Server cluster load balancing cooperative processing system and method and electronic equipment

Abstract

The invention discloses a server cluster load balancing cooperative processing system and method and electronic equipment, and relates to the technical field of servers, and the method comprises the steps: receiving node operation data sent by each edge node in a region; carrying out load prediction in the region according to the node operation data and a pre-trained neural network model to obtain a prediction result; determining a dynamic scheduling weight; solving calculation is carried out according to the multiple pre-stored node mapping schemes, preset solving parameters and a pre-established target function, and a multi-target Pareto solution set is obtained; performing preferential processing according to the Pareto solution set and the dynamic scheduling weight to obtain a target node mapping scheme; and according to the target node mapping scheme, controlling each edge node in the region to execute load balancing cooperative operation. The problem that the load balancing performance of the server cluster is poor is solved, and the server load balancing performance of the server cluster is improved by selecting a proper target node mapping scheme based on the node operation data collected by each edge node of the edge layer.

Description

Technical Field

[0001] The present application relates to the field of server technology, and in particular to a server cluster load balancing collaborative processing system, method and electronic device. Background Art

[0002] As the number of internet users increases, their demands for server response speeds are becoming increasingly stringent. Consequently, server clusters are increasingly being used to meet these demands. However, in practice, the load on server clusters can fluctuate over time. During periods of high traffic volume and changing business demands, poor server load balancing can lead to significant increases in request latency and server energy consumption.

[0003] Therefore, there is an urgent need for a server cluster load balancing coordination processing method to improve the above technical problems. Summary of the Invention

[0004] The present application provides a server cluster load balancing collaborative processing system, method and electronic device to at least solve the problem of poor server load balancing in the server cluster during special usage periods such as traffic peaks and changes in business needs in the related art.

[0005] The present application provides a server cluster load balancing collaborative processing system, comprising: an edge layer, a regional layer, and a global layer;

[0006] The edge layer includes multiple edge nodes, which are used to: respond to user-specified tasks and collect node operation data in real time during task execution;

[0007] The regional layer includes multiple regional coordinators, which are data-connected to at least one corresponding edge node. The regional coordinator is used to: receive node operation data sent by the corresponding edge node, and perform load forecasting in the region based on the node operation data and the pre-trained neural network model to obtain the forecast result; determine the dynamic scheduling weight based on the forecast result, node operation data, preset weight vector and preset feedback gain coefficient; perform solution calculation based on multiple pre-stored node mapping schemes and pre-built objective functions to obtain a multi-objective Pareto solution set; perform optimal processing based on the Pareto solution set and dynamic scheduling weight to obtain a target node mapping scheme; and control each edge node in the region to perform load balancing collaborative operations based on the target node mapping scheme.

[0008] The global layer is connected to the data of each regional coordinator in the regional layer. The global layer is used to: receive cross-regional scheduling requests and prediction results sent by each regional coordinator, and determine the cross-regional load balancing scheduling strategy in response to the cross-regional scheduling request based on the prediction results, the whole network resource view and the policy library. The cross-regional scheduling request is generated when the regional coordinator detects that the load in the region exceeds the preset threshold; send the cross-regional scheduling strategy to each regional coordinator to complete the cross-regional load balancing collaborative operation.

[0009] The present application also provides a server cluster load balancing collaborative processing method, including:

[0010] Receive node operation data sent by each edge node in the region, where the node operation data is collected in real time by the edge node;

[0011] Based on the node operation data and pre-trained neural network model, the load in the area is predicted to obtain the prediction results;

[0012] Determine the dynamic scheduling weight based on the prediction results, node operation data, preset weight vector and preset feedback gain coefficient;

[0013] Perform calculations based on multiple pre-existing node mapping schemes, preset solution parameters, and pre-built objective functions to obtain a multi-objective Pareto solution set;

[0014] The target node mapping solution is obtained by performing optimal processing based on the Pareto solution set and dynamic scheduling weights;

[0015] According to the target node mapping scheme, each edge node in the region is controlled to perform load balancing collaborative operations.

[0016] The present application also provides a method for coordinating server cluster loads, including:

[0017] Real-time collection of node operation data;

[0018] The node operation data is sent to the regional coordinator in the regional layer of the corresponding area, so that the regional coordinator performs load forecasting in the area based on the node operation data and the pre-trained neural network model to obtain the forecast result. The dynamic scheduling weight is determined based on the forecast result, the node operation data, the preset weight vector and the preset feedback gain coefficient. The solution is calculated based on the preset solution parameters and pre-built objective functions of multiple pre-stored node mapping schemes to obtain a multi-objective Pareto solution set. The optimal processing is performed based on the Pareto solution set and the dynamic scheduling weight to obtain the target node mapping scheme;

[0019] Receive the target node mapping plan sent by the regional coordinator of the corresponding area, and perform load balancing collaborative operations according to the target node mapping plan.

[0020] The present application also provides a server cluster load balancing collaborative processing method, including:

[0021] Receive cross-regional scheduling requests and prediction results sent by each regional coordinator in the regional layer. The prediction results are obtained by each regional coordinator in the regional layer based on the node operation data and pre-trained neural network model to predict the load in the region. The node operation data is collected in real time by each edge node corresponding to each regional coordinator and sent to each regional coordinator;

[0022] Determining a cross-region load scheduling strategy based on the prediction result, the network-wide resource view, and the strategy library in response to a cross-region scheduling request, wherein the cross-region scheduling request is generated when the regional coordinator detects that the load in the region exceeds a preset threshold;

[0023] Send the cross-region scheduling strategy to each regional coordinator to complete the cross-region load balancing collaborative operation.

[0024] The present application also provides a server cluster load balancing collaborative processing method, including:

[0025] Each edge node in each area of ​​the edge layer collects node operation data in real time;

[0026] Each edge node sends node operation data to the corresponding regional coordinator at the regional layer;

[0027] Each regional coordinator predicts the load in its region based on the node operation data and the pre-trained neural network model, obtains the prediction results, and determines the dynamic scheduling weight based on the prediction results, node operation data, preset weight vector and preset feedback gain coefficient;

[0028] Each regional coordinator performs calculations based on multiple pre-stored node mapping schemes, preset solution parameters, and pre-built objective functions to obtain a multi-objective Pareto solution set. It then performs optimal processing based on the Pareto solution set and dynamic scheduling weights to obtain a target node mapping scheme.

[0029] Each regional coordinator controls the edge nodes in the region to perform load balancing coordination operations according to the target node mapping scheme;

[0030] Each regional coordinator determines the regional load corresponding to each region based on the node operation data, and generates a cross-region scheduling request when it detects that the regional load corresponding to any region exceeds the preset threshold;

[0031] Each regional coordinator sends a cross-region scheduling request and the corresponding prediction results of each region to the global layer;

[0032] The global layer responds to cross-region scheduling requests and determines the cross-region load balancing scheduling strategy based on the prediction results, the whole network resource view and the policy library;

[0033] The global layer sends the cross-region load balancing scheduling strategy to the regional coordinator corresponding to each region in the regional layer;

[0034] The regional coordinators corresponding to each region complete cross-region load balancing collaborative operations according to the cross-region load balancing scheduling strategy.

[0035] The present application also provides an electronic device, comprising:

[0036] Memory for storing computer programs;

[0037] A processor is used to implement the steps of any of the above-mentioned server cluster load balancing collaborative processing methods when executing a computer program.

