Computer resource allocation management method and system based on deep learning

Abstract

The invention discloses a computer resource allocation management method and system based on deep learning, and the method comprises the steps: obtaining multi-source heterogeneous data, constructing a system topological graph based on the multi-source heterogeneous data, enabling server nodes, virtual machine instances and network connections to serve as nodes of the graph, and enabling a connection relation between the nodes to serve as edges, thereby obtaining a dynamic graph sequence; constructing a GCN-GAT mixed graph neural network model based on a GCN graph convolutional network and a GAT graph attention network, and adaptively learning an important relationship between nodes by using multilayer graph convolution and an attention mechanism to obtain a target GCN-GAT mixed graph neural network model; and establishing a multi-objective optimization function according to the prediction index and the current system topological structure, performing resource allocation decision through an improved DQN deep reinforcement learning algorithm, and outputting an optimal allocation strategy. The resource demand prediction error is reduced, and the resource utilization rate is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer resource management, in particular to a computer resource allocation management method and system based on deep learning. BACKGROUND

[0002] With the rapid development of cloud computing and big data technology, efficient allocation and management of computer resources has become a key problem. Traditional resource allocation methods mainly rely on static rules or simple threshold judgments, which cannot accurately predict changes in business load, resulting in low resource utilization, high response delay, increased cost and other problems. Although there are some resource prediction methods based on machine learning in the prior art, these methods usually only focus on single-dimensional prediction, lack comprehensive analysis capability of multi-dimensional and multi-temporal characteristics, and are difficult to adapt to dynamic business scenarios. Unable to capture the resource demand law under complex business scenarios, the efficiency of computer resource allocation management is not high. SUMMARY

[0003] The purpose of the present application is to solve the above problems, and a computer resource allocation management method and system based on deep learning are designed.

[0004] To achieve the above purpose, the technical scheme of the present application is as follows: further, in the computer resource allocation management method based on deep learning, the computer resource allocation management method comprises the following steps:

[0005] Obtain multi-source heterogeneous data, construct a system topology graph based on the multi-source heterogeneous data, take server nodes, virtual machine instances and network connections as nodes of the graph, take the connection relationship between nodes as edges, and obtain a dynamic graph sequence;

[0006] Construct a GCN-GAT hybrid graph neural network model based on a GCN graph convolution network and a GAT graph attention network, use multi-layer graph convolution and attention mechanism to adaptively learn the important relationship between nodes, and obtain a target GCN-GAT hybrid graph neural network model;

[0007] Input the dynamic graph sequence into the target GCN-GAT hybrid graph neural network model, and output a prediction index of each type of resource demand;

[0008] Establish a multi-objective optimization function according to the prediction index and the current system topology structure, make resource allocation decisions through an improved DQN deep reinforcement learning algorithm, and output an optimal allocation strategy.

[0009] Further, in the computer resource allocation management method based on deep learning, the multi-source heterogeneous data is obtained, the system topology graph is constructed based on the multi-source heterogeneous data, the server nodes, virtual machine instances and network connections are taken as the nodes of the graph, the connection relationship between the nodes is taken as the edges, and the dynamic graph sequence is obtained:

[0010] acquiring multi-source heterogeneous data, including at least historical resource usage data, service load data, system state data, network topology data and user behavior data;

[0011] adopting linear interpolation to complete short-term missing data in the multi-source heterogeneous data, and deleting long-term missing abnormal data;

[0012] scaling the data to the interval [0, 1] after removing extreme values beyond the normal range by the 3σ principle to obtain preprocessed data.

[0013] Further, in the above computer resource allocation management method based on deep learning, the multi-source heterogeneous data is acquired, and a system topology graph is constructed based on the multi-source heterogeneous data, taking server nodes, virtual machine instances and network connections as nodes of the graph, and taking connection relationships between nodes as edges, to obtain a dynamic graph sequence:

[0014] The server nodes include CPU core number, memory capacity, residual bandwidth and health status, the virtual machine instances include allocated CPU, current load rate and belonging service type, and the network connection nodes include bandwidth upper limit, real-time utilization rate and connection stability;

[0015] The server and virtual machine edges represent that the virtual machine is deployed on the server, the server and server edges represent that two servers are connected through the network, and the virtual machine and service edges represent that the virtual machine supports the service, to construct the system topology graph;

[0016] Based on the system topology graph, a time window is divided by 10 minutes per slice, a static graph is generated for each slice, the sequence length is fixed, and a dynamic graph sequence is obtained.

[0017] Further, in the above computer resource allocation management method based on deep learning, the GCN-GAT hybrid graph neural network model is constructed based on the GCN graph convolution network and the GAT graph attention network, the important relationship between nodes is adaptively learned by using multi-layer graph convolution and attention mechanism, and the target GCN-GAT hybrid graph neural network model is obtained:

[0018] The input layer of the hybrid graph neural network model is used to receive the feature matrix and the adjacency matrix of the dynamic graph sequence;

[0019] The feature fusion layer splices the sequence features in the time dimension and the topology features in the space dimension, and outputs a fusion feature matrix, and the full connection layer maps the extracted features to the resource demand prediction dimension.

