Intelligent operation and maintenance method and system for data center based on digital twinning

Abstract

The application discloses a data center intelligent operation and maintenance method and system based on digital twinning, relates to the technical field of data processing, and is used for solving the problem of secondary disaster caused by lagging behind in micro-resource competition identification and blind scheduling in a data center across layers. First, the bottom micro-architecture and micro-service tracking data are acquired and mapped to a cross-layer digital twinning topology graph to generate software and hardware running characteristics. Then, the graph attention network is used to aggregate the delay gradient and the cache miss rate to extract micro-resource competition characteristics and accurately locate abnormal nodes. Further, in a virtual sandbox, combined with high-concurrency traffic, fault propagation evolution is carried out along the call edge to output the affected nodes and the cascading failure path. Finally, multi-objective cooperative game analysis is carried out by comprehensively considering the business protocol attribute and the computing power bandwidth cost to generate and execute a dynamic control script to realize hot migration and safety isolation, and an operation and maintenance closed-loop system from digital mapping to physical self-healing is constructed.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a data center intelligent operation and maintenance method and system based on digital twins. Background Technology

[0002] With the rapid development of the digital economy and the widespread adoption of cloud computing technology, data centers, as the core infrastructure for handling massive amounts of data computing and storage, are experiencing exponential growth in scale and complexity. Especially in vertical sectors such as financial information technology outsourcing, high-concurrency payment settlement, and massive credit data processing, modern data centers not only need to handle the sudden surge in massive distributed business requests, but also ensure that the underlying physical hardware resources can be efficiently and stably scheduled and allocated to meet the stringent service level agreements (SLAs) required for financial institutions. Under this high-load operating environment, the coupling between hardware and software within data centers is increasingly deep, and even minor resource competition can easily trigger large-scale business congestion or even transaction interruptions. This places extremely high demands on the intelligent operation and maintenance and refined management of data centers.

[0003] Existing data center operation and maintenance methods exhibit significant lag and limitations when handling high-concurrency scenarios such as financial transactions or batch data transfers. Traditional operation and maintenance approaches typically operate independently of physical hardware performance monitoring and upper-layer software business link tracing. This results in situations like cross-layer resource contention, such as a batch data encryption security process instantly preempting the underlying cache of a real-time transaction link, where the monitoring system can only issue alarm signals after overall resource utilization spikes or widespread business interruptions occur, failing to locate micro-level anomalies during the fault latency period. Furthermore, when performing server resource scheduling and business migration, existing methods mostly execute static, fixed rule allocations directly in the physical environment. This operation is highly susceptible to secondary system disasters under complex high-concurrency traffic surges, leading to uncontrollable risks during business recovery and seriously threatening the security and continuity of core data processing operations. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a data center intelligent operation and maintenance method and system based on digital twins, which solves the problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a data center intelligent operation and maintenance method based on digital twins, comprising the following steps: S1, obtaining the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice link tracing sequence, mapping the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice link tracing sequence to a pre-constructed cross-layer digital twin topology graph, generating hardware and software operating status slice features; S2, performing graph attention network aggregation on the data processing microservice response latency gradient and hardware last-level cache miss rate in the hardware and software operating status slice features, extracting micro-resource contention feature vectors, and calculating cross-container resource preemption within the physical host based on the micro-resource contention feature vectors. S3. For cross-layer digital twin topology nodes with abnormal node location markings, combined with the real-time high-concurrency data traffic load characteristics, the sandbox fault propagation evolution is carried out along the call dependency edge of the distributed data processing microservice link tracing sequence, and the set of affected microservice nodes and cascaded failure network topology paths are output; S4. The data processing service level protocol attributes of the affected microservice node set are extracted, and combined with the micro-computing power margin of the backup physical machine cluster and the cross-node migration bandwidth cost of the cascaded failure network topology path, a dynamic control script is generated through multi-objective collaborative game analysis. The dynamic control script is then distributed to the physical data center to execute the cross-region hot migration of the target microservice and the security isolation of abnormal containers.