[0038] Through this application, since regional coordinators are added between the edge server nodes and the global manager node, when the edge nodes in each region can coordinate their loads, the regional coordinator can perform load predictions within the region based on the node operation data collected in real time by the edge nodes in each region and the pre-trained neural network model, obtaining prediction results. The scheduling weights are then dynamically adjusted based on the prediction results, existing node operation data, preset weight vectors, and feedback gain coefficients. A multi-objective Pareto solution is then performed to obtain a Pareto solution set consisting of more optimal pre-stored node mapping solutions. Finally, the final target node mapping solution is selected based on the dynamic scheduling weights. All edge nodes in the region are then controlled to perform load balancing collaborative operations based on the target node mapping solution. Therefore, this adaptive collaborative operation, due to the dynamic scheduling weights, can make the final target node mapping coordination operation plan more closely aligned with the actual needs and operating status of the business site. This can solve the technical problem of poor server load balancing in server clusters during special usage periods such as traffic peaks and changes in business demand in related technologies, achieving the technical effect of improving server load balancing in server clusters, improving load response efficiency, and reducing request response latency. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0040] Figure 1 A schematic diagram of the structure of a server cluster load balancing collaborative processing system provided in an embodiment of the present application;

[0041] Figure 2 Schematic diagram of the process of server cluster load balancing collaborative processing method provided in the application embodiment Figure 1 ;

[0042] Figure 3 Schematic diagram of the process of server cluster load balancing collaborative processing method provided in the embodiment of the application Figure 2 ;

[0043] Figure 4 Schematic diagram of the process of server cluster load balancing collaborative processing method provided in the embodiment of the application Figure 3 ;

[0044] Figure 5 A schematic diagram of the interactive process of the server cluster load balancing collaborative processing method provided in an embodiment of the present application;

[0045] Figure 6 Schematic diagram of the structure of the server cluster load balancing collaborative processing device provided in the embodiment of the application Figure 1 ;

[0046] Figure 7 Schematic diagram of the structure of the server cluster load balancing collaborative processing device provided in the embodiment of the application Figure 2 ;

[0047] Figure 8 Schematic diagram of the structure of the server cluster load balancing collaborative processing device provided in the application embodiment Figure 3 ;

[0048] Figure 9 This is a schematic diagram of the structure of the electronic device provided in this application. DETAILED DESCRIPTION

[0049] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0050] It should be noted that, in the description of this application, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. The terms "first," "second," etc., in this application are used to distinguish similar objects, and are not used to describe a particular order or sequence.

[0051] Based on the problems encountered in the use of server clusters in related technologies, an embodiment of the present invention proposes the following concept: the user's load flow coordination task is distributed to the edge nodes of the edge layer, the regional coordinators of the regional layer, and the global layer, so that the entire server cluster can meet the needs of quickly responding to user requests while also performing load collaborative scheduling at the regional layer and the global layer, significantly reducing scheduling delays and improving the load coordination performance of the entire server cluster.

[0052] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0053] Figure 1 A schematic diagram of the structure of a server cluster load balancing collaborative processing system provided in an embodiment of the present application.

[0054] like Figure 1 As shown, the server cluster load balancing collaborative processing system includes: an edge layer, a regional layer and a global layer.

[0055] Specifically, the edge layer includes multiple edge nodes, which are used to respond to user-specified tasks and collect node operation data in real time during task execution.

[0056] In this embodiment, an edge node refers to an edge server that can directly respond to user requests. When a user sends a request to an edge node, the edge node accepts the corresponding user request and executes the corresponding task. In this embodiment, node operation data can be the edge node's operating parameters, operating environment, or relevant parameters when executing tasks, such as network bandwidth and request latency.

[0057] The edge layer is divided into multiple regions, each of which includes at least one edge node, usually multiple edge nodes, such as Figure 1 As shown, there are edge nodes A1 to An in region A, edge nodes B1 to Bn in region B, and edge nodes M1 to Mn in region M, where n is a natural number greater than 0.

[0058] The regional layer includes multiple regional coordinators, which are data-connected to at least one corresponding edge node. The regional coordinator is used to: receive node operation data sent by the corresponding edge node, and perform load prediction in the region based on the node operation data and the pre-trained neural network model to obtain the prediction result; determine the dynamic scheduling weight based on the prediction result, node operation data, preset weight vector and preset feedback gain coefficient; perform solution calculation based on multiple pre-stored node mapping schemes and pre-built objective functions to obtain a Pareto solution set of multiple objectives; perform optimal processing based on the Pareto solution set and dynamic scheduling weight to obtain a target node mapping scheme; and control the edge nodes in the region to perform load balancing collaborative operations based on the target node mapping scheme.

[0059] In this embodiment, the regional layer includes multiple regional coordinators, which can be servers or hardware devices with the same functions as servers. The regional coordinators are responsible for load coordination between edge nodes in the corresponding region.

[0060] like Figure 1 As shown, in an optional embodiment of the present application, the regional coordinator includes a prediction module, a decision engine, and a communication interface, and the prediction module, decision engine, and communication interface are communicatively connected. The regional coordinator receives node operating data sent by each edge node in the corresponding region via the communication interface. The prediction module then uses the node operating data and a pre-trained neural network model to perform load forecasting within the region, obtaining a forecast result. The forecast result here can be a short-term load curve with a confidence interval. The pre-trained neural network can be a special recurrent neural network model, namely a long short-term memory (LSTM) model, obtained by training the model using pre-collected server operating parameters of each server in a server cluster when responding to user requests. This pre-trained neural network model introduces three gates to control memory: one gate is used to output entries from a memory cell, referred to as the output gate. Another gate is used to determine when to read data into the memory cell, referred to as the input gate. Another gate is used to control whether the contents of the memory cell are reset, referred to as the forget gate.

[0061] In this embodiment, in the LSTM model architecture, Represents the input information at the current moment,

[0062] Represents the hidden state at the previous moment, Represents the hidden state passed to the next moment. Represents the sigmoid activation function, through which data can be transformed Tanh represents the hyperbolic tangent function tanh function, through which the data can be transformed into The memory unit of LSTM at time 𝑡−1 Working together with the forget gate, input gate, and output gate, it controls the retention and discarding of information to obtain the memory unit of the LSTM at time 𝑡 This structure enables information to be transmitted in memory cells.

[0063] The information is transmitted and first passes through the forget gate. The forget gate acts on the cell state. , which controls the selective forgetting of information in the cell state, and decides which parts need to be discarded and which parts need to be retained. The formula of the forget gate is as follows:

[0064]

[0065] It represents the calculation result after the information at time 𝑡 passes through the forget gate, is the weight matrix of the forget gate, is the bias term of the forget gate, represents the output at time 𝑡−1, Denotes the input at time 𝑡. is the Sigmod activation function, The nonlinear Sigmund activation function is used to map the data to a range of 0-1, where 0 means all are discarded and 1 means all are retained.

[0066] in It represents the calculation result after the information at time 𝑡 passes through the input gate, is the weight matrix of the input gate. After the information passes through the input gate, the input gate determines which information is added to the cell state as new memory. The input gate performs two parts of calculation. The calculation formula for the first part is as follows:

[0067]

[0068] in, is the bias term of the input gate.

[0069] Then proceed to the second part of the calculation, the formula is as follows:

[0070]

[0071] This part constructs a candidate cell state , which saves and The information in The value of multiplication is used to determine which memories are useful. 0 represents complete discard, and 1 represents complete retention. The retained information is added to the new cell state as a new memory. It is called the candidate cell state. It is just a state waiting to be selected, and only useful ones will be added as new memories. is the weight matrix of the candidate memory element, is the bias term of the candidate memory element, and 𝑡𝑎𝑛ℎ is the hyperbolic tangent function.

[0072] The next step is to update the memory. The memory in the new cell state consists of two parts. One part is the old memory left after the memory cell forgets the useless memory at the previous moment. The other part is the useful information filtered out by the input gate as the new memory. The formula is as follows:

[0073]

[0074] symbol represents the Hadamard operation between matrices, This calculation determines what information to discard. Represents the updated information. The forgotten information plus the updated information can get the unit state at the time 𝑡 after the update This design is introduced to alleviate the vanishing gradient problem and better capture long-range dependencies in sequences.