[0020] Further, in the above computer resource allocation management method based on deep learning, the GCN graph convolution network and the GAT graph attention network are used to construct a GCN-GAT hybrid graph neural network model, and the important relationship between nodes is adaptively learned by using multi-layer graph convolution and attention mechanism to obtain a target GCN-GAT hybrid graph neural network model:

[0021] The first layer of the GCN graph convolution network extracts local topological features, and the second layer fuses global topological features.

[0022] The attention mechanism of the GAT graph attention network calculates the attention weight of each node's neighbor node, and the multi-head attention sets 3 attention heads to improve the robustness of feature extraction.

[0023] Further, in the above computer resource allocation management method based on deep learning, the dynamic graph sequence is input into the target GCN-GAT hybrid graph neural network model, and the prediction index of each type of resource demand is output, including:

[0024] 12 pieces of data are intercepted from the dynamic graph sequence as input of the target GCN-GAT hybrid graph neural network model, and the model outputs the prediction index of each type of resource demand in the future 30 minutes, 1 hour and 2 hours.

[0025] Further, in the above computer resource allocation management method based on deep learning, the multi-objective optimization function is established according to the prediction index and the current system topology structure, the resource allocation decision is made by using the improved DQN deep reinforcement learning algorithm, and the optimal allocation strategy is output, including:

[0026] The double DQN structure is set, the evaluation network and the target network of the DQN deep reinforcement learning algorithm are set separately, and the improved DQN deep reinforcement learning algorithm is obtained.

[0027] The current system topology structure, the prediction index and the allocated resource data are obtained, the state vector is constructed, the action space is defined, including the resource adjustment action and the topology adjustment action, and the optimal allocation strategy is output.

[0028] Further, in the computer resource allocation management system based on deep learning, the computer resource allocation management system comprises the following modules:

[0029] The dynamic graph sequence acquisition module is used to acquire multi-source heterogeneous data, construct a system topology graph based on the multi-source heterogeneous data, take server nodes, virtual machine instances and network connections as nodes of the graph, take the connection relationship between nodes as edges, and obtain a dynamic graph sequence.

[0030] The hybrid network model establishing module is configured to establish a GCN-GAT hybrid graph neural network model based on a GCN graph convolution network and a GAT graph attention network, to use multi-layer graph convolution and attention mechanism to adaptively learn the important relationships between nodes, and to obtain a target GCN-GAT hybrid graph neural network model.

[0031] The resource demand prediction module is configured to input the dynamic graph sequence into the target GCN-GAT hybrid graph neural network model, and to output a prediction index of each type of resource demand.

[0032] The allocation strategy generation module is configured to establish a multi-objective optimization function according to the prediction index and the current system topology structure, to make a resource allocation decision by using an improved DQN deep reinforcement learning algorithm, and to output an optimal allocation strategy.

[0033] Further, in the deep learning-based computer resource allocation management system, the dynamic graph sequence acquisition module comprises the following sub-modules:

[0034] The acquisition sub-module is configured to acquire multi-source heterogeneous data, including at least historical resource usage data, service load data, system state data, network topology data, and user behavior data.

[0035] The deletion sub-module is configured to use linear interpolation to complete short-term missing data in the multi-source heterogeneous data, and to delete long-term missing abnormal data.

[0036] The scaling sub-module is configured to scale the data to the interval [0, 1] after removing extreme values that exceed the normal range by the 3σ principle, to obtain preprocessed data.

[0037] Further, in the deep learning-based computer resource allocation management system, the allocation strategy generation module comprises the following sub-modules:

[0038] The setting sub-module is configured to set a double-DQN structure, to separately set an evaluation network and a target network of the DQN deep reinforcement learning algorithm, and to obtain an improved DQN deep reinforcement learning algorithm.

[0039] The output sub-module is configured to acquire current system topology structure, prediction index, and allocated resource data, to construct a state vector, to define an action space including resource adjustment actions and topology adjustment actions, and to output an optimal allocation strategy.

[0040] The beneficial effect is that the system topology and the relationship between nodes can be effectively represented through the graph neural network model, and the resource demand prediction error is reduced by more than 38% compared with traditional methods. The resource utilization is greatly improved, the allocation strategy considering the system topology makes the resource allocation more reasonable, the resource utilization is improved, and the key connection overload is effectively reduced. The system response speed is accelerated, the forward-looking resource allocation based on the topology structure reduces the key business delay by more than 25%, and the overall performance of the system is improved. The system stability is enhanced, and by considering the connection relationship between nodes, the key path overload is avoided, and the system stability is improved. BRIEF DESCRIPTION OF DRAWINGS

[0041] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to be limiting of the present application.

[0042] Figure 1 The first embodiment of the computer resource allocation management method based on deep learning in the embodiments of the present application is shown in the schematic diagram.

[0043] Figure 2 The second embodiment of the computer resource allocation management method based on deep learning in the embodiments of the present application is shown in the schematic diagram.