[0006] Furthermore, the specific process for obtaining the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice tracing sequence is as follows: A kernel-level probe pre-installed on the physical server operating system collects the hardware microinstruction retirement rate, bus bandwidth usage, and cache hit allocation status at each level from the central processing unit's performance monitoring unit. The hardware microinstruction retirement rate, bus bandwidth usage, and cache hit allocation status at each level are then concatenated according to timestamps to construct the underlying server microarchitecture telemetry timing matrix. A non-intrusive tracing agent is injected into the distributed data processing microservice mesh to capture the request enqueue timestamps and execution completion timestamps of cross-container remote procedure calls. These timestamps are then topologically assembled according to the microservice call dependency hierarchy to generate the distributed data processing microservice tracing sequence.

[0007] Furthermore, the specific process of mapping the underlying server microarchitecture telemetry time series matrix and the distributed data processing microservice tracing sequence to a pre-constructed cross-layer digital twin topology graph to generate hardware and software runtime status slice features is as follows: Parse the infrastructure i.e., code configuration file of the physical data center to construct a multi-level entity graph containing physical host nodes, container nodes, and microservice application nodes; extract the hardware host identifier from the underlying server microarchitecture telemetry time series matrix and the container runtime environment identifier from the distributed data processing microservice tracing sequence; establish cross-layer resource binding edges in the multi-level entity graph based on the hardware host identifier and the container runtime environment identifier to generate a cross-layer digital twin topology graph; perform a sliding window truncation of the cross-layer digital twin topology graph along the time axis at a set step size; and perform tensor concatenation of the underlying server microarchitecture telemetry time series matrix features and the distributed data processing microservice tracing sequence features along the node feature dimension within the same time segment to generate hardware and software runtime status slice features.

[0008] Furthermore, the specific process of using graph attention network aggregation to extract micro-resource contention feature vectors from the data processing microservice response latency gradient and hardware last-level cache miss rate in the hardware and software operation status slice features is as follows: First-order difference operations are performed along the time dimension on the distributed data processing microservice link tracing sequence features in the hardware and software operation status slice features to extract the dynamic response time change rate of each microservice application node and construct the data processing microservice response latency gradient; the cache line eviction event sequence and multi-core parallel access conflict frequency of physical host nodes are separated from the underlying server microarchitecture telemetry time series matrix features in the hardware and software operation status slice features, and combined with the latency penalty coefficient of the micro-instruction retirement cycle, the hardware last-level cache miss rate is constructed; the data processing microservice response latency gradient and hardware last-level cache miss rate are used as the initial node features input to the graph attention network layer, and information aggregation weights are assigned to adjacent nodes in the cross-layer digital twin topology graph through a masked self-attention mechanism. The features of adjacent nodes are then weighted and summed to activate them according to the information aggregation weights, outputting the micro-resource contention feature vector.

[0009] Furthermore, the specific process of calculating the cross-container resource preemption interference index within the physical host and marking abnormal nodes based on the micro-resource competition feature vector is as follows: A multilayer perceptron is used to perform a nonlinear mapping on the micro-resource competition feature vector, outputting a cross-container resource preemption interference index that characterizes the probability of hardware bus queuing conflicts between multiple data processing microservices within the same physical host; the exponential moving average and standard deviation of the cross-container resource preemption interference index are calculated within a time sliding window to construct a dynamic adaptive interference threshold; the difference between the current cross-container resource preemption interference index and the dynamic adaptive interference threshold is compared, and microservice container entities whose differences exceed the tolerance range multiple times consecutively are extracted, and abnormal node location marking is performed in the cross-layer digital twin topology graph.

[0010] Furthermore, for cross-layer digital twin topology nodes with abnormal node location markers, and combined with real-time high-concurrency data traffic load characteristics, the specific process of sandbox fault propagation evolution along the call dependency edges of the distributed data processing microservice link tracing sequence is as follows: The total number of concurrent connections and the underlying request queue backlog depth in the real-time high-concurrency data traffic load characteristics are mapped to the corresponding node edges of the cross-layer digital twin topology, establishing a microservice dynamic load capacity model; the cross-layer digital twin topology nodes with abnormal node location markers are extracted as directed graph propagation source points, and the asynchronous non-blocking call protocol and exponential backoff retry network parameters in the call dependency edges of the distributed data processing microservice link tracing sequence are analyzed; based on the directed graph propagation source point, the microservice dynamic load capacity model, and the asynchronous non-blocking call protocol and exponential backoff retry network parameters, a node capacity overflow transfer function based on queuing theory is constructed. In the virtual space, the call chain request congestion state transition caused by the exhaustion of underlying hardware resources is iteratively simulated at time steps, generating a link congestion state evolution sequence in the virtual sandbox environment.

