Operator edge computing node optimization method and system

Abstract

The invention relates to the technical field of edge computing, in particular to an operator edge computing node optimization method and system, and the method comprises the steps: obtaining local load features and cross-node association features of edge nodes in real time, and constructing a dynamic load vector and an association vector respectively; based on a preset time sequence prediction model, taking the load vector and the association vector as input, and obtaining a predicted load trend in a future preset time period; generating an optimal service migration strategy through a preset reinforcement learning model according to the predicted load trend in combination with the dynamic load vector, the association vector and the task attribute; wherein the migration strategy comprises a migration target node and a migration resource allocation scheme; wherein the task attributes comprise task priorities; and on the basis of the migration strategy, a resource pre-allocation scheme of the target node is generated in combination with a preset resource constraint condition, and the method has the advantages of improving service response, resource utilization rate, task success rate and the like in a high dynamic scene.

Description

Technical Field

[0001] This application relates to the technical field of edge computing, and in particular to a method and system for optimizing operator edge computing nodes. Background Technology

[0002] With the rapid development of 5G communication technology and IoT applications, edge computing, as a computing power deployment mode that is close to user terminals, is widely used in latency-sensitive scenarios such as intelligent transportation, industrial control, and real-time interaction.

[0003] Edge computing nodes, such as base station edge servers and regional edge clouds, need to handle massive distributed tasks. Their load status is significantly affected by dynamic factors such as user mobility, business bursts, and network topology changes. Traditional static management methods are no longer sufficient to meet service quality requirements.

[0004] Current optimization schemes for edge computing nodes suffer from the following core problems:

[0005] First, the load perception dimension is too narrow, relying only on static indicators such as local CPU utilization and memory usage of nodes, ignoring the topological connections between nodes and the characteristics of collaborative load fluctuations, resulting in a one-sided prediction of load changes.

[0006] Second, service migration decisions are prone to local optima traps. Based on fixed rules (such as "lowest load priority"), migration targets are selected without comprehensively considering the balance between migration costs (such as bandwidth consumption and latency) and task priorities, which can easily lead to the "ping-pong effect" (tasks are repeatedly migrated between nodes).

[0007] Third, the resource pre-allocation strategy lacks dynamic adaptability, and the static resource quota cannot respond to real-time changes in task priority, causing high-priority tasks to be interrupted due to failure to preempt resources. Summary of the Invention

[0008] To address the aforementioned issues, this application provides a method and system for optimizing operator edge computing nodes.

[0009] Firstly, this application provides a method for optimizing operator edge computing nodes, employing the following technical solution:

[0010] An optimization method for operator edge computing nodes includes the following steps:

[0011] Real-time acquisition of local load characteristics and cross-node association characteristics of edge nodes, and construction of dynamic load vectors and association vectors respectively; among which, local load characteristics include CPU utilization, memory utilization and network bandwidth utilization, and cross-node association characteristics include topological distance between nodes and historical collaborative load fluctuations between nodes;

[0012] Based on a pre-defined time-series forecasting model, using load vectors and correlation vectors as inputs, the predicted load trend within a pre-defined time period is obtained.

[0013] Based on the predicted load trend, combined with dynamic load vectors, correlation vectors, and task attributes, an optimal service migration strategy is generated through a pre-defined reinforcement learning model; wherein, the migration strategy includes the migration target node and the migration resource allocation scheme; wherein, the task attributes include task priority;

[0014] Based on the migration strategy and combined with preset resource constraints, a resource pre-allocation scheme for the target node is generated.

[0015] In one embodiment: the temporal prediction model is an LSTM model with an embedded spatiotemporal attention mechanism, the spatiotemporal attention mechanism introducing spatial attention weights based on association vectors; the prediction method of the temporal prediction model includes the following steps:

[0016] The basic LSTM unit based on the LSTM model outputs the hidden state of each node at the current time.

[0017] Based on spatial attention weights, the hidden states of predicted nodes and associated nodes are spatiotemporally fused.

[0018] Based on the fused hidden state, the predicted load trend is output.

[0019] In one embodiment: the prediction method of the time series prediction model further includes error correction, which specifically includes:

[0020] The prediction error is decomposed into a trend term and a periodic term, where the trend term represents the long-term change pattern and the periodic term represents the periodic fluctuation pattern.

[0021] Based on trend and periodic terms, the future load trend prediction results are adjusted by correction coefficients, and the correction coefficients are dynamically updated according to the error minimization objective.