[0075] Finally, the hidden state is calculated. The hidden state contains information about the current unit and can be used for prediction or decision making. The output gate first passes the unit state through a sigmoid layer to determine which part of the unit state will be output. Then, the unit state is processed through a tanh layer and multiplied by the output of the sigmoid gate to determine the final output. First, information flows to the output gate to calculate which part of the unit state will be output. The formula is as follows:

[0076]

[0077] in Represents the result after the information at time 𝑡 passes through the output gate, is the weight matrix of the output gate , is the bias term of the output gate.

[0078] Then the new cell state is The new hidden state is obtained by joint operation, and the formula is as follows:

[0079] At each time step, the LSTM unit takes the hidden state of the previous time step and the input of the current time step, updates its cell state, and then outputs a new hidden state and a prediction result.

[0080] like Figure 1 As shown, the global layer is connected to the data of each regional coordinator in the regional layer. The global layer is used to: receive cross-regional scheduling requests and prediction results sent by each regional coordinator, and respond to the cross-regional scheduling request based on the prediction results, the whole network resource view and the policy library to determine the cross-regional load balancing scheduling strategy, where the cross-regional scheduling request is generated when the regional coordinator detects that the load in the region exceeds the preset threshold; send the cross-regional scheduling strategy to each regional coordinator to complete the cross-regional load balancing collaborative operation.

[0081] In this embodiment, the global layer includes a network-wide resource view, a policy library, a cross-regional scheduling module, and a log audit and security module to coordinate the issuance of cross-regional scheduling policies. The global layer connects to the regional coordinators in the regional layer via a communication bus.

[0082] It should be noted that the data connection in the above system embodiment can be achieved through a wired connection or a wireless communication connection, which is limited in this embodiment.

[0083] In this embodiment, the edge layer, regional layer, and global layer can achieve coordinated optimization of low-latency response and network-wide load balancing through an online feedback loop, thereby improving the load balancing performance of the entire server cluster in special scenarios where traffic surges or business demands change.

[0084] Figure 2 Schematic diagram of the process of server cluster load balancing collaborative processing method provided in the embodiment of the application Figure 1 .like Figure 2 As shown, the embodiment of the present application provides a server cluster load balancing collaborative processing method, which is described in detail as follows:

[0085] S201: Receive node operation data sent by each edge node in the region, wherein the node operation data is collected in real time by the edge node.

[0086] In this embodiment, the node operation data may include data such as the CPU utilization, memory usage, network bandwidth, request response delay, and energy consumption of the edge node itself.

[0087] S202: Perform load prediction in the area based on the node operation data and the pre-trained neural network model to obtain a prediction result.

[0088] In this embodiment, the pre-trained neural network model is a long short-term memory (LSTM) model, which has prediction and decision-making functions. After inputting node operation data, it can output a short-term load curve for the future.

[0089] Based on the above embodiment, in an optional embodiment of the present application, step S202 includes:

[0090] S202a: Perform data preprocessing based on the node operation data to obtain normalized node data.

[0091] In this embodiment, data preprocessing can include data screening and data normalization. Data screening refers to removing obviously erroneous data from node operation data. Data normalization refers to normalizing all data into a data set within a specific numerical range. For example, this can be done by normalizing a digital data set into a uniform format or normalizing a data set to a value less than 1. This facilitates subsequent batch data reading and processing.

[0092] S202b: performing data batch reading processing according to the normalized node data and a preset window size to obtain data to be predicted for multiple periods arranged in time sequence.

[0093] In this embodiment, step S202b refers to the process of batch reading data using a sliding window method, wherein the preset window size can be a pre-set sliding window size. For example, the preset window size is 1-5 minutes. That is, the normalized node operation data for the most recent 1-5 minutes of the normalized node data is read using the sliding window method. The data to be predicted for multiple periods arranged in time sequence is obtained so that the dynamic scheduling weight can be adjusted in real time based on each period.

[0094] S202c: Input the data to be predicted in each cycle into the pre-trained neural network model in sequence to obtain a predicted load curve including a confidence interval as a prediction result.

[0095] In this embodiment, the specific input and output processes of the pre-trained neural network model have been described in detail in the aforementioned system embodiment and will not be repeated here. After sequentially inputting the data to be predicted for each period into the pre-trained neural network model, the pre-trained neural network model predicts the load trends of each edge node over the next Δ (1–5 minutes) period and ultimately outputs a short-term load curve with confidence intervals.

[0096] S203: Determine the dynamic scheduling weight according to the prediction result, the node operation data, the preset weight vector and the preset feedback gain coefficient.

[0097] In this embodiment, the preset weight vector can be a multi-dimensional indicator weight vector input by the user based on business needs. This multi-dimensional indicator weight vector can provide data support for adjusting subsequent load balancing collaborative processing. The dynamic scheduling weight refers to the specific weight values ​​of the key indicators required for load balancing collaborative processing, calculated during the current cycle. For example, the dynamic scheduling weights are: bandwidth usage weight 0.1, energy consumption weight 0.1, and request response delay weight 0.8.

[0098] In this embodiment, the preset feedback gain coefficient can be manually set to any decimal within a fixed range. For example, the preset feedback coefficient can be any decimal between 0.1 and 0.3.

[0099] Specifically, in an optional embodiment of the present application, the preset weight vector includes a delay weight, an energy consumption weight, and a plurality of different node operation weights; accordingly, step S203 includes:

[0100] S203a: Determine actual delays and actual energy consumption of multiple cycles arranged in time sequence according to the node operation data.

[0101] In this embodiment, the corresponding actual delay and actual energy consumption are extracted based on the CPU utilization, memory usage, network bandwidth, request response delay, and energy consumption data collected in real time in a time series in the node operation data. The period here can also be 1-5 minutes. The operation weights of multiple different nodes can include the weights of operation indicators related to edge node operation, such as CPU benefit weight and network bandwidth weight.

[0102] S203b: Determine the predicted delay of the current cycle and the predicted energy consumption value of the current cycle according to the prediction result of the current cycle.

[0103] In this embodiment, the prediction results refer to future predicted numerical curves encompassing multiple dimensions, including CPU utilization, memory usage, network bandwidth, request-response latency, and energy consumption. These curves comprise the short-term load curve described in the above embodiments. The confidence intervals can be used to determine whether the corresponding values ​​in the short-term load curve are usable, thereby deriving the predicted latency and energy consumption values ​​for the current period.

[0104] S203c: Determine a delay error based on the predicted delay and the actual delay of the current cycle, and determine an energy consumption error based on the actual energy consumption and the predicted energy consumption of the current cycle.

[0105] In this embodiment, the delay error can be obtained by performing a difference calculation between the actual delay and the predicted delay, and the energy consumption error can be obtained by performing a difference calculation between the actual energy consumption and the predicted energy consumption.

[0106] S203d: Determine an adjusted delay weight according to the delay error, the delay weight in the preset weight vector, and the preset feedback gain parameter.

[0107] In this embodiment, the adjusted delay weight can be obtained by adding the delay weight in the preset weight vector to the product of the delay error and the preset feedback gain coefficient.

[0108] S203e: Determine an adjusted energy consumption weight according to the energy consumption error, the energy consumption weight in the preset weight vector, and the preset feedback gain parameter.

[0109] In this embodiment, the adjusted energy consumption weight can be obtained by adding the energy consumption weight in the preset weight vector to the product of the delay error and the preset feedback gain coefficient.

[0110] S203f: Perform weight adjustment based on the adjusted delay weight, the adjusted energy consumption weight, and multiple different node operation weights to obtain an adjusted weight vector, and perform normalization processing based on the adjusted weight vector to obtain a dynamic scheduling weight.