[0011] Furthermore, the specific process of outputting the affected microservice node set and the cascading failure network topology path is as follows: From the link congestion state evolution sequence, trace the derived failure nodes whose simulated request backlog exceeds the capacity limit of the microservice dynamic load level model; perform deduplication and aggregation operations on the derived failure nodes to generate the affected microservice node set; extract the connected subgraph connecting the directed graph propagation source point and the affected microservice node set from the cross-layer digital twin topology graph; remove normal bypass branches in the connected subgraph that do not exhibit congestion transfer characteristics to generate an abnormal propagation connected subgraph; convert the abnormal propagation connected subgraph into a directed acyclic graph containing failure evolution timestamps to generate the cascading failure network topology path.

[0012] Furthermore, the data processing service level protocol attributes of the affected microservice node set are extracted. Combined with the micro-level computing power margin of the backup physical machine cluster and the cross-node migration bandwidth cost of the cascading failure network topology path, the specific process of generating a dynamic management script through multi-objective collaborative game analysis is as follows: The data processing service level protocol attributes corresponding to the affected microservice node set are analyzed to extract the tolerance latency cap and business degradation priority weights for core financial data processing tasks; the available last-level cache capacity and idle network interface controller throughput in the micro-level computing power margin of the backup physical machine cluster are collected to evaluate the cross-node migration bandwidth cost between the source host and the target backup node in the cascading failure network topology path; a mixed-integer programming scheduling model is constructed with the optimization objectives of minimizing the tolerance latency cap default rate and minimizing the cross-node migration bandwidth cost. The business degradation priority weights, available last-level cache capacity, and idle network interface controller throughput are input, and a heuristic constraint solver is used to calculate the optimal resource allocation solution that satisfies the hard constraint of the micro-level computing power margin. The dynamic management script is then compiled based on the optimal resource allocation solution.

[0013] Furthermore, the specific process of distributing the dynamic management script to the physical data center to perform cross-region hot migration of the target microservice and secure isolation of abnormal containers is as follows: Parse the abnormal container security isolation instructions in the dynamic management script, call the control group kernel interface of the host operating system in the physical data center, and implement hard limits on last-level cache allocation and CPU instruction issuance quota compression for microservice container entities with abnormal node location markers; Parse the target microservice cross-region hot migration instructions in the dynamic management script, perform incremental memory page synchronization and checkpoint snapshot capture for high-priority microservice containers in the affected microservice node set, and transmit the checkpoint snapshot to the backup physical machine cluster for container state recovery; After the backup physical machine cluster completes container state recovery, update the virtual LAN routing table and distributed microservice registration and discovery center of the physical data center to complete the cross-region hot migration of the target microservice.

[0014] The data center intelligent operation and maintenance system based on digital twins is used to execute the aforementioned data center intelligent operation and maintenance method based on digital twins. It includes: a twin mapping module, used to acquire the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice link tracing sequence, mapping the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice link tracing sequence to a pre-constructed cross-layer digital twin topology graph to generate hardware and software operating status slice features; and an anomaly localization module, used to perform graph attention network aggregation on the data processing microservice response latency gradient and hardware last-level cache miss rate in the hardware and software operating status slice features, extracting micro-resource contention feature vectors, and calculating cross-container resource preemption interference within the physical host based on the micro-resource contention feature vectors. The system performs several steps: First, it establishes a disturbance index and marks abnormal nodes. Second, it uses a sandbox evolution module to analyze cross-layer digital twin topology nodes with abnormal node markings. This module, combined with real-time high-concurrency data traffic load characteristics, traces the call dependency edges along the distributed data processing microservice link to perform sandbox fault propagation evolution, outputting the set of affected microservice nodes and the cascading failure network topology path. Third, it uses a dynamic management module to extract the data processing service level protocol attributes of the affected microservice node set. Combining the micro-computing power margin of the backup physical machine cluster and the cross-node migration bandwidth cost of the cascading failure network topology path, it generates a dynamic management script through multi-objective collaborative game analysis. This script is then distributed to the physical data center to execute cross-region hot migration of the target microservice and secure isolation of abnormal containers.