[0022] In one embodiment: the reinforcement learning model is a deep Q-network, which calculates the target Q-value by combining the experience samples stored by experience replay with the target network, and updates the model parameters with the goal of minimizing the error between the predicted Q-value and the target Q-value, thereby generating the optimal service migration strategy; wherein, the experience samples include the current state space, the current action space, the current reward, and the next state space.

[0023] In one embodiment: the state space, action space and reward are all obtained based on predicted load trends, dynamic load vectors and task attributes, wherein the state space includes the current node load prediction value, the resource status of neighboring nodes, network link quality and task priority, and the action space includes the selection of migration target node and migration bandwidth allocation.

[0024] The formula for calculating the reward is:

[0025] ;

[0026] In the formula, B alloc For migration bandwidth allocation, B max Where T is the maximum bandwidth of the link, ΔT is the migration delay, and T is the maximum bandwidth of the link. max The maximum allowable latency is P, where P is the task priority and ω1, ω2, and ω3 are preset weight coefficients.

[0027] In one embodiment: the resource constraints include the target node having enough remaining CPU cores to meet the task's computational requirements, the bandwidth allocated to the migration task not exceeding the target node's currently idle bandwidth, and the migration completion time being less than the task deadline. The task's computational requirements are obtained based on the task priority and the target node's CPU frequency.

[0028] In one embodiment: the optimal pre-assignment scheme is generated by a genetic algorithm through selection, crossover, and mutation operations.

[0029] In one embodiment: the optimization objective of the genetic algorithm is to maximize the weighted sum of task success rate and resource utilization rate, that is:

[0030] ;

[0031] In the formula, S represents the task success rate, Ures represents the resource utilization rate, and γ1 and γ2 are preset weights.

[0032] In one embodiment: after the step of generating the resource pre-allocation scheme for the target node, the method further includes real-time collection of actual performance data after migration, and feedback and adjustment of the parameters in the two steps of optimizing the service migration strategy and the resource pre-allocation scheme based on the actual performance data;

[0033] The actual performance data includes migration latency and task success rate.

[0034] Secondly, this application provides an operator edge computing node optimization system, which adopts the following technical solution:

[0035] An operator edge computing node optimization system includes:

[0036] The data acquisition and processing module is used to acquire the local load characteristics and cross-node association characteristics of edge nodes in real time, and construct dynamic load vectors and association vectors respectively.

[0037] The load forecasting module, based on a preset time-series forecasting model, uses load vectors and correlation vectors as inputs to obtain the predicted load trend within a preset future time period.

[0038] The strategy generation module is used to generate an optimal service migration strategy based on the predicted load trend, combined with dynamic load vectors, correlation vectors, and task attributes, through a preset reinforcement learning model; wherein the migration strategy includes the migration target node and the migration resource allocation scheme; wherein the task attributes include task priority;

[0039] The resource pre-allocation module is used to generate a resource pre-allocation scheme for the target node based on the migration strategy and preset resource constraints.

[0040] In summary, this application has the following beneficial effects:

[0041] 1. Breaking through the traditional single-node monitoring mode, it integrates local indicators with cross-node correlation features for the first time, solving the problem of one-sided prediction of load changes;

[0042] 2. Employ spatiotemporal joint prediction, which integrates the dynamic influence of neighboring nodes to predict the node load trend within a preset time period in the future, to solve the problem of load prediction lag and alleviate the problem of service response latency in highly dynamic scenarios.

[0043] 3. Generate optimal service migration strategies through reinforcement learning, comprehensively considering the balance between migration costs and task priorities, and solve the local optimum problem in migration decisions;

[0044] 4. The resource pre-allocation strategy can respond to real-time changes in task priority to achieve dynamic adaptation, solve the problem of blind resource allocation, and improve resource utilization and task success rate in highly dynamic scenarios. Attached Figure Description

[0045] Figure 1 This is a flowchart of the operator edge computing node optimization method in this embodiment;

[0046] Figure 2 This is a logical block diagram of the operator edge computing node optimization system in this embodiment.

[0047] In the diagram, 10 is the data acquisition and processing module; 20 is the load prediction module; 30 is the strategy generation module; and 40 is the resource pre-allocation module. Detailed Implementation

[0048] The present application will be further described in detail below with reference to the accompanying drawings.

[0049] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. In some cases, to avoid obscuring various aspects of this application due to unnecessary description, well-known methods, processes, systems, components, and / or circuits already described at a higher level will not be elaborated upon. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0050] It should be noted that the descriptions of these embodiments are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0051] In the description of this application, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0052] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples.