[0111] In this embodiment, the dynamic scheduling weight can be obtained by integrating multiple different node operation weights, adjusted delay weights and adjusted energy consumption weights in the initial weight vector, and then normalizing the integrated weight vector. The normalization here can be to adjust the weight values ​​in the integrated vector so that the final sum of the weights is 1. The final weight values ​​are used as dynamic scheduling weights.

[0112] In an optional embodiment of the present application, when calculating the adjusted delay weight and adjusted energy consumption weight of the next cycle, the delay weight and energy consumption weight in the dynamic scheduling weight calculated in the current cycle may be used to achieve dynamic and continuous optimization and adjustment.

[0113] S204: performing solution calculations according to a plurality of pre-stored node mapping schemes, preset solution parameters, and pre-built objective functions to obtain a multi-objective Pareto solution set.

[0114] In this embodiment, the pre-stored node mapping schemes can be candidate task-to-edge node mapping schemes pre-set and stored based on different business scenarios. For example, when the load on edge node A1 reaches a certain upper load threshold, the candidate task is mapped to any edge node from A2 to An. In this embodiment, the multi-objective Pareto solution set can be a set of pre-stored node mapping schemes that have relatively high feasibility among multiple pre-stored node mapping schemes.

[0115] The specific calculation method of step S204 can be a non-dominated sorting genetic algorithm II (hereinafter referred to as NSGA-II algorithm), and the preset solution parameters can be the population size N, number of iterations G, crossover rate Cr and mutation rate Mr required for the NSGA-II algorithm to be initialized.

[0116] Specifically, in an optional embodiment of the present application, the pre-built objective functions include an average response delay objective function, a network bandwidth occupancy objective function, and a total energy consumption objective function. The preset solution parameters include population size, number of iterations, crossover rate, and mutation rate. Step S204 includes:

[0117] S204a: Establish multiple chromosome codes according to each pre-stored node mapping scheme.

[0118] S204b: The response delay objective function, the network bandwidth occupancy objective function, and the total energy consumption objective function are taken as three functions to be solved, and an iterative solution process is performed through a non-dominated sorting genetic algorithm based on the three functions to be solved, multiple chromosome codes, population size, number of iterations, crossover rate, and mutation rate to obtain a multi-objective Pareto solution set, wherein the multi-objective Pareto solution set includes at least one pre-existing node mapping scheme.

[0119] In this embodiment, the NSGA-II algorithm is used as the core, chromosome encoding is established for each pre-existing node mapping scheme, the population size N (50-200), iteration number G (30-100), crossover rate Cr (0.8) and mutation rate Mr are set, and "average response delay", "network bandwidth occupancy" and "total system energy consumption" are used as three objective functions. First, a fast non-dominated sort is performed and then the non-dominated solutions at the forefront of the Pareto solution set are selected as the Pareto solution set according to the crowding distance until the solution process reaches the number of iterations or meets other iteration stop conditions.

[0120] S205: Perform optimal processing according to the Pareto solution set and the dynamic scheduling weight to obtain a target node mapping solution.

[0121] In this embodiment, the optimal processing may be to select, according to the dynamic scheduling weight, a pre-stored node mapping solution that better matches the dynamic scheduling weight from the Pareto solution set as the target node mapping solution.

[0122] Specifically, in an optional embodiment of the present application, step S205 includes:

[0123] S205a: Calculate a solution score according to at least one pre-stored node mapping solution and a dynamic scheduling weight to obtain a node score for each pre-stored node mapping solution.

[0124] In this embodiment, the score calculation may be to calculate the scores of nodes according to each pre-stored node mapping scheme based on dynamic weights, and specifically to calculate the running scores of edge nodes of corresponding weights based on different weights in the dynamic scheduling weights.

[0125] S205b: According to the node scores of the pre-existing node mapping solutions, the pre-existing node mapping solution corresponding to the lowest node score is selected and determined as the target node mapping solution.

[0126] In this embodiment, the lower the node score is, the lower the probability that the load of the corresponding node mapping solution exceeds the load threshold is, and therefore the node is more suitable for subsequent load balancing collaborative operations.

[0127] S206: Control each edge node in the region to perform load balancing collaborative operations according to the target node mapping solution.

[0128] In this embodiment, according to the target node mapping scheme, the current load or subsequent load of the fully loaded edge nodes in each edge node in the region is controlled to migrate to the unloaded edge nodes, so that the loads of each edge node tend to be close or even completely equal.

[0129] Specifically, in an optional embodiment of the present application, step S206 includes:

[0130] S206a: Determine the load to be migrated and the load migration route in this area according to the target node mapping solution.

[0131] In this embodiment, determining the load to be migrated and the load migration route in the region may include extracting the device identity of the edge node from which the load is to be transferred and the device identity of the edge node from which the load is to be transferred, according to the target node mapping scheme, and generating a load migration route based on the device identity of the edge node from which the load is to be transferred and the device identity of the edge node from which the load is to be transferred. The load to be migrated is then retrieved from the edge node based on the device identity of the edge node from which the load is to be transferred.

[0132] S206b: Send the load to be migrated and the load migration route to each edge node in this area, so that each edge node migrates the load to be migrated according to the load migration route to complete the load balancing collaborative operation, or deploys a processing container and inputs the load to be migrated into the processing container for processing to complete the load balancing collaborative operation.

[0133] In this embodiment, the CRIU (Checkpoint / Restore In Userspace) method can be used to freeze and restore the application or container state within user space, enabling the migration of containers from one edge node to another without user awareness, thereby completing the container deployment and processing process. Alternatively, VM Live-Migration can be used to migrate virtual machines from one edge node to another without interrupting edge node service to users, thereby achieving load balancing coordination.

[0134] The above embodiments are intended to describe the load balancing collaborative processing process when the regional load or network condition does not exceed a preset threshold.

[0135] Based on the above embodiment, a server cluster load coordination processing method provided in an optional embodiment of the present application further includes:

[0136] Step A: Determine the regional load corresponding to each area based on the node operation data.

[0137] In this embodiment, the regional-level load may be the sum of the loads currently processed by the edge nodes in the region. The sum of the loads may be the sum of the traffic. That is, the loads of the edge nodes are numerically added to obtain the regional-level load corresponding to the region.

[0138] Step B: When it is detected that the regional load corresponding to any area exceeds a preset threshold, a cross-region scheduling request is generated.

[0139] In this embodiment, the preset threshold value may be a pre-set upper limit on the load that the edge nodes corresponding to the region can handle. When the regional load corresponding to any region exceeds the preset threshold value, it indicates that the regional coordinator corresponding to that region is no longer able to effectively migrate load between edge nodes. In this case, edge nodes in other regions must be called upon to handle the excess load in the region.

[0140] Step C: Send the cross-region scheduling request and the prediction results corresponding to each region to the global layer, so that the global layer responds to the cross-region scheduling request and determines the cross-region load balancing scheduling strategy based on the prediction results, the whole network resource view and the policy library.

[0141] In this embodiment, the entire network resource library can be the relevant data of all edge nodes in the entire server cluster. Based on the prediction results, the load of edge nodes in each region in the future period can be determined. Based on the predicted load and the entire network resource library, the corresponding cross-region node mapping strategy is selected according to the pre-set strategy library as the cross-region load balancing scheduling strategy.

[0142] Step D: Receive the cross-region load balancing scheduling strategy sent by the global layer, and complete the cross-region load balancing collaborative operation according to the cross-region load balancing scheduling strategy.

[0143] In this embodiment, after receiving the cross-region load balancing scheduling policy sent by the global layer, the regional coordinator implements load migration using a method similar to that in step S205. The specific process of this embodiment will not be repeated here. The difference is that the load migration in this embodiment is load migration between edge nodes in different regions.

