RAN-side Slice Resource Allocation Method and System Based on SLA Assurance

Abstract

A method for allocating slice resources on the RAN side based on SLA guarantee, including a data set collection stage and a slice resource prediction stage, wherein: through the OAI system, information interaction data generated during the wireless communication process between the user terminal and the base station in the actual wireless communication scenario is collected, and after preprocessing, it is matched to the data format input to the prediction network; the information interaction data is input into the trained prediction network to obtain a resource allocation scheme. Through the long short-term memory (LSTM) algorithm based on RNN, the present invention can better represent the requirement indicators for SLA guarantee and resource utilization rate, so as to obtain a better resource orchestration scheme for the RAN side network slice when the prediction network training is completed.

Description

Technical Field

[0001] The present invention relates to a technology in the field of wireless communication, specifically a method and system for slice resource allocation on the RAN side based on ensuring SLA. Background Art

[0002] Network slices can be mainly divided into two categories according to the wireless communication system architecture, namely core network (CN) slices and radio access network (RAN) slices. In the field of RAN side slice resource allocation, the resource allocation scheme can be divided into two major categories: soft slices and hard slices. Soft slices rely on the upper layer of the network to execute slice logic, which also leads to its inability to directly control the underlying network physical resources, leaving some technical challenges in terms of security and flexibility. Therefore, in the existing research field, it is more ideal to achieve complete isolation of resources through hard slices. In order to improve the performance of hard slices, the existing work mainly focuses on enriching the application scenarios of slices and improving the prediction accuracy of slice resources. However, the significant differences in indicators such as throughput, latency, and packet loss rate between different types of services pose complex performance requirements for the research and design of the RAN side network slice framework considering the SLA guarantee and resource utilization guarantee of different types of services simultaneously. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, such as the lack of actual scenario verification, the lack of SLA guarantee, and the influence of resource utilization rate on slice resource allocation, that is, when the network resource allocation is insufficient, the poor user experience affects the SLA, etc., the present invention proposes a method and system for slice resource allocation on the RAN side based on ensuring SLA. Through the long short-term memory (LSTM) algorithm based on RNN, it can better represent the required indicators of SLA guarantee and resource utilization rate, so as to obtain a better RAN side network slice resource orchestration scheme at the end of the prediction network training.

[0004] The present invention is realized through the following technical solutions:

[0005] The present invention relates to a method for slice resource allocation on the RAN side based on ensuring SLA, including a data set collection stage and a slice resource prediction stage, wherein: through the OAI system, collect the information interaction data generated during the wireless communication process between the user terminal and the base station in the actual wireless communication scenario, and match the data format input to the prediction network after preprocessing; input the information interaction data into the trained prediction network to obtain a resource allocation scheme.

[0006] The present invention relates to a system for implementing the above method, including: a data set acquisition and preprocessing unit, a slice resource prediction unit, and a slice resource orchestration unit, where: the data set acquisition and preprocessing unit collects corresponding communication data according to the interaction process between the terminal device and the base station in the actual communication system, and performs data preprocessing such as normalization to obtain a data set that meets the input requirements of the prediction network. The slice resource prediction unit performs slice resource prediction processing according to the established prediction network and the redefined loss function convergence algorithm, and obtains the slice resource prediction result at the next moment according to the historical resource allocation data. The slice resource orchestration unit performs the slice resource orchestration at the next moment according to the result of the resource prediction and the mapping rule between resource blocks, and obtains a better slice resource orchestration result.