[0044] Figure 3 The first embodiment of the computer resource allocation management system based on deep learning in the embodiments of the present application is shown in the schematic diagram. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical scheme and advantages of the present application clearer, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0046] Those skilled in the art can understand that, unless specifically stated, the singular form "one", "an" and "the" used herein can also include the plural form. It should be further understood that the wording used in the specification of the present application includes the presence of features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups.

[0047] The present application will be specifically described below in combination with the drawings, Figure 1 The computer resource allocation management method based on deep learning is shown, which includes the following steps:

[0048] In step 101, multi-source heterogeneous data is acquired, a system topology graph is constructed based on the multi-source heterogeneous data, server nodes, virtual machine instances and network connections are taken as nodes of the graph, and connection relationships between the nodes are taken as edges, to obtain a dynamic graph sequence. Specifically, in the embodiment, the multi-source heterogeneous data acquired at least includes historical resource usage data, service load data, system state data, network topology data and user behavior data. Short-term missing data in the multi-source heterogeneous data is completed by linear interpolation, and long-term missing abnormal data is deleted. After extreme values exceeding a normal range are removed by a 3σ principle, the data is scaled to an interval of [0, 1] to obtain preprocessed data. The server nodes include CPU core number, memory capacity, residual bandwidth and health status, the virtual machine instances include allocated CPU, current load rate and belonging service type, and the network connection nodes include bandwidth upper limit, real-time utilization rate and connection stability. The server and virtual machine edges represent that the virtual machine is deployed on the server, the server and server edges represent that two servers are connected through the network, the virtual machine and service edges represent that the virtual machine supports the service, and the system topology graph is constructed. Based on the system topology graph, a time window is divided at 10 minutes per slice, a static graph is generated every slice, a fixed sequence length is determined, and the dynamic graph sequence is obtained.

[0049] The historical resource usage data is used as a historical basis for resource allocation. The core indicators of CPU usage, memory occupancy, disk IO and bandwidth utilization directly reflect the resource consumption law of the server. Prometheus or Zabbix is selected as the data source because it has the ability of multi-node monitoring, index visualization and alarm. The core business collection frequency is set to 1 minute / time, and the non-core business is set to 5 minutes / time. This meets the high real-time demand of core business for resource changes, and avoids resource waste caused by frequent collection of non-core business. The business load data is related to the business orientation of resource allocation. The business request volume, request response time and task priority can accurately depict the change of business pressure. Relying on the log collection and analysis capabilities of ELKStack, Elasticsearch, Logstash and Kibana, synchronous collection with business requests is realized to ensure that data can match business dynamics in real time. The system state data is related to the stability bottom line of resource allocation. The server health status, virtual machine running state and network connectivity are the key to determine whether the system has resource scheduling conditions. Through the system kernel interface, / proc can directly obtain the underlying hardware and process information. Combined with the cloud platform API, the unified management of virtual resources is realized. The real-time monitoring frequency of 30 seconds / time can quickly find server failures, virtual machine downtime and other abnormalities to gain time for resource adjustment. The network topology data affects the network feasibility of resource allocation. The server IP topology, switch connection relationship and bandwidth upper limit determine the network path and transmission capacity of cross-node resource scheduling. SolarWinds, as a professional network management tool, can accurately draw the topology structure. The regular collection of 1 hour / time combined with real-time updating when the topology changes can reduce repeated calculation and timely synchronize the topology changes such as the addition of new servers and switch expansion. User behavior data supports demand prediction for resource allocation. User access period, business operation type and resource application frequency can mine the time distribution law of resource demand. Application logs record user operation details, and Baidu statistics supplement macro access features. The collection frequency of 5 minutes / time can capture user behavior trends while controlling data storage and processing costs. In the data preprocessing stage, for short-term missing, linear interpolation is used to complete <3 collection periods. This is because such missing data is mostly caused by data continuity breakage due to collection delay, and linear interpolation can maximize the retention of data trends. Long-term missing, >10 collection periods of abnormal samples are deleted to avoid interference with model training due to insufficient data integrity. All indicators are scaled to the [0, 1] interval for feature normalization. The core is to solve the problem of different index magnitude differences. For example, the numerical range of memory occupancy, unit: GB and request volume, unit: times is very different. Without normalization, the model will be biased towards the index with large magnitude, affecting the prediction accuracy.