[0144] Based on the above embodiment, a server cluster load coordination processing method provided as an optional embodiment of the present application further includes:

[0145] Step E: Receive execution feedback results of cross-region load coordination operations from each edge node in each region.

[0146] In this embodiment, the execution feedback result refers to indicator data such as actual delay, actual bandwidth usage, and energy consumption collected when each edge node performs corresponding tasks for the cross-region load coordination operation.

[0147] Step F: Adjust the model parameters of the pre-trained neural network model according to the execution feedback result to obtain an adjusted neural network model.

[0148] In this embodiment, by referring to the actual delay, actual bandwidth occupancy and energy consumption in the execution feedback results, the adjusted model parameters can be manually or automatically generated using the corresponding parameter adjustment instructions, and updated to the pre-trained neural network model to obtain the adjusted neural network model. The adjusted model parameters may include the weight matrix 、 and at least one of the learning rate η.

[0149] Step G: Adjust the algorithm parameters of the preset solution parameters according to the execution feedback result to obtain the adjusted preset solution parameters.

[0150] In this embodiment, referring to the actual delay, actual bandwidth occupancy and energy consumption and other indicator data in the execution feedback results, the corresponding preset solution parameters can also be adjusted manually or by calling, and the adjusted preset solution parameters may include at least one of the population size, number of iterations, crossover rate and mutation rate.

[0151] Based on the above embodiment, a server cluster load coordination processing method provided in an optional embodiment of the present application further includes:

[0152] Step H: regularly obtain business demand information, and adjust the preset weight vector according to the business demand information to obtain the weight vector required by the current business.

[0153] In this embodiment, the service demand information can be the load demand during different service time periods. The weights of different indicators are adjusted to obtain the weight vector required for the current service. For example, the traffic load at night is greater than the traffic load during the day. Therefore, to meet the user's usage needs, the energy consumption weight can be reduced and the delay weight can be increased.

[0154] Step I: Determine the dynamic scheduling weight required for the current business based on the prediction results, node operation data, the weight vector required for the current business, and the preset feedback gain coefficient.

[0155] In this embodiment, the method and principle in step I are similar to those in step S203 in the above embodiment, so they will not be described again in detail in this embodiment.

[0156] In summary, the server cluster load balancing collaborative processing method provided by the embodiment of the present application first predicts the load in the region based on the node operation data collected in real time by the edge nodes in each region and the pre-trained neural network model to obtain the prediction result. Then, based on the prediction result, the existing node operation data, the preset weight vector and the feedback gain coefficient, the scheduling weight is dynamically adjusted, and then a multi-objective Pareto solution is performed to obtain a Pareto solution set consisting of more optimal pre-stored node mapping schemes. Finally, the final target node mapping scheme is obtained based on the dynamic scheduling weight, and then all edge nodes in the region are controlled to perform load balancing collaborative operations according to the target node mapping scheme. Therefore, this adaptive collaborative operation, due to the dynamic scheduling weight, can make the final target node mapping coordination operation scheme more in line with the actual needs and operating status of the business site. It can solve the technical problem of poor server load balancing of the server cluster in special use periods such as traffic peaks and changes in business needs in the related technology, and achieve a better technical effect of improving the server load balancing of the server cluster, which can improve load response efficiency and reduce request response delay.

[0157] At the same time, by adjusting the weight vector for a special business demand according to the business demand information, and adjusting the dynamic scheduling weight required by the current business based on this weight vector to better meet the different business needs, we can then obtain a cross-regional scheduling strategy or target node mapping solution that is more in line with the current business needs, thereby improving the load balancing performance of the entire server cluster.

[0158] At the same time, by subsequently receiving the execution feedback results of cross-regional load collaborative operations from each edge node in each region, the model parameters of the pre-trained neural network model and the algorithm parameters of the preset solution parameters are adjusted to form a feedback-driven load balancing collaborative processing method, so that the entire server cluster can maintain high reliability, high performance and low power consumption in multi-business and multi-regional scenarios.

[0159] Figure 3 Schematic diagram of the process of server cluster load balancing collaborative processing method provided in the embodiment of the application Figure 2 .

[0160] like Figure 3 As shown, the embodiment of the present application also provides a server cluster load balancing collaborative processing method, which is described in detail as follows:

[0161] S301: Collect node operation data in real time.

[0162] In this embodiment, the real-time collection of node operation data may be obtained through device operation detection software pre-deployed in each edge node.

[0163] S302: Send the node operation data to the regional coordinator in the regional layer of the corresponding area, so that the regional coordinator performs load prediction in the area based on the node operation data and the pre-trained neural network model to obtain the prediction result, and determines the dynamic scheduling weight according to the prediction result, the node operation data, the preset weight vector and the preset feedback gain coefficient, and performs solution calculation according to the preset solution parameters and pre-built objective functions of multiple pre-stored node mapping schemes to obtain a multi-objective Pareto solution set, and performs optimal processing according to the Pareto solution set and the dynamic scheduling weight to obtain the target node mapping scheme.

[0164] In this embodiment, the node operation data may be sent to the regional coordinator in the regional layer of the corresponding region through the communication interface.

[0165] S303: Receive the target node mapping plan sent by the regional coordinator of the corresponding area, and perform load balancing collaborative operations according to the target node mapping plan.

[0166] In an optional embodiment of the present application, step S303 includes:

[0167] S303a: Performing load balancing coordination operations according to the target node mapping solution, including:

[0168] S303b: Determine the load to be migrated and the load migration route in this area according to the target node mapping solution.

[0169] S303c: Migrate the load to be migrated according to the load migration route to complete the load balancing collaborative operation.

[0170] S303d: Alternatively, a processing container is deployed, and the load to be migrated is input into the processing container for processing to complete the load balancing collaborative operation.

[0171] In this embodiment, the target node mapping solution in step S303 performs load balancing collaborative operation. Figure 2The step S206 in the illustrated embodiment has been described, and thus will not be described again in this embodiment.

[0172] In this embodiment, the principles of steps S301 to S303 related to the regional layer are as follows: Figure 2 This has been explained in the illustrated embodiment, so it will not be described again in detail in this embodiment.

[0173] In an optional embodiment of the present application, after performing the load balancing coordination operation according to the target node mapping scheme in step S303, the following steps are further included:

[0174] S304: Send the execution feedback results for the load balancing collaborative operation to the regional coordinator of the corresponding area, so that the regional coordinator adjusts the model parameters of the pre-trained neural network model according to the execution feedback results to obtain the adjusted neural network model, and adjusts the algorithm parameters of the preset solution parameters according to the execution feedback results to obtain the adjusted preset solution parameters.

[0175] In this embodiment, step S304 and Figure 2 The principles and effects of steps E to G in the illustrated method embodiment are similar, so they will not be described in detail in this embodiment.

[0176] Figure 4 Schematic diagram of the process of server cluster load balancing collaborative processing method provided in the embodiment of the application Figure 3 .

[0177] like Figure 4 As shown, the embodiment of the present application also provides a server cluster load balancing collaborative processing method, which is described in detail as follows:

[0178] S401: Receive cross-region scheduling requests and prediction results sent by each regional coordinator in the regional layer, where the prediction results are obtained by each regional coordinator in the regional layer based on node operation data and a pre-trained neural network model to predict the load in the region. The node operation data is collected in real time by each edge node corresponding to each regional coordinator and sent to each regional coordinator;

[0179] S402: Determine a cross-region load scheduling policy in response to a cross-region scheduling request based on the prediction result, the network-wide resource view, and the policy library, wherein the cross-region scheduling request is generated when the regional coordinator detects that the load in the local region exceeds a preset threshold;

[0180] S403: Send the cross-region scheduling strategy to each regional coordinator to complete the cross-region load balancing collaborative operation.