[0007] Technical effects

[0008] Compared with the prior art, the actual communication data set acquisition accuracy of the present invention reaches the millisecond level, and can customize the output of the required parameters of the physical layer, including the number of resource blocks, channel quality indicators, hybrid automatic repeat request indicators, modulation and coding methods, and other data. The redefined loss function convergence algorithm fully considers SLA guarantee and ensures a high resource utilization rate, and can customize the proportion of the two in the final slice resource orchestration ratio, transforming the traditional resource prediction into service satisfaction prediction, and more fully considering the user experience in the actual communication scenario. Brief description of the drawings

[0009] Figure 1 It is a schematic diagram of the method of the present invention;

[0010] Figure 2 It is a schematic diagram of the device layout for collecting real data in the actual wireless communication system of the embodiment;

[0011] Figure 3 It is a schematic diagram of the loss function of the embodiment;

[0012] Figure 4 It is a schematic diagram of the comparison of the difference between the normalized predicted resource number and the actual required resource number under different algorithms in the embodiment;

[0013] Figure 5 It is a schematic diagram of the comparison of the normalized SLA satisfaction index and the resource utilization rate under different algorithms in the implementation. Detailed implementation manners

[0014] As Figure 1 shown, this embodiment relates to a RAN-side slice resource allocation method based on SLA guarantee, which is implemented in a real communication scenario based on the OAI system as Figure 2 shown. This scenario includes: evolved packet core EPC, evolved Node B, i.e., eNodeB, radio frequency front end RF, and two user equipments UE.

[0015] The described resource allocation method specifically includes:

[0016] Step 1, data collection, specifically including:

[0017] Step 1.1, set the eNodeB to a fixed radio spectrum bandwidth: Use two UEs with different service types, set different UEs under different slices, and correspond the difference in service load to the traffic load of different slices. A higher service load corresponds to a heavier slice transmission traffic load.

[0018] During the system operation, the interaction data between the UE and the eNodeB is saved by a pre-written log information saving program on the eNodeB, so as to obtain the actual slice data set in this wireless communication scenario.

[0019] Step 1.3, perform data cleaning and formatting on the actual slice data set, and convert the slice data into a data format that matches the input of the prediction network.

[0020] Step 2, construct a prediction network based on LSTM and perform offline training, specifically including:

[0021] Step 2.1, adopt LSTM based on the recurrent neural network (RNN) as the main part of the prediction network to explore features such as users' behavior habits on a long time scale. In the definition of the prediction time scale, use the historical interaction information between the UE and the eNodeB to predict the resource orchestration strategy of the slice in the next moment.

[0022] Step 2.2, design a loss function as shown in Figure 3 : To better reflect the consideration of the proposed prediction network for ensuring SLA and improving resource utilization, convert the SLA guarantee index and resource utilization into the loss function of the prediction network, specifically: Where: x is the ratio of the predicted resource number to the required resource number, 0 ≤ x. β (β > 0) represents a penalty term, and its value can be customized to characterize the situation where the eNodeB provides unacceptable services to users, which will result in the inability to meet the SLA guarantee. α (0 < α < 1) is used to describe the user's tolerance for services that cannot fully meet the service quality requirements due to insufficient slice resources. The slope k (k > 0) is a constant, used as a penalty term in the case of over-allocation of resources. However, due to the limitations of the stochastic gradient descent (SGD) method used to train the neural network, a loss function with a constant or step function cannot be used. To solve this problem, add a minimum value ε to the original loss function to not affect the shape of the loss function and avoid the situation where the slope is zero or infinite. The adjusted loss function is: The required loss function is thus defined. When the data is calculated by the prediction network, the output value calculated in the current step is measured by the performance of the loss function to obtain the corresponding loss value.

[0023] Step 2.3, Judgment of the convergence of the prediction network: The training set obtained in Step 1 is used to train the prediction network. The prediction network updates the weight parameters of each layer according to the loss value calculated in each step. When the network has not converged, the output of the network will be used as the network input to re - train the prediction network in the next round, and the update of the prediction network parameters will continue. When the network converges, the training is completed, and the output of the prediction network can provide a basis for slice resource orchestration.

[0024] Step 3, Slice resource orchestration: When the network is input with the slice data of the previous 1 second, the data is calculated by the prediction network, and the resource prediction result of the next 100 milliseconds is output. The slice resource orchestration scheme can be completed according to this result.