[0053] like Figure 1 As shown in the figure, this embodiment discloses an optimization method for operator edge computing nodes, which specifically includes the following steps:

[0054] S100: Real-time acquisition of local load characteristics and cross-node association characteristics of edge nodes, and construction of dynamic load vectors and association vectors respectively.

[0055] In this step, local load characteristics include CPU utilization, memory utilization, and network bandwidth utilization; cross-node association characteristics are extracted based on graph neural network (GNN), which include topological distance between nodes and historical collaborative load fluctuation between nodes. Historical collaborative load fluctuation is characterized as the amount of historical load fluctuation between nodes within a set time period.

[0056] After normalizing the edge node features and cross-node association features, the local load features are constructed into a dynamic load vector through feature concatenation, and the cross-node association features are constructed into an association vector through feature concatenation.

[0057] The normalization of local load characteristics is achieved through the following formula:

[0058] ;

[0059] The normalization of topological distance is achieved using the following formula:

[0060] ;

[0061] The normalization of historical collaborative load fluctuations is achieved using the following formula:

[0062] ;

[0063] S200: Based on a preset time-series forecasting model, using load vector and correlation vector as input, obtain the predicted load trend within a preset future time period.

[0064] In this step, the aforementioned time series prediction model is an LSTM model with embedded spatiotemporal attention mechanism, which introduces spatial attention weights based on the correlation vector.

[0065] The prediction method of the time series prediction model includes the following steps:

[0066] The basic LSTM unit based on the LSTM model outputs the hidden state of each node at the current time.

[0067] Based on spatial attention weights, the hidden states of predicted nodes and associated nodes are spatiotemporally fused.

[0068] Based on the fused hidden state, the predicted load trend is output.

[0069] In this embodiment, the LSTM unit calculation formula is as follows:

[0070] Forgotten Gate:

[0071] f t =σ(W f ·[h t-1 ,x t ]+bf );

[0072] Among them, W f h is the forget gate weight matrix. t-1 The hidden state of the previous moment, x t The input at the current moment is the historical load sequence composed of historical local load characteristics, b f σ is the forget gate bias, and σ is the sigmoid activation function with an output range of [0,1].

[0073] Input Gate:

[0074] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0075] C ~ t =tanh(W C ·[h t-1 ,x t ]+b C );

[0076] Among them, W i W C Let b be the input gate weight matrix. i b C For the corresponding bias, tanh is the hyperbolic tangent activation function with an output range of [-1, 1].

[0077] Cell status update:

[0078] C t =f t ⊙C t-1 +i t ⊙C ~ t ;

[0079] Where ⊙ represents element-wise multiplication, and C t This represents the current state of the cell.

[0080] Output gate:

[0081] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0082] h t =o t ⊙tanh(C t );

[0083] Among them, W o Let b be the output gate weight matrix. o For output gate bias, h t Hide the current state.

[0084] Formula for calculating spatial attention weights:

[0085] e ij =MLP([h t (i) ,h t (j) ,d ij ,ΔL ij ]);

[0086] α ij =exp(e ij ) / ∑k∈Nexp(e ik );

[0087] Among them, h t (i) h represents the hidden state of the current node. t (j) Let d be the hidden state of neighboring node j, MLP is a multilayer perceptron, and d ij Let ΔL be the topological distance between node j and node i. ij For the historical collaborative load fluctuations between node j and node i, α ij Let be the spatial attention weight of node j to node i.

[0088] The formula for spatiotemporal fusion in the hidden state is:

[0089] h^ t =h t +∑ j∈N (αij·ht(j));

[0090] Among them, h^ t The hidden state after being integrated into spatial attention.

[0091] The formula for future load forecasting is:

[0092] L^ t+k =W y ·h^ t +b y ;

[0093] Among them, W y b y The output layer weights and biases are defined by k, which is the prediction step size and is set according to business requirements, such as 5s, 10s, etc.

[0094] In one embodiment, to further improve prediction accuracy, the prediction method of the time series prediction model further includes error correction, which specifically includes:

[0095] The prediction error is decomposed into a trend term and a periodic term, where the trend term represents the long-term change pattern and the periodic term represents the periodic fluctuation pattern.

[0096] Based on trend and periodic terms, the future load trend prediction results are adjusted by correction coefficients, and the correction coefficients are dynamically updated according to the error minimization objective.

[0097] Trend item decomposition:

[0098] T t =MA(e t-m ,...,e t );

[0099] In the formula, Tt is the trend term, m is the window size, MA is the moving average, and e t The prediction error is the difference between the actual load fluctuation and the predicted load trend of the LSTM.