[0015] The present invention has the following beneficial effects: (1) A data center intelligent operation and maintenance method based on digital twins maps the telemetry timing matrix of the underlying server microarchitecture and the link tracing sequence of the distributed data processing microservices to a pre-constructed cross-layer digital twin topology graph, realizing deep integration and slice feature extraction of the operating status of software and hardware. On this basis, a graph attention network is used to aggregate the microservice response latency gradient and the hardware last-level cache miss rate, which can accurately extract the micro-resource competition feature vector and then calculate the cross-container resource preemption interference index. This design breaks the barrier of software and hardware data isolation in traditional operation and maintenance, effectively solves the problem of lagging identification of micro-resource competition and cross-layer interference in existing methods, and realizes accurate location and marking of abnormal nodes in the incubation period.

[0016] (2) The data center intelligent operation and maintenance system based on digital twins, for cross-layer digital twin topology nodes with abnormal node location markers, combines real-time high-concurrency data traffic load characteristics, and performs sandbox fault propagation evolution along the call dependency edges in the virtual space. It can losslessly deduce and output the set of affected microservice nodes and the cascading failure network topology path. Subsequently, it comprehensively extracts service level protocol attributes, micro-computing power margin of physical machine clusters, and cross-node migration bandwidth costs, generates dynamic control scripts through multi-objective collaborative game analysis, and distributes the scripts to the physical environment to execute target microservice cross-region hot migration and abnormal container security isolation. This mechanism overcomes the secondary disaster risks caused by blind scheduling in the physical environment directly by existing methods, and achieves globally optimal resource reallocation and secure isolation control of high-concurrency services.

[0017] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intelligent data center operation and maintenance method based on digital twins according to the present invention.

[0019] Figure 2 A schematic diagram illustrating the dynamic identification of cross-container resource preemption interference.

[0020] Figure 3 This is a schematic diagram illustrating the cascading failure evolution of sandbox links due to congestion.

[0021] Figure 4 This is a schematic diagram of scheduling optimization in a multi-objective collaborative game.

[0022] Figure 5 This is a flowchart of the intelligent data center operation and maintenance system based on digital twins, as described in this invention. Detailed Implementation

[0023] This application's embodiments address the problems of isolated software and hardware monitoring leading to difficulty in detecting micro-anomalies and the potential for secondary disasters caused by direct scheduling of the physical environment during existing data center operations and maintenance processes, through a data center intelligent operation and maintenance method and system based on digital twins.

[0024] The overall approach of the scheme in this application embodiment is as follows: First, obtain the running data of the underlying microarchitecture and the upper-layer microservice link and map it to a cross-layer digital twin topology graph to generate hardware and software running status slice features. Then, perform graph attention network aggregation on the latency gradient and cache miss rate to extract micro features and complete the abnormal node location and marking. Subsequently, perform fault propagation evolution along the link dependency edge in a virtual sandbox environment to output the affected nodes and cascade failure paths. Finally, perform multi-objective game analysis based on the cost features such as business protocol attributes and computing power margin to generate dynamic management scripts, and distribute them to the physical environment to execute microservice hot migration and abnormal container security isolation.