[0050] Dynamic graph sequence construction is a key link to convert multi-source heterogeneous data into a format that can be processed by a graph neural network. The core is to express the association between resources, nodes and services through a structured topology, and to form a dynamic sequence combined with the time dimension. In the definition of the graph structure, the design of node types and attributes needs to accurately map the actual system composition: the attributes of the CPU core number, memory capacity, etc. of the server node reflect the hardware resource endowment, and the remaining bandwidth and health status, 0 = normal, 1 = warning, 2 = failure directly determine the ability of the node to participate in resource allocation, and the numerical coding of the health status facilitates the model to quickly identify the availability of the node; the allocated CPU / memory, current load rate, etc. of the virtual machine instance can directly present the resource usage status, and the type of the business to which it belongs establishes the association between the virtual machine and the business, providing a basis for resource scheduling based on business priority; the bandwidth upper limit, real-time utilization and connection stability of the network connection node, 0 = stable, 1 = fluctuation, then supplement the network constraints in the resource transmission process to avoid resource scheduling failure due to network bottlenecks. The definition and weight design of the edge further quantify the association strength between nodes: the weight of the server-virtual machine edge is set to the virtual machine resource proportion, such as the allocated CPU / total CPU of the server, which can allow the model to clearly identify the resource allocation density of a single server and avoid excessive concentration of resources; the weight of the server-server edge is set to the network bandwidth utilization rate, which can reflect the network cost of cross-server scheduling, and when the bandwidth utilization rate is too high, the model can reduce cross-node migration to avoid network congestion; the weight of the virtual machine-business edge is set to the business request proportion, which can ensure that the virtual machine corresponding to the core business is given priority to resource support to prevent business interruption. In the dynamic sequence generation link, a static graph is generated by dividing the time window into 10 minutes per piece, which is the result of considering the timeliness and computational efficiency of the data; a time window that is too short, such as 1 minute, will result in an explosive increase in the number of graphs, increasing the cost of model training and inference; a time window that is too long, such as 30 minutes, will lose the detailed features of resource changes; a fixed sequence length of 12, i.e. 2 hours of historical data, is because a time span of 2 hours can cover both short-term resource fluctuation patterns and provide sufficient historical basis for resource demand prediction; and when the system topology changes, such as the addition of a server or the migration of a virtual machine, an immediate update is triggered to ensure that the dynamic graph sequence always matches the actual system structure, avoiding input distortion caused by topology lag and laying a solid foundation for the subsequent feature extraction and resource prediction of the graph neural network.

[0051] Step 102, based on the GCN graph convolution network and the GAT graph attention network, a GCN-GAT hybrid graph neural network model is constructed, and the important relationship between nodes is learned adaptively by using multi-layer graph convolution and attention mechanism to obtain a target GCN-GAT hybrid graph neural network model;

[0052] Specifically, in the embodiment, the input layer of the mixed graph neural network model is configured to receive the feature matrix and the adjacency matrix of the dynamic graph sequence; the feature fusion layer is configured to splice the sequence feature in the time dimension and the topological feature in the space dimension, and output a fused feature matrix; and the full connection layer is configured to map the extracted feature to the resource demand prediction dimension. The first layer of the GCN graph convolution network extracts local topological features, and the second layer fuses global topological features; the attention mechanism of the GAT graph attention network calculates attention weights for neighbor nodes of each node, and the multi-head attention sets three attention heads to improve the robustness of feature extraction.

[0053] The overall structure design of the GCN-GAT hybrid graph neural network model closely fits the feature extraction requirements of the dynamic graph sequence. The functions of each layer are progressive, realizing accurate conversion from data input to prediction output. The dynamic graph sequence feature matrix and adjacency matrix received by the input layer are the basis for the model to perceive the system state. The feature matrix integrates the core attributes of the three types of nodes, such as server CPU core number, virtual machine load rate, and the adjacency matrix quantifies the association strength between nodes, such as server-virtual machine resource occupancy weight. Together, they build a data entry for the model to understand the system topology and resource distribution. The feature fusion layer concatenates the time dimension sequence features and the space dimension topology features. The core is to solve the collaborative utilization problem of dynamic graph time and space attributes: the time dimension features reflect the trend of resource changes over time, such as the fluctuation of CPU usage in 12 time slices, and the space dimension features reflect the static association between nodes, such as the deployment relationship between servers and virtual machines. The generated feature matrix after fusion can capture the resource features of when and where, laying a comprehensive data foundation for subsequent feature extraction. The GCN layer is set to 2 layers based on the gradient transmission efficiency of topology feature extraction: the first layer focuses on local topology features, capturing the distribution density and resource occupancy of virtual machines around a single server through convolution operation, allowing the model to understand the small-scale resource pattern first; the second layer further integrates global topology features, extending the local features to the resource distribution and network connection state of the entire server cluster, avoiding the model from being trapped in a local perspective; the ReLU activation function not only effectively alleviates the gradient vanishing problem of deep network, but also improves the model's expression ability for complex topology relationships by introducing non-linear transformation. The 1-layer design of the GAT layer is to strengthen the feature weight of key nodes: the attention mechanism calculates the attention degree of each node to its neighbor nodes, such as the connection server of high-load virtual machine and the network node with high bandwidth utilization, which will obtain higher weight, allowing the model to automatically focus on core nodes that have greater impact on resource demand, avoiding irrelevant node interference; setting 3 attention heads to extract features in parallel and then fusing the results through average pooling can not only reduce the feature deviation of a single attention head and improve the robustness of feature extraction, but also cover more node association scenarios; the fully connected layer in the final output layer maps the multi-dimensional features extracted before to the CPU, memory, and bandwidth resource demand prediction dimensions. These three dimensions are the core objects of computer resource allocation, ensuring that the prediction results can directly serve subsequent resource scheduling decisions.

