Data center resource intelligent scheduling method and system based on digital twinborn and space-time diagram neural network
By constructing a digital twin graph model and a spatiotemporal graph neural network, the scheduling lag and prediction error problems in data center resource scheduling are solved, achieving efficient and accurate resource scheduling and reducing operating costs and failure rates.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-03
AI Technical Summary
Existing data center resource scheduling methods suffer from severe scheduling lag, neglect of spatial characteristics, difficulty in multi-dimensional resource coordination, insufficient physical constraints, and lack of security verification, resulting in increased service response time, large prediction errors, and high scheduling failure rates in high-concurrency scenarios.
A digital twin graph model is constructed and combined with a spatiotemporal graph neural network. By collecting multi-dimensional features, load prediction is performed, multiple optimization objectives are set, a scheduling optimization model is constructed, and simulation verification and feedback optimization are carried out to realize intelligent scheduling of resource scheduling strategies.
It significantly reduces scheduling lag, improves resource utilization, reduces prediction errors, reduces scheduling failure rate, reduces energy consumption, and supports adaptive optimization and real-time scheduling.
Smart Images
Figure CN121785781A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cloud computing technology, and in particular to a method and system for intelligent scheduling of data center resources based on digital twins and spatiotemporal graph neural networks. Background Technology
[0002] Current data center resource scheduling mainly relies on two types of methods: threshold-based passive scheduling and machine learning-based active scheduling. Threshold-based methods trigger scheduling by setting thresholds such as Central Processing Unit (CPU) utilization and memory occupancy, but they suffer from significant lag. Machine learning-based methods generally use Autoregressive Integrated Moving Average (ARIMA) models or Long Short-Term Memory (LSTM) networks for load prediction, but they only consider time-dimensional features and ignore the spatial relationships between data center nodes.
[0003] Existing technologies suffer from several shortcomings, including severe scheduling lag. Passive scheduling strategies have an average latency of 180 to 300 seconds, increasing service response time by 200% to 500% in high-concurrency scenarios. Spatial feature modeling is lacking; existing prediction methods only consider the temporal dependence of load and fail to capture the spatial correlation between nodes. In actual operation, the load correlation coefficient between adjacent servers can reach 0.6 to 0.8. Ignoring spatial features leads to a mean absolute percentage error (MAPE) exceeding 25%. Multi-dimensional resource coordination is difficult; traditional methods optimize for a single resource dimension, resulting in an overall resource utilization improvement of less than 10%. Physical constraints are not adequately considered; scheduling algorithms do not fully model constraints such as server physical location, rack power consumption limits, and network topology latency. Approximately 15% to 20% of scheduling schemes cannot be deployed in practice due to violations of physical constraints. There is a lack of security verification mechanisms; strategy optimization requires trial and error in a production environment, resulting in a scheduling failure rate of 5% to 8%, with each failure potentially causing service interruptions for hundreds of virtual machines. Summary of the Invention
[0004] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides an intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks to solve the above problems.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a data center resource intelligent scheduling method based on digital twins and spatiotemporal graph neural networks, comprising: constructing a digital twin graph model, collecting real-time data of various resources in the data center, and converting the real-time status data of the resources into a graph structure; Based on the digital twin graph model, a spatiotemporal load prediction model is constructed to extract multidimensional features of various resources in the data center in order to identify dynamic resource patterns and predict resource load. Based on the multidimensional characteristics of the resources, a scheduling optimization model is constructed by setting multiple optimization objectives, and a resource scheduling strategy is generated. The scheduling scheme of the resource scheduling strategy is simulated and verified, the execution result of the strategy is evaluated, and the scheduling model is optimized based on the evaluation result.
[0006] As a preferred embodiment of the intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks described in this invention, the construction of the digital twin graph model includes: Real-time data for various resources includes resource performance data, physical environment data, and network status data; In the virtual space, a digital twin corresponding to the physical entity is created, a set of nodes is generated, and the collected resource data is divided into node static attributes, node dynamic attributes, and edge attributes, and associated with the digital twin. Based on actual physical connections and logical dependencies, the association between the digital twins is established, a digital twin graph model is constructed, the state sequence of the digital twin graph at each time moment is obtained, and the state is updated synchronously.