[0025] Through specific actual experiments, under the parameter settings shown in Table 1, since there is a lot of redundant information in the data set collected by the actual system, before sending the data into the prediction network, it is necessary to pre - process the data, such as data cleaning and formatting, to convert the slice data into a data format that matches the input of the prediction network. After the above operations, a data set that can be directly input into the prediction network defined by the present invention can be obtained.

[0026] Table 1 Specific components of the OAI system

[0027]

[0028]

[0029] Set the eNodeB to have a 10MHz (50 PRBs) wireless spectrum bandwidth. Two UEs with different service types are used. The service of one UE is set to play a live video with 360p clarity, and the service of the other UE is set to play a live video with 1080p clarity. Each UE is arranged under a different slice. The low - load slice corresponds to the video service with 360p clarity, and the high - load slice corresponds to the video service with 1080p clarity. When the system runs, the interaction data between the UE and the eNodeB will be saved on the eNodeB through a pre - written log information saving program, and a slice data set on the RAN side that meets the requirements can be obtained.

[0030] The prediction network in this embodiment uses LSTM based on a recurrent neural network as the main part of the prediction network to explore features such as users' behavior habits on a long - time scale, and its structural parameters are shown in Table 2.

[0031] Table 2 Neural network architecture

[0032] Network level Type Dimension Input layer 1×4 Hidden layer 1 Conv1D+ReLU 4×48 Hidden layer 2 LSTM 48×48 Hidden layer 3 LSTM 48×48 Hidden layer 4 FC+ReLU 48×32 Hidden layer 5 FC+ReLU 32×4 Hidden layer 6 FC 4×1 Output layer 1×1

[0033] In the definition of the prediction time scale, the interaction information between the UE and the eNodeB in the previous second is utilized, including hybrid automatic repeat request (HARQ), the number of physical resource blocks (RBs), channel quality indicator (CQI), modulation and coding strategy (MCS), transport block size (TBS), etc., to predict the slice resource orchestration strategy within the next 100 milliseconds. A brand-new loss function is designed, and a performance verification experiment of the prediction network is completed based on the measured data set. The results show that for the slice resource orchestration on the RAN side, the network architecture of this embodiment can ensure the resource utilization rate as high as possible on the basis of meeting the SLA guarantee.

[0034] Through Figure 2 the acquisition and verification of the actual wireless scenario data set shown, compared with the prior art, this method improves the resource utilization rate as much as possible on the premise of ensuring the SLA. As Figure 4 shown, affected by the user's mobile behavior pattern and the change of the channel environment, the difference between the predicted value and the test value fluctuates according to different algorithms. On the premise of ensuring the SLA, this method makes sufficient punishment for the situation of over-allocation of resources. The predicted value is consistent with the test value, and most of them remain in a state greater than 0, which means it has very good SLA satisfaction and high resource utilization rate. The prediction results of MAE and MSE are below the x-axis in many cases, which is likely to cause insufficient SLA guarantee when the resources are not satisfied. Compared with the DeepCog technology (D. Bega, M. Gramaglia, M. Fiore, A. Banchs, and X. Costa-Perez, “Deepcog: Cognitive network management in sliced 5g networks with deep learning,” in IEEE INFOCOM 2019 - IEEE Conference on Computer Communications, 2019, pp. 280–288.) which is also above the x-axis, this method is closer to the x-axis and has a higher resource utilization rate.

[0035] As Figure 5As shown, the prediction results of this method and other comparison algorithms in terms of SLA satisfaction index and resource utilization are presented. As can be seen from the figure, the SLA satisfaction index of this method reaches 0.99 because the user satisfaction is fully considered when designing the loss function, which is much higher than 0.43 under MAE and 0.69 under MSE. In terms of resource utilization, the resource utilization of the proposed RNN-based algorithm reaches 0.98 due to its appropriate penalty design for resource over-allocation, which is much better than the resource utilization of DeepCog at 0.82. Considering both the SLA satisfaction index and resource utilization comprehensively, the proposed invention has better performance than other inventions.

[0036] The above specific implementation can be locally adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific implementation, and all implementation solutions within its scope are subject to the present invention.

