executable instructions.

Images