[0007] As a preferred embodiment of the intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks described in this invention, the method includes: constructing a spatiotemporal load prediction model and extracting multidimensional features of various data center resources, including: Based on the state sequence of the digital twin graph, a spatiotemporal load prediction model is constructed through a spatiotemporal graph neural network, and the model is trained and optimized to predict the node load of various resources at multiple future time steps. The spatiotemporal load prediction model extracts temporal features through a time encoder, extracts spatial features through a graph attention network, obtains spatiotemporal features through a gating fusion mechanism, and decodes the extracted features to obtain the prediction result.
[0008] As a preferred embodiment of the intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks described in this invention, the scheduling optimization model is constructed by setting multiple optimization objectives, including: Define decision variables, construct a scheduling optimization model with response time, resource utilization, energy consumption and migration cost as objectives, and introduce resource capacity, service quality, resource power consumption and anti-affinity constraints; Using the predicted node load and real-time resource status data as input to the scheduling optimization model, an improved non-dominated sorting genetic algorithm is used to solve the scheduling optimization model to obtain a resource scheduling scheme.
[0009] As a preferred embodiment of the intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks described in this invention, the method includes: simulating and verifying the scheduling scheme of the resource scheduling strategy, including: A discrete event simulation engine is constructed, event types are defined, scheduling schemes are transformed into simulation event sequences, resource working processes are simulated, and scheduling simulation results are obtained. Based on the scheduling simulation results, performance indicators are calculated, and a performance evaluation report of the scheduling scheme in the simulation environment is generated. The scheduling scheme is decomposed into a set of operations, an operation dependency graph is constructed, the scheduling scheme is executed based on the operation dependency graph, the prediction error is monitored in real time, and if it exceeds a set threshold, online learning is triggered to adaptively adjust the parameters of the spatiotemporal load prediction model.
[0010] As a preferred embodiment of the intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks described in this invention, the digital twin graph model is represented as follows: ; in, Represents a set of nodes. , Represents a set of server nodes. , Represents a set of storage nodes. }, Represents a set of network device nodes; This represents an edge set, which includes physical connections and logical dependencies. Represents the set of node attributes; Represents the set of edge attributes; Represents a set of timestamps.
[0011] As a preferred embodiment of the intelligent data center resource scheduling method based on digital twin and spatiotemporal graph neural network described in this invention, the spatiotemporal load prediction model includes a time encoder, a spatial encoder, a spatiotemporal fusion module, and a decoder. The time encoder updates the hidden state of each node step by step through the gated loop unit update equation to extract time features, namely the node load time series. The spatial encoder calculates the attention weights between the current node and its neighbors through a graph attention network, aggregates the neighbor features, and obtains a spatial feature vector. The spatiotemporal fusion module obtains spatiotemporal features by dynamically balancing temporal and spatial features through a fusion gate; The decoder maps the spatiotemporal features to load predictions for the next H steps using a three-layer fully connected network.
[0012] Secondly, the present invention provides a data center resource intelligent scheduling system based on digital twins and spatiotemporal graph neural networks, comprising: The first construction module is used to build a digital twin graph model, collect real-time data of various resources in the data center, and transform the real-time status data of the resources into a graph structure. The second construction module is used to construct a spatiotemporal load prediction model based on the digital twin graph model, extract multi-dimensional features of various resources in the data center, identify dynamic resource patterns, and predict resource load. The third construction module is used to construct a scheduling optimization model and generate a resource scheduling strategy based on the multi-dimensional characteristics of the resources and by setting multiple optimization objectives. The simulation optimization module is used to simulate and verify the scheduling scheme of the resource scheduling strategy, evaluate the strategy execution results, and optimize the scheduling model based on the evaluation results.
[0013] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the intelligent scheduling method for data center resources based on digital twins and spatiotemporal graph neural networks.
[0014] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks.