Claims

1. A method for allocating slice resources on the RAN side based on ensuring SLA, characterized in that, it includes a data set collection stage and a slice resource prediction stage, wherein: through the OAI system, collect the information interaction data generated during the wireless communication process between the user terminal and the base station in the actual wireless communication scenario, and match the data format of the input prediction network after preprocessing; input the information interaction data into the trained prediction network to obtain a resource allocation plan; The slice resource prediction stage includes: constructing a prediction network based on LSTM and performing offline training and slice resource orchestration; The constructing a prediction network based on LSTM and performing offline training includes: Step 2.1: Adopt a prediction network of LSTM containing a recurrent neural network to explore the behavioral habit characteristics of users on a long time scale; in the definition of the prediction time scale, use the historical interaction information between the UE and the eNodeB to predict the resource orchestration strategy of the slice in the next moment; Step 2.

2. Design the loss function: Transform the SLA guarantee index and resource utilization rate into the loss function of the prediction network, specifically as follows: Where: x is the ratio of the predicted number of resources to the required number of resources, 0 ≤ x, β > 0 represents a penalty term to characterize the situation where the eNodeB provides unacceptable services to users; 0 < α < 1 is used to describe the user's tolerance for services that cannot fully meet the quality of service requirements due to insufficient slice resources; the slope k > 0 is used as a penalty term in the case of over-allocation of resources; Step 2.3: Prediction network convergence judgment: Use the training set obtained in the data set collection stage to train the prediction network. This prediction network updates the weight parameters of each layer of the prediction network according to the loss value calculated in each step. When the network has not converged, the output of the network will be used as the network input to re-perform the next round of prediction network training, and continuously update the prediction network parameters. When the network converges, the training is completed, and the output of the prediction network can provide a basis for slice resource orchestration.

2. The method for allocating slice resources on the RAN side based on ensuring SLA according to claim 1, characterized in that, The data set collection stage includes: Step 1.1: Set the eNodeB to a fixed wireless spectrum bandwidth: Use two UEs with different service types, set different UEs under different slices, and correspond the difference in service load to the service attachments of different slices; Step 1.2: During the operation of the system, the interaction data between the UE and the eNodeB is saved by a pre-written log information saving program on the eNodeB, so as to obtain the actual slice data set in this wireless communication scenario; Step 1.3: Perform data cleaning and formatting processing on the actual slice data set, and convert the slice data into a data format that matches the input of the prediction network.

3. The method for allocating slice resources on the RAN side based on ensuring SLA according to claim 1, characterized in that, In the aforementioned step 2.2, by adding a minimum value ε to the original loss function, the shape of the loss function is not affected, and the situation where the slope is zero or infinite is avoided; the adjusted loss function is as follows: When the data is calculated through the prediction network, the output value calculated in the current step is measured by the performance of the loss function to obtain the corresponding loss value.

4. The method for allocating slice resources on the RAN side based on ensuring SLA according to claim 1, characterized in that, The slice resource orchestration refers to: when the slice data of the previous 1 second is input into the network, the data is calculated by the prediction network, and the resource prediction result of the next 100 milliseconds is output, and the slice resource orchestration plan can be completed according to this result.

5. A system for implementing the method according to any one of claims 1 to 4, characterized in that, it includes: The data set acquisition and preprocessing unit, the slice resource prediction unit, and the slice resource orchestration unit, where: The data set acquisition and preprocessing unit collects corresponding communication data according to the interaction process between the terminal device and the base station in the actual communication system, and performs normalization processing to obtain a data set that meets the input of the prediction network. The slice resource prediction unit performs slice resource prediction processing according to the established prediction network and the redefined loss function convergence algorithm, and obtains the slice resource prediction result at the next moment according to the historical resource allocation data. The slice resource orchestration unit performs the slice resource orchestration at the next moment according to the result of the resource prediction and the mapping rule between resource blocks, and obtains a better slice resource orchestration result.

Images