[0054] Important relations refer to the inter-node correlations in the dynamic graph sequence that play a key role in predicting resource demand. Technically, the GCN graph convolution network focuses on the local topology in the first layer to capture the local correlations such as the distribution of virtual machines around a single server, and then fuses the global topology in the second layer to grasp the macro correlations such as the resource distribution of the server cluster. The GAT graph attention network calculates the attention weight of each node to the neighbor node by means of the attention mechanism, so that the model focuses on the core nodes such as the servers connected to the high-load virtual machines, and at the same time, three attention heads are set in parallel to extract features, reduce bias, and adaptively learn the important relations that affect resource demand.

[0055] The design of the model training and optimization link focuses on how to make the model stable learning and adapt to the actual production environment. Each step is around the accuracy and practicability. The 12-month historical data is selected in the training data preparation stage to cover different business cycles, such as holiday high load and daily smooth load resource characteristics, to avoid the lack of data timeliness leading to poor model generalization ability. The training set, validation set and test set are divided according to 7:2:1, which not only ensures that the training set has enough data to make the model fully learn, and 70% of the data, but also realizes the parameter adjustment in the training process through the validation set, 20% of the data, and objectively evaluates the generalization ability of the model on the unseen data through the test set, 10% of the data, to avoid model overfitting to training data. The label is defined as the actual resource usage peak in the future 30 minutes, which is designed in combination with the execution cycle of resource adjustment. A 30-minute time window can give the model enough prediction lead time to allow the operation and maintenance or automatic scheduling system to have time to execute the resource allocation strategy, and can also ensure the timeliness of the prediction target to avoid large deviation between demand and actual value due to too long prediction window. The selection of training parameters and strategies takes into account the convergence efficiency and model stability: the mean square error, MSE, is used as the loss function because it is sensitive to the deviation between the predicted value and the actual value, can accurately measure the error of resource demand prediction, and meets the accuracy requirements of resource allocation; the Adam optimizer is more suitable for training complex graph models than the traditional SGD optimizer due to its adaptive learning rate characteristics, an initial learning rate of 0.001 can ensure rapid convergence in the early stage, and a decay to 0.8 times of the previous value every 5 epochs can fine-tune the parameters in the later training stage to avoid oscillation; L2 regularization with a weight decay coefficient of 0.0001 can prevent overfitting by limiting the absolute value of the model weights; the early stopping strategy stops training when the validation set loss does not decrease for 3 consecutive epochs, which can save the optimal model that has not been overfitted and avoid wasting computing resources on invalid training. TensorFlow Serving is selected for model deployment and update because it supports high-concurrency and low-latency API service calls, which can meet the real-time resource prediction requirements in production environment. Incremental training is performed every week using new historical data to avoid the resource waste of full retraining and to enable the model to learn the latest resource variation patterns such as the load characteristics of new businesses, ensuring long-term prediction accuracy stability.

[0056] Step 103, input the dynamic graph sequence into the target GCN-GAT hybrid graph neural network model to output the prediction index of each type of resource demand;

[0057] Specifically, in this embodiment, 12 pieces of data are extracted from the dynamic graph sequence as the input of the target GCN-GAT hybrid graph neural network model, and the model outputs the prediction index of each type of resource demand in the future 30 minutes, 1 hour and 2 hours.

[0058] Input data preprocessing is the basis for ensuring prediction accuracy. The sequence extraction link selects the latest 12 pieces of data from the real-time dynamic graph sequence, corresponding to 2 hours. This design is fully matched with the time slice standard of 10 minutes / piece of dynamic graph sequence and the input length during model training. The 12 pieces of data can not only completely cover the historical topology and resource change characteristics required by the model, but also avoid the increase in calculation cost caused by feature missing or data redundancy due to insufficient data, ensuring the adaptability of the input format to the model training and laying a format foundation for prediction accuracy. The feature alignment link emphasizes the use of the mean and standard deviation of the training set to ensure consistent normalization standards. The core is to solve the scale unification problem of real-time data and training data. During model training, all features are normalized in the [0, 1] interval based on the statistical parameters of the training set. If the real-time data uses new statistical parameters such as its own mean and standard deviation, it will cause the feature magnitude to deviate from the feature distribution learned by the model, and then cause the prediction result to be distorted. Therefore, by reusing the statistical parameters of the training set, the scale of the real-time input features and the features recognized by the model can be ensured to be consistent, eliminating the prediction error caused by the difference in normalization standards. The anomaly detection link performs the operation of excluding faulty nodes and reconstructing temporary topology graphs for servers with a state = 2. This is because the resource data of the faulty node, such as CPU usage and memory occupation, has lost its reference value, and its connection relationship may be interrupted due to failure. If it is retained, it will cause the topology graph to be unable to reflect the true available state of the system; when reconstructing the temporary topology graph, it will automatically retain the nodes with a normal state = 0 and a warning state = 1 and their connection relationships, ensuring that the topology structure input into the model is fully matched with the current system available resource distribution, avoiding invalid data interference with the prediction result.