[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention significantly reduces scheduling lag by using ST-GNN to predict load at future times, significantly reducing scheduling response time; it achieves global optimization of multi-dimensional resources, with a digital twin representation based on a graph model fully depicting the multi-dimensional resource relationships, improving CPU utilization, memory utilization, and network bandwidth utilization; it balances prediction real-time performance and accuracy, with single inference time meeting real-time scheduling requirements; it reduces the mean absolute percentage error (MAPE) of load prediction and handles sudden scenarios, significantly outperforming traditional models; it fully considers physical constraints, with the digital twin model fully mapping physical topology, power consumption limits, and heat dissipation capabilities, improving the feasibility of scheduling schemes; simultaneously, it provides efficient simulation verification capabilities, with simulation speeds 20 to 50 times faster than actual execution speeds, reducing simulation result errors and providing a reliable basis for strategy optimization and risk assessment, while supporting continuous adaptive optimization. The online learning mechanism reduces predicted MAPE, improves the optimization target value, reduces energy consumption, and the system automatically adapts to changes in business models; it reduces operating costs, lowers electricity consumption, saves electricity bills, and reduces scheduling failure rates. Attached Figure Description
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the overall process of a data center resource intelligent scheduling method based on digital twin and spatiotemporal graph neural network according to an embodiment of the present invention. Detailed Implementation
[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0019] Reference Figure 1 As one embodiment of the present invention, a data center resource intelligent scheduling method based on digital twins and spatiotemporal graph neural networks is provided, comprising: S100 constructs a digital twin graph model, collects real-time data of various resources in the data center, and transforms the real-time status data of the resources into a graph structure; S200, based on a digital twin graph model, constructs a spatiotemporal load prediction model, extracts multidimensional features of various resources in the data center to identify dynamic resource patterns, and performs load prediction on resources. S300, based on the multi-dimensional characteristics of resources, constructs a scheduling optimization model by setting multiple optimization objectives and generates resource scheduling strategies; S400 performs simulation verification of the resource scheduling strategy, evaluates the strategy execution results, and optimizes the scheduling model based on the evaluation results.
[0020] In a preferred embodiment, constructing a digital twin graph model includes: Real-time data for various resources includes resource performance data, physical environment data, and network status data, and the collected data undergoes standardized processing. In the virtual space, a digital twin corresponding to the physical entity is created, a set of nodes is generated, and the collected resource data is divided into node static attributes, node dynamic attributes, and edge attributes, and then associated with the digital twin. Based on actual physical connections and logical dependencies, the relationships between digital twins are established, a digital twin graph model is constructed, the state sequence of the digital twin graph at each time step is obtained, and the state is updated synchronously.
[0021] In a preferred embodiment, the digital twin graph is defined as a quintuple, represented as:
[0022] in, Represents a set of nodes. , Represents a set of server nodes. , Represents a set of storage nodes. }, Represents a set of network device nodes; This represents an edge set, which includes physical connections and logical dependencies. Represents the set of node attributes; Represents the set of edge attributes; Represents a set of timestamps.
[0023] In this embodiment, the nodes of the node set in the digital twin graph The attribute vector includes static attributes and dynamic attributes, represented as: ; in, For the current moment, As a static property, it is defined as follows: ; Among them, resource capacity These represent the number of CPU cores, memory capacity (Gigabyte, GB), storage capacity (Terabyte, TB), and network bandwidth (Gigabit per second, Gbps), respectively. For physical location; This is the upper limit of power consumption; For heat dissipation capacity.
[0024] This represents a dynamic attribute, defined as: ; in, The resource utilization vector is represented as follows: ; satisfy ; In running state; Temperature (Celsius, °C); Real-time power consumption (kilowatt, kW); This is a list of virtual machines to be hosted.
[0025] Edges of edge concentration The attribute vector is represented as: ; in For link bandwidth utilization, For real-time traffic, For link capacity; To delay the spread; This refers to the packet loss rate.
[0026] Through a state synchronization mechanism and an event-driven incremental update strategy, when newly arrived monitoring data at time t is acquired... When calculating attribute change , is represented as:
[0027] in, For node objects in a digital twin graph, The current value of the node count attribute. , The threshold value can be dynamically set according to the actual application scenario. An update will then be performed, represented as follows: ; And update the timestamp, represented as: ; Furthermore, a version control mechanism is introduced, where the version number increments after each status update, represented as: ; in, This represents the version number at time t.