[0025] Example 1; please refer to Figure 1 This invention provides a technical solution: a data center intelligent operation and maintenance method based on digital twins, comprising the following steps: S1, obtaining the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice link tracing sequence, mapping the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice link tracing sequence to a pre-constructed cross-layer digital twin topology graph, generating hardware and software operating status slice features; S2, performing graph attention network aggregation on the data processing microservice response latency gradient and hardware last-level cache miss rate in the hardware and software operating status slice features, extracting micro-resource contention feature vectors, and calculating the cross-container resource preemption interference index within the physical host based on the micro-resource contention feature vectors. And mark the abnormal nodes; S3, for the cross-layer digital twin topology nodes with abnormal node location marks, combined with the real-time high-concurrency data traffic load characteristics, trace the call dependency edge along the distributed data processing microservice link to carry out sandbox fault propagation evolution, and output the set of affected microservice nodes and the cascaded failure network topology path; S4, extract the data processing service level protocol attributes of the affected microservice node set, combine the micro-computing power margin of the backup physical machine cluster and the cross-node migration bandwidth cost of the cascaded failure network topology path, generate dynamic management script through multi-objective collaborative game analysis, and send the dynamic management script to the physical data center to execute the cross-region hot migration of the target microservice and the security isolation of abnormal containers.

[0026] In this implementation plan, step S1 involves collecting and integrating the underlying hardware operation data and the upper-layer software business data of the data center to construct a unified virtual mapping model. In this step, the underlying server microarchitecture telemetry timing matrix refers to a data set recording the underlying hardware operating status of various internal components of the central processing unit, buses, and caches in chronological order; the distributed data processing microservice tracing sequence refers to a log chain recording the execution path, sequence, and time consumed when a specific computing task crosses multiple different software service modules; and the cross-layer digital twin topology map refers to a network structure model that establishes physical and logical connections between physical hardware device nodes and the software application nodes running on them in virtual space. The technical significance of this step lies in completely breaking down the data isolation barriers between traditional data center hardware monitoring systems and software business monitoring systems, providing a panoramic data foundation with extremely high data granularity and time synchronization for subsequent anomaly diagnosis.

[0027] Step S2 utilizes a graph neural network algorithm to analyze the fused feature data, identify micro-anomalies caused by contention for underlying hardware resources, and precisely pinpoint the source of the fault. In this step, the data processing microservice response latency gradient refers to the rate of change and fluctuation trend of time spent by the software service processing business requests; the hardware last-level cache miss rate refers to the frequency at which the central processing unit (CPU) cannot find the required data in the outermost cache and is forced to read data from the slower main memory; graph attention network aggregation is a deep learning algorithm that dynamically allocates different weight indicators based on the severity of mutual interference between adjacent nodes in the graph and performs information fusion calculations; the cross-container resource contention interference index is a specific value used to quantitatively assess the severity of performance loss caused by different software operating environments competing for limited underlying hardware resources on the same physical server. The technical role of this step is to accurately identify micro-resource competition phenomena that traditional overall utilization monitoring cannot detect, and to preemptively identify micro-anomaly nodes in their incubation period before large-scale business congestion and lag occur.

[0028] Step S3 uses the located abnormal node as a starting point to simulate and deduce the spread and propagation trend of the abnormal state under the current business traffic in a virtual environment. In this step, the real-time high-concurrency data traffic load characteristics represent the scale and density of massive concurrent user access requests currently being handled by the data center; call dependency edges represent the logical binding relationship in the software architecture where one business module must wait for another business module to complete data processing before it can continue to run; sandbox fault propagation evolution refers to the testing process in a virtual test space completely isolated from the real physical environment, using algorithms to simulate how the depletion of local hardware resources triggers a chain reaction in the upper-layer software links; cascading failure network topology path refers to the propagation route predicted by the system from the initial abnormal source, gradually leading to congestion or collapse of other related hardware and software nodes in the surrounding area. The technical role of this step is to utilize the zero-risk trial-and-error characteristics of digital twins to proactively quantify the business scope and damage boundaries that local micro-faults may affect in the future, providing a scientific and forward-looking basis for formulating precise loss mitigation strategies.