[0181] In this embodiment, the implementation principle and technical effect of steps S401 to S403 are as follows: Figure 2Steps A to D in the illustrated embodiment have been described, and thus will not be described again in detail in this embodiment.

[0182] Figure 5 A schematic diagram of the interactive process of the server cluster load balancing collaborative processing method provided in an embodiment of the present application.

[0183] like Figure 5 As shown, the embodiment of the present application also provides a server cluster load balancing collaborative processing method, which is described in detail as follows:

[0184] S501: Each edge node in each area of ​​the edge layer collects node operation data in real time.

[0185] S502: Each edge node sends node operation data to the corresponding regional coordinator of the regional layer.

[0186] S503: Each regional coordinator predicts the load in the region based on the node operation data and the pre-trained neural network model to obtain the prediction result, and determines the dynamic scheduling weight based on the prediction result, node operation data, preset weight vector and preset feedback gain coefficient.

[0187] S504: Each regional coordinator performs solution calculations based on multiple pre-stored node mapping schemes, preset solution parameters and pre-built objective functions to obtain a multi-objective Pareto solution set, and performs optimal processing based on the Pareto solution set and dynamic scheduling weights to obtain a target node mapping scheme.

[0188] S505: Each regional coordinator controls each edge node in the region to perform load balancing coordination operations according to the target node mapping solution.

[0189] S506: Each regional coordinator determines the regional load corresponding to each region based on the node operation data, and generates a cross-region scheduling request when detecting that the regional load corresponding to any region exceeds a preset threshold.

[0190] S507: Each regional coordinator sends a cross-region scheduling request and the prediction results corresponding to each region to the global layer.

[0191] S508: The global layer determines a cross-region load balancing scheduling strategy in response to the cross-region scheduling request based on the prediction results, the entire network resource view, and the strategy library.

[0192] S509: The global layer sends the cross-region load balancing scheduling policy to the regional coordinator corresponding to each region of the regional layer.

[0193] S5010: The regional coordinators corresponding to each region complete the cross-region load balancing collaborative operation according to the cross-region load balancing scheduling strategy.

[0194] The implementation principle of each step in this embodiment is Figure 2 The steps of the method embodiment have been described in detail, so they will not be repeated here in this embodiment.

[0195] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method.

[0196] Figure 6 Schematic diagram of the structure of the server cluster load balancing collaborative processing device provided in the embodiment of the application Figure 1 .like Figure 6 As shown, an embodiment of the present application further provides a server cluster load balancing collaborative processing device, including: a communication interface module 61, a prediction module 62, a decision engine 63 and a load balancing collaborative module 64.

[0197] The communication interface module 61 is used to receive node operation data sent by each edge node in the region, wherein the node operation data is collected in real time by the edge node;

[0198] The prediction module 62 is used to perform load prediction in the region based on the node operation data and the pre-trained neural network model to obtain a prediction result;

[0199] A decision engine 63 is configured to determine a dynamic scheduling weight based on the prediction result, the node operation data, a preset weight vector, and a preset feedback gain coefficient;

[0200] The decision engine 63 is further configured to perform a solution calculation based on a plurality of pre-stored node mapping schemes, preset solution parameters, and pre-built objective functions to obtain a multi-objective Pareto solution set;

[0201] The decision engine 63 is further used to perform optimal processing based on the Pareto solution set and the dynamic scheduling weight to obtain a target node mapping solution;

[0202] The load balancing coordination module 64 is used to control the edge nodes in the region to perform load balancing coordination operations according to the target node mapping solution.

[0203] In an optional embodiment of the present application, the prediction module 62 is specifically used to: perform data preprocessing based on the node operation data to obtain normalized node data; perform data batch reading and processing according to the preset window size based on the normalized node data to obtain multiple cycles of data to be predicted arranged in time series; input the data to be predicted of each cycle into the pre-trained neural network model in turn to obtain a predicted load curve containing a confidence interval as a prediction result.

[0204] In an optional embodiment of the present application, the preset weight vector includes a delay weight, an energy consumption weight and a plurality of different node operation weights; accordingly, the decision engine 63 is specifically used to: determine the actual delay and actual energy consumption of a plurality of cycles arranged in time sequence according to the node operation data; determine the predicted delay of the current cycle and the predicted energy consumption of the current cycle according to the prediction result of the current cycle; determine the delay error according to the predicted delay and the actual delay of the current cycle, and determine the energy consumption error according to the actual energy consumption and the predicted energy consumption of the current cycle; determine the adjusted delay weight according to the delay error, the delay weight in the preset weight vector and the preset feedback gain parameter; determine the adjusted energy consumption weight according to the energy consumption error, the energy consumption weight in the preset weight vector and the preset feedback gain parameter; perform weight adjustment according to the adjusted delay weight, the adjusted energy consumption weight and a plurality of different node operation weights to obtain an adjusted weight vector, and perform normalization processing according to the adjusted weight vector to obtain a dynamic scheduling weight.

[0205] In an optional embodiment of the present application, the pre-built objective functions include an average response delay objective function, a network bandwidth occupancy objective function and a total energy consumption objective function, and the preset solution parameters include population size, number of iterations, crossover rate and mutation rate; accordingly, the decision engine 63 is further specifically used to: establish multiple chromosome codes according to each pre-stored node mapping scheme; use the response delay objective function, the network bandwidth occupancy objective function and the total energy consumption objective function as three functions to be solved, and perform iterative solution processing through a non-dominated sorting genetic algorithm based on the three functions to be solved, multiple chromosome codes, population size, number of iterations, crossover rate and mutation rate to obtain a Pareto solution set of multiple objectives, wherein the Pareto solution set of multiple objectives includes at least one pre-stored node mapping scheme.

[0206] In an optional embodiment of the present application, the decision engine 63 is further specifically used to: calculate a scheme score based on at least one pre-stored node mapping scheme and a dynamic scheduling weight to obtain a node score for each pre-stored node mapping scheme; and based on the node scores of each pre-stored node mapping scheme, select the pre-stored node mapping scheme corresponding to the lowest node score and determine it as the target node mapping scheme.

[0207] In an optional embodiment of the present application, the load balancing collaboration module 64 is specifically used to: determine the load to be migrated and the load migration route in this area according to the target node mapping scheme; send the load to be migrated and the load migration route to each edge node in this area, so that each edge node migrates the load to be migrated according to the load migration route to complete the load balancing collaboration operation, or deploys a processing container and inputs the load to be migrated into the processing container for processing to complete the load balancing collaboration operation.

[0208] In an optional embodiment of the present application, the load balancing collaboration module 64 is also used to: determine the regional-level load corresponding to each area based on the node operation data; generate a cross-regional scheduling request when it is detected that the regional-level load corresponding to any area exceeds a preset threshold; send the cross-regional scheduling request and the prediction results corresponding to each area to the global layer, so that the global layer responds to the cross-regional scheduling request based on the prediction results, the entire network resource view and the policy library to determine the cross-regional load balancing scheduling strategy; receive the cross-regional load balancing scheduling strategy sent by the global layer, and complete the cross-regional load balancing collaboration operation according to the cross-regional load balancing scheduling strategy.

[0209] In an optional embodiment of the present application, the communication interface module 61 is also used to receive execution feedback results of each edge node in each region for cross-regional load collaborative operations; adjust model parameters of the pre-trained neural network model according to the execution feedback results to obtain an adjusted neural network model; and adjust algorithm parameters of the preset solution parameters according to the execution feedback results to obtain adjusted preset solution parameters.

[0210] Figure 6 The description of the features of the server cluster load balancing collaborative processing device in the embodiment can be found in Figure 2 The relevant description of the embodiment of the server cluster load balancing collaborative processing method is not repeated here.