[0028] In a preferred embodiment, a spatiotemporal load prediction model is constructed to extract multidimensional features of various resources in the data center, including: Based on the state sequence of digital twin graphs, a spatiotemporal load prediction model is constructed through a spatiotemporal graph neural network, and the model is trained and optimized to predict the node load of various resources at multiple future time steps. The spatiotemporal load prediction model extracts temporal features through a time encoder, extracts spatial features through a graph attention network, obtains spatiotemporal features through a gating fusion mechanism, and decodes the extracted features to obtain the prediction results.
[0029] In a preferred embodiment, the spatiotemporal load prediction model includes a time encoder, a spatial encoder, a spatiotemporal fusion module, and a decoder; The time encoder updates the hidden state of each node step by step through the update equation of the gated cyclic unit, and extracts the time features, namely the node load time series. The spatial encoder uses a graph attention network to calculate the attention weights between the current node and its neighbors, aggregates the neighbor features, and obtains a spatial feature vector. The spatiotemporal fusion module dynamically weighs temporal and spatial features through a fusion gate to obtain spatiotemporal features; The decoder maps spatiotemporal features to load predictions for the next H steps using a three-layer fully connected network.
[0030] In this embodiment, the state sequence of the data center over the past T time steps is defined as follows: ,in, This represents the state sequence at time t, predicting the load state at the next H time steps. ,in, This is the load vector for all nodes. Let t be the load value of node N at time t.
[0031] The spatiotemporal load prediction model adopts a spatiotemporal graph neural network architecture, namely an encoder-decoder architecture, which includes a time encoder, a spatial encoder, a spatiotemporal fusion module, and a decoder.
[0032] The time encoder uses a gated recurrent unit (GRU) to extract time-dependent features for each node. Its load time series is .
[0033] The GRU update equation is expressed as: ; ; ; ; in, It is the Sigmoid activation function. This is element-wise multiplication (Hadamard product). This is the weight matrix. The bias vector; the output node of the time encoder. Time feature vector , , Let represent the final hidden state and the candidate hidden state at time t, respectively. This indicates that the door is being reset. This indicates an update to the door.
[0034] The spatial encoder uses a Graph Attention Network (GAT) to capture spatial dependencies for nodes. Its set of neighboring nodes is .
[0035] The attention coefficient is calculated as follows: ; Spatial feature aggregation is represented as: ; in, Represents the spatial feature vector of node i. Attention coefficient It is a linear transformation matrix. For attention vectors, For splicing operations, For activation function, , For node feature vectors, It is a non-linear activation function.
[0036] The spatiotemporal fusion module adopts a gated fusion mechanism, represented as: ; ; in, For the fusion gate, The fused spatiotemporal feature vector, This is the bias vector of the fusion gate.
[0037] The decoder uses a three-layer fully connected network, represented as: ; ; ; in, , for nodes Load forecasts for the next H time steps.
[0038] Furthermore, the model is trained, and the loss function is a combination of mean squared error and prediction uncertainty, expressed as: ; in, This represents the actual load value. To predict variance, quantify uncertainty. This is the regularization coefficient.
[0039] Using the Adam optimizer, the learning rate The batch size B=64, and the model adopts the pre-training-fine-tuning paradigm. After pre-training on a public dataset, it is fine-tuned using historical data (7 to 14 days) from the target data center.
[0040] In a preferred embodiment, constructing a scheduling optimization model by setting multiple optimization objectives includes: Define decision variables, construct a scheduling optimization model with response time, resource utilization, energy consumption and migration cost as objectives, and introduce resource capacity, service quality, resource power consumption and anti-affinity constraints; Using predicted node load and real-time resource status data as inputs to the scheduling optimization model, an improved non-dominated sorting genetic algorithm is employed to solve the scheduling optimization model and obtain a resource scheduling scheme.
[0041] In this embodiment, the decision variable is , indicating virtual machine Should it be assigned to a server? ,as well as , indicating virtual machine Should we migrate from the current server to the new server? .
[0042] Define multiple optimization objectives, specifically including: Minimize service response time , is represented as: ; in, , , They represent virtual machines. On the server Task processing time, network transmission latency, and storage usage. Indicates the number of servers. Indicates the number of virtual machines.
[0043] Maximize resource utilization , is represented as: ; in, , , , These represent the server's CPU utilization, memory utilization, storage utilization, and network utilization, respectively.
[0044] Minimize energy consumption , is represented as: ; in, Indicates server s i Idle power consumption, Indicates server Maximum power consumption.