[0059] Step 104, a multi-objective optimization function is established according to the prediction index and the current system topology structure, resource allocation decisions are made through an improved DQN deep reinforcement learning algorithm, and an optimal allocation strategy is output.

[0060] Specifically, in the present embodiment, a double-DQN structure is provided, and the evaluation network and the target network of the DQN deep reinforcement learning algorithm are separately set to obtain an improved DQN deep reinforcement learning algorithm; the current system topology structure, the prediction index, and the allocated resource data are obtained, a state vector is constructed, an action space is defined, including resource adjustment actions and topology adjustment actions, and an optimal allocation strategy is output.

[0061] The construction of multi-objective optimization function is the core criterion of resource allocation decision, which needs to find a balance between efficient use of resources and stable operation of business, and its target definition and integration method are in line with the actual system management needs. Core target 1 is set to minimize the resource waste rate, which is less than 10%, because resource waste will directly increase operating costs. If the allocated resources far exceed the actual use, the idle CPU and memory will cause hardware resource waste. The waste rate calculation formula is allocated resources-actual use resources / allocated resources, which can directly quantify the degree of waste, and the target value of 10% is set by combining the industry average level and system redundancy demand, which can avoid excessive pursuit of zero waste leading to insufficient resource flexibility, and prevent high waste from increasing costs. Core target 2 focuses on minimizing business delay, which distinguishes core business and non-core business, and the target value is less than 50ms and 200ms, because different businesses have different sensitivity to delay: core businesses such as transaction payment and real-time data analysis will directly affect user experience or business accuracy if the delay is too high, and the target of 50ms can meet the real-time needs; non-core businesses such as log backup and offline statistics are more tolerant to delay, and the target of 200ms can give resources to core businesses without affecting business. Core target 3 aims to maximize resource utilization, with a range of 50%~80%, rather than pursuing 100%, because high utilization, such as more than 90%, will lead to insufficient system resource margin, which is easy to cause lag and crash in the event of sudden load, and the lower limit of 50% avoids low utilization, which causes resource idling, and achieves efficient and stable utilization. In terms of target integration method, a weighted sum function is used, and the optimization value = α*(1-waste rate)+β*(1-delay / maximum allowed delay)+γ*utilization, where * represents multiplication, α represents the weight coefficient of resource waste rate in the overall optimization target, which reflects the importance of resource waste control, the larger the α value, the more attention is paid to reducing resource waste in resource allocation decision-making, in order to reduce operating costs and avoid waste of idle hardware resources; γ represents the weight coefficient of resource utilization in the overall optimization target, which reflects the importance of pursuing efficient use of resources, the larger the γ value, the more efficient use of resources is expected in the decision-making process, and the idle resources are avoided under the premise of ensuring the stable operation of the system; β is the weight coefficient of business delay in the overall optimization target, which reflects the consideration degree of business response speed, the larger the β value, the more attention is paid to reducing business delay in resource allocation. The weight allocation reflects the priority of the target: in core business, β=0.5 is the highest, because delay has the greatest impact on core business; α=0.3, γ=0.2 considers both resource waste and utilization; in non-core business, α=0.4, γ=0.4 is higher than β=0.2, because non-core business values resource cost control more and delay can be compromised.The constraints, CPU / memory ≤ 90%, bandwidth ≤ 85%, and migration ≤ 5 times / hour, are the safety boundaries for decision-making: the upper limit of resource utilization prevents system overload, and the limit of the number of virtual machine migrations avoids service interruption or network fluctuations caused by frequent migration, ensuring that the decision is executed within a safe range.

[0062] The improved DQN deep reinforcement learning algorithm improves decision-making accuracy and efficiency through technical optimization, and the decision execution process and monitoring mechanism ensure that the algorithm is stable in practical scenarios. The algorithm improvement points specifically address the shortcomings of traditional DQN: priority experience replay prioritizes decision samples with large errors, such as a waste rate > 20%, into the experience pool, because these samples contain model learning errors, and priority learning can quickly correct model bias and improve training efficiency; the double DQN structure separates the evaluation network and the target network to avoid the overestimation of action value caused by the same network evaluating actions and updating target values in traditional DQN; the reward function design, reward = optimization value - 0.1 * migration cost, considers both optimization effect and execution cost, and deducting 0.1 for each migration prevents the model from frequently migrating to pursue high optimization values, balancing decision benefits and costs. The design of the decision execution process meets the real-time scheduling requirements: state initialization integrates the current topology, predicted demand, and allocated resources to ensure that the model has a complete understanding of the system; in the action space definition, CPU adjustment step size is 0.5 core and memory is 1 GB, because these are common resource adjustment granularities for cloud platforms, which can accurately adapt to demand changes and avoid frequent decisions caused by small step sizes; executing a decision every 10 minutes is the result of balancing real-time and computational cost; the experience pool capacity is 10,000, and 32 samples are extracted and updated each time, which can store enough historical experience and ensure training efficiency through small batch samples; the model convergence standard, reward ≥ 0.8 for 5 consecutive times, ensures that the optimal action output has stable effects. The decision monitoring and adjustment mechanism forms a closed-loop optimization: resource usage is checked every 2 minutes to quickly identify abnormalities after decision implementation, such as utilization exceeding 90%, triggering emergency adjustments to prevent problems from escalating; decision effectiveness is analyzed and parameters are adjusted every week, because system resource demand changes with business changes, such as new business and user growth, and regular iteration allows the algorithm to continuously adapt to actual scenarios, ensuring the long-term effectiveness of resource allocation decisions.