[0211] Figure 7 Schematic diagram of the structure of the server cluster load balancing collaborative processing device provided in the embodiment of the application Figure 2 .like Figure 7 As shown, an embodiment of the present application further provides a server cluster load balancing collaborative processing device, including: a data acquisition module 71, a data sending module 72 and a receiving module 73.

[0212] The data collection module 71 is used to collect node operation data in real time.

[0213] A data transmission module 72 is configured to transmit node operation data to a regional coordinator in a regional layer of a corresponding region, so that the regional coordinator performs load forecasting within the region based on the node operation data and a pre-trained neural network model, obtains a forecast result, determines a dynamic scheduling weight based on the forecast result, the node operation data, a preset weight vector, and a preset feedback gain coefficient, performs a solution calculation based on preset solution parameters and a pre-built objective function of multiple pre-stored node mapping schemes, obtains a multi-objective Pareto solution set, performs optimal processing based on the Pareto solution set and the dynamic scheduling weight, and obtains a target node mapping scheme;

[0214] The receiving module 73 is configured to receive a target node mapping solution sent by a regional coordinator of a corresponding region, and perform a load balancing collaborative operation according to the target node mapping solution.

[0215] In an optional embodiment of the present application, the receiving module 73 is specifically used to: determine the load to be migrated and the load migration route in this area according to the target node mapping scheme; migrate the load to be migrated according to the load migration route to complete the load balancing collaborative operation; or deploy a processing container, and input the load to be migrated into the processing container for processing to complete the load balancing collaborative operation.

[0216] In an optional embodiment of the present application, the data sending module 72 is specifically used to: send the execution feedback results for the load balancing collaborative operation to the regional coordinator of the corresponding area, so that the regional coordinator adjusts the model parameters of the pre-trained neural network model according to the execution feedback results to obtain the adjusted neural network model, and adjusts the algorithm parameters of the preset solution parameters according to the execution feedback results to obtain the adjusted preset solution parameters.

[0217] Figure 7 The description of the features of the server cluster load balancing collaborative processing device in the embodiment can be found in Figure 3 The relevant description of the embodiment of the server cluster load balancing collaborative processing method is not repeated here.

[0218] Figure 8 Schematic diagram of the structure of the server cluster load balancing collaborative processing device provided in the embodiment of the application Figure 3 .like Figure 8 As shown, an embodiment of the present application further provides a server cluster load balancing collaborative processing device, including: a communication bus module 81 and a cross-region decision module 82.

[0219] Among them, the communication bus module 81 is used to receive cross-regional scheduling requests and prediction results sent by each regional coordinator in the regional layer, where the prediction results are obtained by each regional coordinator in the regional layer based on the node operation data and the pre-trained neural network model to predict the load in the region, and the node operation data is collected in real time by each edge node corresponding to each regional coordinator and sent to each regional coordinator.

[0220] The cross-region decision module 82 is used to determine the cross-region load scheduling strategy in response to the cross-region scheduling request based on the prediction results, the whole network resource view and the policy library, wherein the cross-region scheduling request is generated when the regional coordinator detects that the load in the local area exceeds the preset threshold.

[0221] The communication bus module 81 is also used to send the cross-region scheduling strategy to each regional coordinator to complete the cross-region load balancing collaborative operation.

[0222] Figure 9 This is a schematic diagram of the structure of the electronic device provided in this application. Figure 9As shown, the electronic device 90 provided in this embodiment includes: at least one processor 901 and a memory 902. Optionally, the electronic device 90 further includes a communication component 903. The processor 901, the memory 902 and the communication component 903 are connected via a bus.

[0223] During the specific implementation process, at least one processor 901 executes the computer execution instructions stored in the memory 902, so that the at least one processor 901 executes any of the above-mentioned server cluster load balancing collaborative processing method embodiments.

[0224] The specific implementation process of the processor 901 can be found in the above method embodiment. Its implementation principle and technical effects are similar and will not be repeated here in this embodiment.

[0225] In the above embodiments, it should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the application may be directly executed by a hardware processor or by a combination of hardware and software modules within the processor.

[0226] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage.

[0227] A bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. Buses can be categorized as address buses, data buses, and control buses. For ease of illustration, the buses in the drawings of this application are not limited to just one bus or just one type of bus.

[0228] An embodiment of the present application further provides a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps of any of the above-mentioned server cluster load balancing collaborative processing method embodiments when running.

[0229] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0230] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of any of the above-mentioned server cluster load balancing collaborative processing method embodiments are implemented.

[0231] An embodiment of the present application also provides another computer program product, including a non-volatile computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps in any of the above-mentioned server cluster load balancing collaborative processing method embodiments.

[0232] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0233] The above is a detailed introduction to a server cluster load balancing collaborative processing system, method and electronic device provided by the present application. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core ideas of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A server cluster load balancing collaborative processing system, characterized in that: include: edge layer, regional layer, and global layer; The server cluster load balancing collaborative processing system is characterized by comprising: an edge layer, a regional layer and a global layer; The edge layer includes a plurality of edge nodes, which are used to: respond to user-specified tasks and collect node operation data in real time during the execution of tasks; The regional layer includes multiple regional coordinators, each of which is data-connected to at least one corresponding edge node. The regional coordinator is used to: receive node operation data sent by the corresponding edge node, and perform load prediction in the region based on the node operation data and a pre-trained neural network model to obtain a prediction result; determine a dynamic scheduling weight based on the prediction result, the node operation data, a preset weight vector, and a preset feedback gain coefficient; perform a solution calculation based on multiple pre-stored node mapping schemes and pre-built objective functions to obtain a multi-objective Pareto solution set; perform optimal processing based on the Pareto solution set and the dynamic scheduling weight to obtain a target node mapping scheme; and control each edge node in the region to perform load balancing collaborative operations based on the target node mapping scheme. The global layer is data-connected with each regional coordinator in the regional layer. The global layer is used to: receive the cross-regional scheduling request and the prediction result sent by each regional coordinator, and determine the cross-regional load balancing scheduling strategy in response to the cross-regional scheduling request based on the prediction result, the whole network resource view and the policy library, wherein the cross-regional scheduling request is generated when the regional coordinator detects that the load in the local area exceeds a preset threshold; send the cross-regional scheduling strategy to each regional coordinator to complete the cross-regional load balancing collaborative operation.

2. A server cluster load balancing collaborative processing method, characterized in that: include: Receive node operation data sent by each edge node in the local area, wherein the node operation data is collected in real time by the edge node; Perform load forecasting in the region based on the node operation data and the pre-trained neural network model to obtain a forecast result; Determining a dynamic scheduling weight according to the prediction result, the node operation data, a preset weight vector and a preset feedback gain coefficient; Perform calculations based on multiple pre-existing node mapping schemes, preset solution parameters, and pre-built objective functions to obtain a multi-objective Pareto solution set; Performing optimal processing according to the Pareto solution set and the dynamic scheduling weight to obtain a target node mapping solution; According to the target node mapping scheme, each edge node in the region is controlled to perform load balancing collaborative operations.

3. The method according to claim 2, characterized in that The load prediction within the region is performed based on the node operation data and the pre-trained neural network model to obtain a prediction result, including: Performing data preprocessing according to the node operation data to obtain normalized node data; Performing data batch reading processing according to the normalized node data in a preset window size to obtain data to be predicted for multiple periods arranged in time sequence; The data to be predicted in each cycle is input into the pre-trained neural network model in turn, and a predicted load curve including a confidence interval is obtained as the prediction result.