[0045] Minimize virtual machine migration cost, expressed as: ; in, This represents the fixed overhead of virtual machine migration, including state synchronization, connection reconstruction, etc. Represents virtual machine Memory size (GB) This indicates network bandwidth (Gbps).
[0046] In summary, multi-objective functions Represented as: ; in, Let be the weighting coefficient, satisfying .
[0047] Define constraints, specifically including: Resource capacity constraints are expressed as: ; ; ; in, Represents virtual machine Required number of CPU cores Indicates server CPU resources Represents virtual machine Required memory, representing Indicates server memory capacity, Represents virtual machine Required storage capacity Indicates server Storage capacity.
[0048] The unique allocation constraint for a virtual machine is represented as follows: ; Service quality constraints are represented as: ; in, Indicates service quality, Represents virtual machine The maximum response time specified in the service level agreement.
[0049] Power consumption constraints (rack level) are expressed as: ; in, Indicates power consumption. Indicates the first k A rack contains a collection of servers. This represents the upper limit of power consumption for the k-th rack, which is limited by power supply and heat dissipation capabilities.
[0050] Anti-affinity constraint (high availability deployment) is represented as: ; in, For a collection of virtual machine pairs that require anti-affinity deployment.
[0051] Furthermore, an improved Non-Dominated Sorting Genetic Algorithm II (NSGA-II) is used to solve the scheduling optimization model. Chromosomes are encoded with integers and have a length of [missing information]. M (Number of virtual machines), the value of the j-th gene is k express Assigned to The crossover operation uses two-point crossover, with crossover probability... The mutation operation uses random reset mutation, with a mutation probability. .
[0052] Perform non-dominated sorting, individuals Dominant Individual If and only if: ; in, express In the k The function value on the optimization objective.
[0053] Crowding is calculated and expressed as: ; in, and In the first k On the goal Adjacent individuals, , These represent the population at the th... k The maximum and minimum values on each target.
[0054] In a preferred embodiment, the scheduling scheme of the resource scheduling strategy is simulated and verified, including: A discrete event simulation engine is constructed, event types are defined, scheduling schemes are transformed into simulation event sequences, resource working processes are simulated, and scheduling simulation results are obtained. Based on the scheduling simulation results, performance indicators are calculated, and a performance evaluation report of the scheduling scheme in the simulation environment is generated. The scheduling scheme is decomposed into a set of operations, an operation dependency graph is constructed, the scheduling scheme is executed based on the operation dependency graph, the prediction error is monitored in real time, and if it exceeds a set threshold, online learning is triggered to adaptively adjust the parameters of the spatiotemporal load prediction model.
[0055] In this embodiment, the simulation engine employs discrete event simulation technology, and the event types are... definition: ; in, This indicates that the virtual machine has started working. Indicates a virtual machine migration event. Indicates the event that the task has arrived. This indicates a task completion event.
[0056] Simulation engine maintenance event list (sorted by timestamp), simulation clock from the current moment Advance to Then, performance metrics are calculated, including: Average response time , is represented as: ; in, This represents the set of all tasks completed during the simulation. , These represent the task completion time and the task arrival time, respectively.
[0057] Average resource utilization rate , is represented as: ; in, Indicates the total simulation duration. Indicates server exist t The overall resource utilization rate at any time This represents the time differential element.
[0058] Average power consumption , is represented as: ; Service Level Agreement (SLA) violation rate , is represented as: ; in, This represents the set of tasks whose response time exceeds the SLA requirement.
[0059] Furthermore, parallel simulation technology and adaptive time step are used to accelerate the simulation, making the simulation speed 20 to 50 times faster than that of the actual system.
[0060] Furthermore, a coordination mechanism is implemented to construct an Operation Dependency Graph (ODG), defined as a Directed Acyclic Graph (DAG), represented as follows: ; in, For operation sets, For a set of dependencies, topological sorting is used to determine the execution order, while operations without dependencies are executed in parallel. Then, online adaptive optimization is performed; when the prediction error exceeds a threshold, online learning is triggered to adjust the prediction error. The calculation formula is: ; like ,in, The prediction threshold is set at 20%, and the model is fine-tuned using the latest observation data, with the learning rate adaptively adjusted. Represented as: ; Freeze the encoder parameters and fine-tune only the decoder (5 epochs) to enable the model to continuously optimize.