[0063] The beneficial effects are that, through the graph neural network model, the system topology structure and the relationship between nodes can be effectively represented, the resource demand prediction error is reduced by more than 38% compared with traditional methods, the resource utilization is greatly improved, the allocation strategy considering the system topology structure makes the resource allocation more reasonable, the resource utilization is improved, and the key connection overload is effectively reduced. The system response speed is accelerated, the forward-looking resource allocation based on the topology structure reduces the key business delay by more than 25%, and the overall performance of the system is improved. The system stability is enhanced, the connection relationship between nodes is considered to avoid key path overload, and the system stability is improved.

[0064] Please refer to Figure 2 In the computer resource allocation management method based on deep learning, multi-source heterogeneous data is acquired, a system topology graph is constructed based on the multi-source heterogeneous data, server nodes, virtual machine instances and network connections are taken as nodes of the graph, and the connection relationship between nodes is taken as edges, and a dynamic graph sequence is obtained, including the following steps:

[0065] Step 201, acquiring multi-source heterogeneous data, at least including historical resource usage data, business load data, system state data, network topology data and user behavior data;

[0066] Step 202, linear interpolation is used to complete short-term missing data in the multi-source heterogeneous data, and long-term missing abnormal data is deleted;

[0067] Step 203, after removing extreme values exceeding the normal range by the 3σ principle, the data is scaled to the [0, 1] interval to obtain preprocessed data.

[0068] The above describes the embodiment of the computer resource allocation management method, please refer to Figure 3 In the computer resource allocation management system based on deep learning, the computer resource allocation management system includes the following modules:

[0069] The dynamic graph sequence acquisition module is used to acquire multi-source heterogeneous data, construct a system topology graph based on the multi-source heterogeneous data, take server nodes, virtual machine instances and network connections as nodes of the graph, take the connection relationship between nodes as edges, and obtain a dynamic graph sequence;

[0070] The hybrid network model establishment module is used to construct a GCN-GAT hybrid graph neural network model based on a GCN graph convolution network and a GAT graph attention network, to adaptively learn the important relationship between nodes by using multi-layer graph convolution and attention mechanism, and to obtain a target GCN-GAT hybrid graph neural network model;

[0071] The resource demand prediction module is used to input the dynamic graph sequence into the target GCN-GAT hybrid graph neural network model, and output the prediction index of each type of resource demand;

[0072] The allocation strategy generation module is configured to establish a multi-objective optimization function according to the prediction index and the current system topology structure, make a resource allocation decision through an improved DQN deep reinforcement learning algorithm, and output an optimal allocation strategy.

[0073] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and equivalents thereof.