4. The method according to claim 2, characterized in that The preset weight vector includes a delay weight, an energy consumption weight and a plurality of different node operation weights; Accordingly, determining the dynamic scheduling weight according to the prediction result, the node operation data, the preset weight vector and the preset feedback gain coefficient includes: determining, based on the node operation data, actual delay and actual energy consumption of a plurality of cycles arranged in time sequence; Determining, according to the prediction result of the current cycle, a predicted delay of the current cycle and a predicted energy consumption of the current cycle; Determining a delay error based on the predicted delay of the current cycle and the actual delay, and determining an energy consumption error based on the actual energy consumption of the current cycle and the predicted energy consumption; Determining an adjusted delay weight according to the delay error, the delay weight in the preset weight vector, and the preset feedback gain parameter; Determining an adjusted energy consumption weight according to the energy consumption error, the energy consumption weight in the preset weight vector and the preset feedback gain parameter; Weight adjustment is performed according to the adjusted delay weight, the adjusted energy consumption weight and a plurality of different node operation weights to obtain an adjusted weight vector, and normalization is performed according to the adjusted weight vector to obtain a dynamic scheduling weight.

5. The method according to claim 2, characterized in that The pre-built objective functions include an average response delay objective function, a network bandwidth occupancy objective function, and a total energy consumption objective function, and the preset solution parameters include population size, number of iterations, crossover rate, and mutation rate; Accordingly, the multi-objective Pareto solution set is obtained by performing solution calculations based on multiple pre-stored node mapping schemes, preset solution parameters, and pre-built objective functions, including: Establish multiple chromosome codes according to each pre-existing node mapping scheme; The response delay objective function, the network bandwidth occupancy objective function and the total energy consumption objective function are taken as three functions to be solved, and an iterative solution process is performed through a non-dominated sorting genetic algorithm based on the three functions to be solved, the multiple chromosome codes, the population size, the number of iterations, the crossover rate and the mutation rate to obtain a multi-objective Pareto solution set, wherein the multi-objective Pareto solution set includes at least one pre-existing node mapping scheme.

6. The method according to claim 5, characterized in that The performing optimal processing according to the Pareto solution set and the dynamic scheduling weight to determine the target node mapping scheme includes: Calculating a solution score according to the at least one pre-stored node mapping solution and the dynamic scheduling weight to obtain a node score for each pre-stored node mapping solution; According to the node scores of the pre-existing node mapping solutions, the pre-existing node mapping solution corresponding to the lowest node score is selected and determined as the target node mapping solution.

7. The method according to claim 2, characterized in that The controlling the edge nodes in the region to perform load balancing collaborative operations according to the target node mapping scheme includes: Determine the load to be migrated and the load migration route in the region according to the target node mapping scheme; The load to be migrated and the load migration route are sent to each edge node in this area, so that each edge node migrates the load to be migrated according to the load migration route to complete the load balancing collaborative operation, or deploys a processing container and inputs the load to be migrated into the processing container for processing to complete the load balancing collaborative operation.

8. The method according to any one of claims 2 to 7, characterized in that Also includes: Determine the regional load corresponding to each region based on the node operation data; When it is detected that the regional load corresponding to any area exceeds a preset threshold, a cross-region scheduling request is generated; Sending the cross-region scheduling request and the prediction results corresponding to each region to the global layer, so that the global layer determines the cross-region load balancing scheduling strategy in response to the cross-region scheduling request based on the prediction results, the whole network resource view and the strategy library; Receive the cross-region load balancing scheduling strategy sent by the global layer, and complete the cross-region load balancing collaborative operation according to the cross-region load balancing scheduling strategy.

9. The method according to claim 8, characterized in that After the cross-region load coordination operation is completed according to the cross-region load balancing scheduling strategy, the method further includes: receiving execution feedback results of the cross-region load collaborative operation from each edge node in each region; Adjusting model parameters of the pre-trained neural network model according to the execution feedback result to obtain an adjusted neural network model; Algorithm parameters are adjusted for the preset solution parameters according to the execution feedback result to obtain adjusted preset solution parameters.

10. A server cluster load coordination processing method, characterized in that: include: Real-time collection of node operation data; Sending the node operation data to a regional coordinator in a regional layer of a corresponding region, so that the regional coordinator performs load forecasting in the region based on the node operation data and a pre-trained neural network model to obtain a forecast result; determining a dynamic scheduling weight based on the forecast result, the node operation data, a preset weight vector, and a preset feedback gain coefficient; performing a solution calculation based on preset solution parameters and a pre-built objective function of a plurality of pre-stored node mapping schemes to obtain a multi-objective Pareto solution set; performing optimal processing based on the Pareto solution set and the dynamic scheduling weight to obtain a target node mapping scheme; Receive the target node mapping plan sent by the regional coordinator of the corresponding area, and perform the load balancing collaborative operation according to the target node mapping plan.

11. The method according to claim 10, characterized in that The performing of the load balancing collaborative operation according to the target node mapping scheme includes: Determine the load to be migrated and the load migration route in the region according to the target node mapping scheme; Migrating the load to be migrated according to the load migration route to complete the load balancing collaborative operation; or A processing container is deployed, and the load to be migrated is input into the processing container for processing to complete a load balancing collaborative operation.

12. The method according to claim 10, characterized in that After performing the load balancing collaborative operation according to the target node mapping scheme, the method further includes: The execution feedback results of the load balancing collaborative operation are sent to the regional coordinator of the corresponding area, so that the regional coordinator adjusts the model parameters of the pre-trained neural network model according to the execution feedback results to obtain the adjusted neural network model, and adjusts the algorithm parameters of the preset solution parameters according to the execution feedback results to obtain the adjusted preset solution parameters.

13. A server cluster load balancing collaborative processing method, characterized in that: include: Receive cross-regional scheduling requests and prediction results sent by each regional coordinator in the regional layer, where the prediction results are obtained by each regional coordinator in the regional layer based on node operation data and a pre-trained neural network model to predict the load in the region, and the node operation data is collected in real time by each edge node corresponding to each regional coordinator and sent to each regional coordinator; Determining a cross-region load scheduling strategy in response to the cross-region scheduling request based on the prediction result, the network-wide resource view, and the strategy library, wherein the cross-region scheduling request is generated when the regional coordinator detects that the load in the local region exceeds a preset threshold; The cross-region scheduling strategy is sent to each regional coordinator to complete the cross-region load balancing collaborative operation.

14. A server cluster load balancing collaborative processing method, characterized in that: include: Each edge node in each area of ​​the edge layer collects node operation data in real time; Each edge node sends the node operation data to the corresponding regional coordinator of the regional layer; Each regional coordinator performs load forecasting within the region based on the node operation data and the pre-trained neural network model to obtain a forecast result, and determines a dynamic scheduling weight based on the forecast result, the node operation data, a preset weight vector, and a preset feedback gain coefficient; Each regional coordinator performs a solution calculation based on multiple pre-stored node mapping schemes, preset solution parameters, and pre-built objective functions to obtain a multi-objective Pareto solution set, and performs optimal processing based on the Pareto solution set and the dynamic scheduling weight to obtain a target node mapping scheme; Each regional coordinator controls each edge node in the region to perform load balancing collaborative operations according to the target node mapping scheme; Each regional coordinator determines the regional load corresponding to each region based on the node operation data, and generates a cross-region scheduling request when detecting that the regional load corresponding to any region exceeds a preset threshold; Each regional coordinator sends the cross-region scheduling request and the prediction results corresponding to each region to the global layer; The global layer determines, in response to the cross-region scheduling request, a cross-region load balancing scheduling strategy based on the prediction result, the entire network resource view, and the strategy library; The global layer sends a cross-region load balancing scheduling strategy to the regional coordinator corresponding to each region of the regional layer; The regional coordinators corresponding to the respective regions complete cross-region load balancing collaborative operations according to the cross-region load balancing scheduling strategy.

15. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the server cluster load balancing collaborative processing method as described in any one of claims 2 to 9, any one of 10 to 12, 13 or 14 when executing the computer program.

Images