[0061] It should be noted that this invention significantly reduces scheduling lag by using ST-GNN to predict load at future time points, thus significantly reducing scheduling response time. It achieves global optimization of multi-dimensional resources, with a graph-based digital twin representation fully depicting multi-dimensional resource relationships, improving CPU utilization, memory utilization, and network bandwidth utilization. It balances prediction real-time performance and accuracy, with single inference time meeting real-time scheduling requirements. It reduces the mean absolute percentage error (MAPE) of load prediction and handles sudden scenarios, significantly outperforming traditional models. It fully considers physical constraints; the digital twin model can fully map physical topology, power consumption limits, and heat dissipation capabilities, improving the feasibility of scheduling schemes. Simultaneously, it provides efficient simulation verification capabilities, with simulation speeds 20 to 50 times faster than actual execution speeds, reducing simulation result errors and providing a reliable basis for strategy optimization and risk assessment. It also supports continuous adaptive optimization, with an online learning mechanism reducing prediction MAPE, improving the optimization target value, reducing energy consumption, and allowing the system to automatically adapt to changes in business models. Finally, it reduces operating costs, lowers electricity consumption, saves electricity bills, and reduces scheduling failure rates.
[0062] The above is an illustrative scheme of a data center resource intelligent scheduling method based on digital twins and spatiotemporal graph neural networks according to this embodiment. It should be noted that the technical solution of this data center resource intelligent scheduling system based on digital twins and spatiotemporal graph neural networks belongs to the same concept as the technical solution of the data center resource intelligent scheduling method based on digital twins and spatiotemporal graph neural networks described above. Details not described in detail in the technical solution of the data center resource intelligent scheduling system based on digital twins and spatiotemporal graph neural networks in this embodiment can be found in the description of the technical solution of the data center resource intelligent scheduling method based on digital twins and spatiotemporal graph neural networks described above.
[0063] This embodiment provides a data center resource intelligent scheduling system based on digital twins and spatiotemporal graph neural networks, including: The first construction module is used to build a digital twin graph model, collect real-time data of various resources in the data center, and transform the real-time status data of the resources into a graph structure. The second construction module is used to construct a spatiotemporal load prediction model based on the digital twin graph model, extract multi-dimensional features of various resources in the data center, identify dynamic resource patterns, and predict resource load. The third construction module is used to construct a scheduling optimization model and generate a resource scheduling strategy based on the multi-dimensional characteristics of the resources and by setting multiple optimization objectives. The simulation optimization module is used to simulate and verify the scheduling scheme of the resource scheduling strategy, evaluate the strategy execution results, and optimize the scheduling model based on the evaluation results.
[0064] This embodiment also provides a computer device suitable for intelligent scheduling of data center resources based on digital twins and spatiotemporal graph neural networks, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement a data center resource intelligent scheduling method based on digital twins and spatiotemporal graph neural networks, as proposed in the above embodiments.
[0065] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a data center resource intelligent scheduling method based on digital twins and spatiotemporal graph neural networks as proposed in the above embodiment.
[0066] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for intelligent scheduling of data center resources based on digital twins and spatiotemporal graph neural networks proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.
[0067] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.
[0068] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A data center resource intelligent scheduling method based on digital twins and spatiotemporal graph neural networks, characterized in that, include: Construct a digital twin graph model, collect real-time data of various resources in the data center, and transform the real-time status data of the resources into a graph structure; Based on the digital twin graph model, a spatiotemporal load prediction model is constructed to extract multidimensional features of various resources in the data center in order to identify dynamic resource patterns and predict resource load. Based on the multidimensional characteristics of the resources, a scheduling optimization model is constructed by setting multiple optimization objectives, and a resource scheduling strategy is generated. The scheduling scheme of the resource scheduling strategy is simulated and verified, the execution result of the strategy is evaluated, and the scheduling model is optimized based on the evaluation result.