[0029] Step S4 automatically calculates and executes optimal service scheduling and resource isolation commands based on the aforementioned evolutionary deduction results and actual physical resource conditions. In this step, the data processing service level agreement attributes refer to the pre-defined service degradation priorities, maximum tolerable latency times, and availability assurance indicators based on the importance of different services; cross-node migration bandwidth cost refers to the amount of underlying network transmission resources consumed when transferring the running software application state from one physical server to another; multi-objective collaborative game analysis is an algorithmic approach that uses mathematical solutions to find the optimal balance among multiple conflicting optimization objectives, such as ensuring core business stability, reducing network bandwidth consumption, and fully utilizing backup computing power; cross-region hot migration refers to the operational technique of smoothly transferring threatened software applications to a secure backup server cluster without interrupting current user service access. The technical advantage of this step is that it avoids the secondary downtime risks caused by blind scheduling based on manual experience in traditional operations and maintenance, achieving intelligent adaptive reorganization of computing resources. While ensuring that core high-concurrency data processing services are not affected, it securely isolates the source of abnormal underlying resource consumption.

[0030] Specifically, the process of obtaining the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice tracing sequence is as follows: A kernel-level probe pre-installed on the physical server operating system collects the hardware microinstruction retirement rate, bus bandwidth usage, and cache hit allocation status of each level of the central processing unit. The hardware microinstruction retirement rate, bus bandwidth usage, and cache hit allocation status are then aligned and concatenated according to timestamps to construct the underlying server microarchitecture telemetry timing matrix. A non-intrusive tracing agent is injected into the distributed data processing microservice mesh to capture the request enqueue timestamps and execution completion timestamps of cross-container remote procedure calls. These timestamps are then topologically assembled according to the microservice call dependency hierarchy to generate the distributed data processing microservice tracing sequence.

[0031] In this implementation scheme, to achieve precise spatiotemporal synchronization between underlying physical resources and upper-layer software calls and to eliminate the performance overhead of traditional application-layer monitoring, the underlying hardware registers of the central processing unit's performance monitoring unit are first accessed directly through an operating system kernel-level probe. The collected hardware operating states are strictly aligned chronologically to construct a telemetry timing matrix for the underlying server microarchitecture. Specifically, at any collection timestamp t, a telemetry feature vector is constructed. ,in, This represents the hardware microinstruction retirement rate characteristic value under the data collection timestamp; This indicates the characteristic value of bus bandwidth utilization; This represents the feature vector indicating the cache hit allocation status at each level. The telemetry feature vectors from multiple consecutive timestamps are combined to form a complete telemetry timing matrix for the underlying server microarchitecture. Simultaneously, a non-intrusive tracing plugin is injected into the data plane sidecar proxy of the microservice mesh to intercept and record remote procedure call data between microservices. For any cross-container call chain in the network mesh, the request enqueue timestamp and execution end timestamp are extracted to calculate the actual processing dwell time of a single call. This is then combined with the logical call hierarchy between business microservices to construct a distributed data processing microservice chain tracing sequence. The purpose of this step is to obtain the independent operating states of the hardware and software at the finest granular level, providing a high-fidelity raw physical and business data source for subsequent twin mapping.

[0032] Specifically, the process of mapping the underlying server microarchitecture telemetry time series matrix and the distributed data processing microservice tracing sequence to a pre-built cross-layer digital twin topology graph to generate hardware and software runtime status slice features is as follows: Parse the infrastructure i.e., code configuration file of the physical data center to construct a multi-level entity graph containing physical host nodes, container nodes, and microservice application nodes; extract the hardware host identifier from the underlying server microarchitecture telemetry time series matrix and the container runtime environment identifier from the distributed data processing microservice tracing sequence; establish cross-layer resource binding edges in the multi-level entity graph based on the hardware host identifier and the container runtime environment identifier to generate a cross-layer digital twin topology graph; perform a sliding window truncation of the cross-layer digital twin topology graph along the time axis at a set step size; and perform tensor concatenation of the underlying server microarchitecture telemetry time series matrix features and the distributed data processing microservice tracing sequence features along the node feature dimension within the same time segment to generate hardware and software runtime status slice features.