Claims

1. A computer resource allocation management method based on deep learning, characterized by, The computer resource allocation management method comprises the following steps: Obtain multi-source heterogeneous data, and construct a system topology graph based on the multi-source heterogeneous data, taking server nodes, virtual machine instances and network connections as nodes of the graph, and taking connection relationships between the nodes as edges, to obtain a dynamic graph sequence; A GCN-GAT hybrid graph neural network model is constructed based on a GCN graph convolution network and a GAT graph attention network, multi-layer graph convolution and attention mechanism are used to adaptively learn important relationships between nodes, and a target GCN-GAT hybrid graph neural network model is obtained; The dynamic graph sequence is input into the target GCN-GAT hybrid graph neural network model, and a prediction index of each type of resource demand is output; A multi-objective optimization function is established according to the prediction index and the current system topology structure, resource allocation decisions are made through an improved DQN deep reinforcement learning algorithm, and an optimal allocation strategy is output. 2.The deep learning based computer resource allocation management method of claim 1, wherein, The multi-source heterogeneous data is obtained, and a system topology graph is constructed based on the multi-source heterogeneous data, taking server nodes, virtual machine instances and network connections as nodes of the graph, and taking connection relationships between the nodes as edges, to obtain a dynamic graph sequence: Multi-source heterogeneous data is obtained, including at least historical resource usage data, business load data, system state data, network topology data and user behavior data; Short-term missing data in the multi-source heterogeneous data is completed by linear interpolation, and abnormal data missing for a long time is deleted; After removing extreme values beyond the normal range through the 3σ principle, the data is scaled to the [0, 1] interval to obtain preprocessed data. 3.The deep learning based computer resource allocation management method of claim 1, wherein, The multi-source heterogeneous data is obtained, and a system topology graph is constructed based on the multi-source heterogeneous data, taking server nodes, virtual machine instances and network connections as nodes of the graph, and taking connection relationships between the nodes as edges, to obtain a dynamic graph sequence: The server nodes include CPU core number, memory capacity, residual bandwidth and health status, the virtual machine instances include allocated CPU, current load rate and business type to which the virtual machine belongs, and the network connection nodes include bandwidth upper limit, real-time utilization rate and connection stability; The server and virtual machine edges represent that the virtual machine is deployed on the server, the server and server edges represent that two servers are connected through the network, and the virtual machine and business edges represent that the virtual machine supports the business, and the system topology graph is constructed; Based on the system topology graph, a time window of 10 minutes / piece is divided, a static graph is generated for each piece, the sequence length is fixed, and a dynamic graph sequence is obtained. 4.The deep learning based computer resource allocation management method of claim 1, wherein, The GCN-GAT hybrid graph neural network model is constructed based on the GCN graph convolution network and the GAT graph attention network, multi-layer graph convolution and attention mechanism are used to adaptively learn important relationships between nodes, and the target GCN-GAT hybrid graph neural network model is obtained: The input layer of the hybrid graph neural network model is used to receive the feature matrix and the adjacency matrix of the dynamic graph sequence; The feature fusion layer splices the sequence features in the time dimension and the topology features in the space dimension, outputs a fusion feature matrix, and the full connection layer maps the extracted features to the resource demand prediction dimension. 5.The deep learning based computer resource allocation management method of claim 1, wherein, The GCN-GAT hybrid graph neural network model is constructed based on the GCN graph convolution network and the GAT graph attention network, important relationships between nodes are adaptively learned by using multi-layer graph convolution and attention mechanism, and a target GCN-GAT hybrid graph neural network model is obtained: The first layer of the GCN graph convolution network extracts local topological features, and the second layer fuses global topological features. The attention mechanism of the GAT graph attention network calculates attention weights for neighbor nodes of each node, three attention heads are set for multi-head attention, and the robustness of feature extraction is improved. 6.The deep learning based computer resource allocation management method of claim 1, wherein, The dynamic graph sequence is input into the target GCN-GAT hybrid graph neural network model, and prediction indexes of various resource demands are output, including: 12 pieces of data are intercepted from the dynamic graph sequence as input of the target GCN-GAT hybrid graph neural network model, and prediction indexes of various resource demands in future 30 minutes, 1 hour and 2 hours are output. 7.The deep learning based computer resource allocation management method of claim 1, wherein, The multi-objective optimization function is established according to the prediction indexes and the current system topological structure, resource allocation decisions are made by using the improved DQN deep reinforcement learning algorithm, and the optimal allocation strategy is output, including: The double-DQN structure is set, the evaluation network and the target network of the DQN deep reinforcement learning algorithm are set separately, and the improved DQN deep reinforcement learning algorithm is obtained; The current system topological structure, the prediction indexes and the allocated resource data are obtained, the state vector is constructed, the action space is defined, including resource adjustment actions and topological adjustment actions, and the optimal allocation strategy is output.

8. A computer resource allocation management system based on deep learning, characterized by, The computer resource allocation management system comprises the following modules: A dynamic graph sequence acquisition module is configured to acquire multi-source heterogeneous data, construct a system topological graph based on the multi-source heterogeneous data, take server nodes, virtual machine instances and network connections as nodes of the graph, take connection relationships between the nodes as edges, and obtain a dynamic graph sequence; A hybrid network model establishment module is configured to construct a GCN-GAT hybrid graph neural network model based on a GCN graph convolution network and a GAT graph attention network, adaptively learn important relationships between nodes by using multi-layer graph convolution and attention mechanism, and obtain a target GCN-GAT hybrid graph neural network model; A resource demand prediction module is configured to input the dynamic graph sequence into the target GCN-GAT hybrid graph neural network model, and output prediction indexes of various resource demands; An allocation strategy generation module is configured to establish a multi-objective optimization function according to the prediction indexes and the current system topological structure, make resource allocation decisions by using the improved DQN deep reinforcement learning algorithm, and output an optimal allocation strategy.

9. The deep learning-based computer resource allocation management system of claim 8, wherein, The dynamic graph sequence acquisition module comprises the following sub-modules: An acquisition sub-module is configured to acquire multi-source heterogeneous data, including at least historical resource usage data, business load data, system state data, network topological data and user behavior data; A deletion sub-module is configured to adopt linear interpolation to complete short-term missing data in the multi-source heterogeneous data, and delete long-term missing abnormal data; A scaling sub-module is configured to remove extreme values beyond a normal range by using a 3σ principle, scale the data to a [0, 1] interval, and obtain preprocessed data.

10. The deep learning-based computer resource allocation management system of claim 8, wherein, The distribution strategy generation module comprises the following sub-modules: The setting sub-module is configured to set a double DQN structure, separate evaluation network and target network of a DQN deep reinforcement learning algorithm, and obtain an improved DQN deep reinforcement learning algorithm; The output sub-module is configured to obtain current system topology structure, prediction index and allocated resource data, construct a state vector, define an action space including resource adjustment actions and topology adjustment actions, and output an optimal distribution strategy.

Images