2. The intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks as described in claim 1, characterized in that, Building a digital twin graph model includes: Real-time data for various resources includes resource performance data, physical environment data, and network status data; In the virtual space, a digital twin corresponding to the physical entity is created, a set of nodes is generated, and the collected resource data is divided into node static attributes, node dynamic attributes, and edge attributes, and associated with the digital twin. Based on actual physical connections and logical dependencies, the association between the digital twins is established, a digital twin graph model is constructed, the state sequence of the digital twin graph at each time moment is obtained, and the state is updated synchronously.
3. The intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks as described in claim 2, characterized in that, Construct a spatiotemporal load prediction model and extract multidimensional features of various resources in the data center, including: Based on the state sequence of the digital twin graph, a spatiotemporal load prediction model is constructed through a spatiotemporal graph neural network, and the model is trained and optimized to predict the node load of various resources at multiple future time steps. The spatiotemporal load prediction model extracts temporal features through a time encoder, extracts spatial features through a graph attention network, obtains spatiotemporal features through a gating fusion mechanism, and decodes the extracted features to obtain the prediction result.
4. The intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks as described in claim 3, characterized in that, By setting multiple optimization objectives, a scheduling optimization model is constructed, including: Define decision variables, construct a scheduling optimization model with response time, resource utilization, energy consumption and migration cost as objectives, and introduce resource capacity, service quality, resource power consumption and anti-affinity constraints; Using the predicted node load and real-time resource status data as input to the scheduling optimization model, an improved non-dominated sorting genetic algorithm is used to solve the scheduling optimization model to obtain a resource scheduling scheme.
5. The intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks as described in claim 1, characterized in that, The scheduling scheme of the resource scheduling strategy is simulated and verified, including: A discrete event simulation engine is constructed, event types are defined, scheduling schemes are transformed into simulation event sequences, resource working processes are simulated, and scheduling simulation results are obtained. Based on the scheduling simulation results, performance indicators are calculated, and a performance evaluation report of the scheduling scheme in the simulation environment is generated. The scheduling scheme is decomposed into a set of operations, an operation dependency graph is constructed, the scheduling scheme is executed based on the operation dependency graph, the prediction error is monitored in real time, and if it exceeds a set threshold, online learning is triggered to adaptively adjust the parameters of the spatiotemporal load prediction model.
6. The intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks as described in claim 2, characterized in that, The digital twin graph model is represented as follows: ； in, Represents a set of nodes. , Represents a set of server nodes. , Represents a set of storage nodes. , Represents a set of network device nodes; This represents an edge set, which includes physical connections and logical dependencies. Represents the set of node attributes; Represents the set of edge attributes; Represents a set of timestamps.
7. The intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks as described in claim 3, characterized in that, The spatiotemporal load prediction model includes a time encoder, a spatial encoder, a spatiotemporal fusion module, and a decoder; The time encoder updates the hidden state of each node step by step through the gated loop unit update equation to extract time features, namely the node load time series. The spatial encoder calculates the attention weights between the current node and its neighbors through a graph attention network, aggregates the neighbor features, and obtains a spatial feature vector. The spatiotemporal fusion module obtains spatiotemporal features by dynamically balancing temporal and spatial features through a fusion gate; The decoder maps the spatiotemporal features to load predictions for the next H steps using a three-layer fully connected network.
8. A data center resource intelligent scheduling system based on digital twins and spatiotemporal graph neural networks, employing the data center resource intelligent scheduling method based on digital twins and spatiotemporal graph neural networks as described in any one of claims 1 to 7, characterized in that, include: The first construction module is used to build a digital twin graph model, collect real-time data of various resources in the data center, and transform the real-time status data of the resources into a graph structure. The second construction module is used to construct a spatiotemporal load prediction model based on the digital twin graph model, extract multidimensional features of various resources in the data center, identify dynamic resource patterns, and predict resource load. The third construction module is used to construct a scheduling optimization model and generate a resource scheduling strategy based on the multi-dimensional characteristics of the resources and by setting multiple optimization objectives. The simulation optimization module is used to simulate and verify the scheduling scheme of the resource scheduling strategy, evaluate the strategy execution results, and optimize the scheduling model based on the evaluation results.
9. A computer device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the intelligent data center resource scheduling method based on digital twin and spatiotemporal graph neural network as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed by a processor, implement the steps of the intelligent data center resource scheduling method based on digital twins and spatiotemporal graph neural networks as described in any one of claims 1 to 7.