[0100] Decomposition of periodic terms:

[0101] C t =FFT(e t );

[0102] The revised update formula is as follows:

[0103] L^ t +1=L^ t +λ·(T t +C t );

[0104] ;

[0105] S300: Based on the predicted load trend, combined with dynamic load vectors, correlation vectors, and task attributes, the optimal service migration strategy is generated through a preset reinforcement learning model.

[0106] In this step, the migration strategy includes the target node to be migrated and the migration resource allocation scheme, and the task attributes include task priority.

[0107] In addition, the reinforcement learning model mentioned above is a deep Q-network, which uses the experience samples stored by experience replay and combines them with the target network to calculate the target Q value. The model parameters are updated with the goal of minimizing the error between the predicted Q value and the target Q value, thereby generating the optimal service migration strategy.

[0108] The experience sample includes the current state space, the current action space, the current reward, and the next state space.

[0109] The aforementioned state space, action space, and reward are all obtained based on predicted load trends, dynamic load vectors, and task attributes. The state space includes the current node's predicted load value, the resource status of neighboring nodes, network link quality, and task priority. The action space includes the selection of the migration target node and the allocation of migration bandwidth.

[0110] The resource status of neighboring nodes includes the remaining number of CPU cores, available memory, and remaining bandwidth of neighboring nodes. The target node is selected based on the topological distance between nodes, and the migration bandwidth allocation value does not exceed the difference between the maximum bandwidth of the link and the current bandwidth utilization rate.

[0111] The formula for calculating the reward is:

[0112] ;

[0113] In the formula, B alloc For migration bandwidth allocation, B max Where T is the maximum bandwidth of the link, ΔT is the migration delay, and T is the maximum bandwidth of the link. max The maximum allowable latency is P, where P is the task priority and ω1, ω2, and ω3 are preset weight coefficients.

[0114] S400: Based on the migration strategy and combined with preset resource constraints, generate a resource pre-allocation scheme for the target node.

[0115] In this step, the resource constraints mentioned above include the target node having enough remaining CPU cores to meet the task's computational requirements, the bandwidth allocated to the migration task not exceeding the target node's currently idle bandwidth, and the migration completion time being less than the task's deadline.

[0116] The task computation requirements are obtained based on the task priority and the CPU frequency of the target node. Specifically, the task computation requirements are obtained by dividing the task priority by the CPU frequency of the target node.

[0117] The migration completion time is calculated by dividing the amount of data to be migrated by the bandwidth allocated to the migration task.

[0118] Furthermore, the optimal pre-assignment scheme is generated using a genetic algorithm through selection, crossover, and mutation operations. The optimization objective of the genetic algorithm is to maximize the weighted sum of task success rate and resource utilization rate, i.e.:

[0119] ;

[0120] In the formula, S represents the task success rate, Ures represents the resource utilization rate, and γ1 and γ2 are preset weights.

[0121] The S500 collects real-time performance data after migration and provides feedback and adjustments to parameters in the two steps of optimizing service migration strategy and resource pre-allocation scheme based on the actual performance data.

[0122] In this step, the aforementioned actual performance data includes migration latency and task success rate. This actual performance data is used to adjust the reward weights and resource constraints in the reinforcement learning model.

[0123] like Figure 2 As shown, this embodiment also provides an operator edge computing node optimization system, including:

[0124] The data acquisition and processing module 10 is used to acquire the local load characteristics and cross-node association characteristics of edge nodes in real time, and construct dynamic load vectors and association vectors respectively.

[0125] The load forecasting module 20, based on a preset time-series forecasting model, uses load vectors and correlation vectors as inputs to obtain the predicted load trend within a preset future time period.

[0126] The strategy generation module 30 is used to generate an optimal service migration strategy based on the predicted load trend, combined with dynamic load vectors, correlation vectors, and task attributes, through a preset reinforcement learning model; wherein the migration strategy includes the migration target node and the migration resource allocation scheme; wherein the task attributes include task priority;

[0127] The resource pre-allocation module 40 is used to generate a resource pre-allocation scheme for the target node based on the migration strategy and in combination with preset resource constraints.

[0128] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for optimizing operator edge computing nodes, characterized in that, Includes the following steps: Real-time acquisition of local load characteristics and cross-node association characteristics of edge nodes, and construction of dynamic load vectors and association vectors respectively; among which, local load characteristics include CPU utilization, memory utilization and network bandwidth utilization, and cross-node association characteristics include topological distance between nodes and historical collaborative load fluctuations between nodes; Based on a pre-defined time-series forecasting model, using load vectors and correlation vectors as inputs, the predicted load trend within a pre-defined time period is obtained. Based on the predicted load trend, combined with dynamic load vectors, correlation vectors, and task attributes, an optimal service migration strategy is generated through a pre-defined reinforcement learning model; wherein, the migration strategy includes the migration target node and the migration resource allocation scheme; wherein, the task attributes include task priority; Based on the migration strategy and combined with preset resource constraints, a resource pre-allocation scheme for the target node is generated.

2. The operator edge computing node optimization method according to claim 1, characterized in that: The time-series prediction model is an LSTM model with an embedded spatiotemporal attention mechanism, which introduces spatial attention weights based on the correlation vector. The prediction method of the time-series prediction model includes the following steps: The basic LSTM unit based on the LSTM model outputs the hidden state of each node at the current time. Based on spatial attention weights, the hidden states of predicted nodes and associated nodes are spatiotemporally fused. Based on the fused hidden state, the predicted load trend is output.

3. The operator edge computing node optimization method according to claim 2, characterized in that: The prediction method of the time series prediction model also includes error correction, which specifically includes: The prediction error is decomposed into a trend term and a periodic term, where the trend term represents the long-term change pattern and the periodic term represents the periodic fluctuation pattern. Based on trend and periodic terms, the future load trend prediction results are adjusted by correction coefficients, which are dynamically updated according to the error minimization objective.

4. The operator edge computing node optimization method according to claim 1, characterized in that: The reinforcement learning model is a deep Q-network, which calculates the target Q-value by storing experience samples through experience replay and combining them with the target network. The model parameters are updated with the goal of minimizing the error between the predicted Q-value and the target Q-value, thereby generating the optimal service migration strategy. The experience samples include the current state space, the current action space, the current reward, and the next state space.

5. The operator edge computing node optimization method according to claim 4, characterized in that: The state space, action space, and reward are all obtained based on predicted load trends, dynamic load vectors, and task attributes. The state space includes the current node load prediction value, the resource status of neighboring nodes, network link quality, and task priority. The action space includes the selection of migration target nodes and the allocation of migration bandwidth. The formula for calculating the reward is: ； In the formula, B alloc For migration bandwidth allocation, B max Where T is the maximum bandwidth of the link, ΔT is the migration delay, and T is the maximum bandwidth of the link. max The maximum allowable latency is P, where P is the task priority and ω1, ω2, and ω3 are preset weight coefficients.

6. The operator edge computing node optimization method according to claim 5, characterized in that: The resource constraints include the target node having enough remaining CPU cores to meet the task's computational requirements, the bandwidth allocated to the migration task not exceeding the target node's currently idle bandwidth, and the migration completion time being less than the task deadline. The task's computational requirements are based on the task priority and the target node's CPU frequency.

7. The operator edge computing node optimization method according to claim 6, characterized in that: The optimal pre-assignment scheme is generated by using a genetic algorithm through selection, crossover, and mutation operations.

8. The operator edge computing node optimization method according to claim 7, characterized in that: The optimization objective of the genetic algorithm is to maximize the weighted sum of task success rate and resource utilization rate, that is: ； In the formula, S represents the task success rate, Ures represents the resource utilization rate, and γ1 and γ2 are preset weights.

9. The operator edge computing node optimization method according to claim 8, characterized in that, The step of generating the resource pre-allocation scheme for the target node also includes real-time collection of actual performance data after migration, and feedback and adjustment of the parameters in the two steps of optimizing the service migration strategy and resource pre-allocation scheme based on the actual performance data. The actual performance data includes migration latency and task success rate.

10. An operator edge computing node optimization system, characterized in that, include: The data acquisition and processing module (10) is used to acquire the local load characteristics and cross-node association characteristics of edge nodes in real time, and construct dynamic load vectors and association vectors respectively. The load forecasting module (20) uses a preset time-series forecasting model, with load vector and correlation vector as input, to obtain the predicted load trend within a preset time period in the future. The strategy generation module (30) is used to generate the optimal service migration strategy based on the predicted load trend, combined with the dynamic load vector, correlation vector and task attributes, through a preset reinforcement learning model; wherein the migration strategy includes the migration target node and the migration resource allocation scheme; wherein the task attributes include the task priority; The resource pre-allocation module (40) is used to generate a resource pre-allocation scheme for the target node based on the migration strategy and in combination with preset resource constraints.

Images