[0033] In this implementation scheme, to establish an accurate physical and logical mapping between fragmented hardware and software data in the digital space, the Infrastructure as Code (ICC) automated deployment configuration file is first parsed to construct a multi-level entity graph consisting of physical host nodes, container nodes, and microservice application nodes from bottom to top. By matching the hardware host identifier in telemetry data with the container runtime environment identifier in link tracing data, vertically related cross-level resource binding edges are established between nodes at different levels of the entity graph, forming a cross-level digital twin topology graph. This topology graph not only includes horizontal microservice call dependencies but also establishes vertical mapping paths from software containers to physical hosts. Subsequently, an observation interval of size W is truncated along the time axis with a set step size L, and feature alignment and tensor concatenation calculations are performed on the multimodal runtime data within the cross-level digital twin topology graph. Specifically, for a specific topology node n within the same time segment, hardware and software runtime state slice features are generated. The calculation formula is expressed as follows: ,in, This represents a sub-feature of the telemetry time series matrix of the underlying server microarchitecture corresponding to node n in the k-th sampling section; This represents the sub-feature of the distributed data processing microservice link tracing sequence corresponding to node n in the same sampling section; This represents the tensor concatenation operator along the feature channel dimension; This represents the time aggregation weighting coefficient. Considering that data closer to the current time truncation point has greater significance in representing the instantaneous operating state of the system, the method disclosed above for determining the time aggregation weighting coefficient is to dynamically calculate it using an exponentially decaying smoothing function, i.e. ,in The time-forgetting decay factor is obtained by fitting autocorrelation analysis data of hardware and software load fluctuations under historical normal operating conditions. This step aims to deeply fuse discrete and heterogeneous hardware and software operating characteristics under a unified digital twin spatiotemporal benchmark, integrating temporal and node dimensions. This generates a tensor representation encompassing the entire link from microservice business logic to the underlying hardware microarchitecture, laying a multi-dimensional input foundation for subsequent feature extraction and hidden resource contention anomaly analysis in deep graph neural networks.

[0034] Specifically, the process of using a graph attention network to aggregate the response latency gradient of data processing microservices and the hardware last-level cache miss rate in the hardware and software operation status slice features to extract the micro-resource competition feature vector is as follows: First-order difference operations are performed along the time dimension on the distributed data processing microservice link tracing sequence features in the hardware and software operation status slice features to extract the dynamic response time change rate of each microservice application node and construct the data processing microservice response latency gradient; the cache line eviction event sequence and multi-core parallel access conflict frequency of physical host nodes are separated from the underlying server microarchitecture telemetry time series matrix features in the hardware and software operation status slice features, and combined with the latency penalty coefficient of the micro-instruction retirement cycle, the hardware last-level cache miss rate is constructed; the data processing microservice response latency gradient and the hardware last-level cache miss rate are used as the initial node features input to the graph attention network layer, and information aggregation weights are assigned to adjacent nodes in the cross-layer digital twin topology graph through a masked self-attention mechanism. The features of adjacent nodes are then weighted and summed to activate them according to the information aggregation weights, outputting the micro-resource competition feature vector.

[0035] In this implementation scheme, to accurately capture the implicit performance degradation of upper-layer services caused by contention for underlying physical resources, the system first performs a first-order difference operation on the link tracing sequence features of the same node within adjacent time segments to extract the response latency gradient of the data processing microservices reflecting the trend of service lag. Simultaneously, it extracts the physical host machine cache eviction event sequence and the frequency of multi-core parallel access conflicts, and combines this with the latency penalty coefficient of the microinstruction retirement cycle to construct the hardware final-level cache miss rate. Since logically independent microservices in a cloud-native environment may run on the same physical machine and share cache and memory bus, the aforementioned latency gradient and miss rate need to be combined as initial features and input into the graph attention network layer for cross-layer correlation analysis. Specifically, for the target microservice node u in the cross-layer digital twin topology graph, the information aggregation weights of adjacent nodes v are calculated. The calculation formula is expressed as: Subsequently, the features of adjacent nodes are weighted and summed for activation to output a micro-resource competition feature vector. The calculation formula is expressed as: In the above formula, u represents the target microservice node in the cross-layer digital twin topology graph; v represents the adjacent node that has a cross-layer association mapping with the target microservice node. This represents the set of all neighboring nodes of the target microservice node; This represents the initial feature vector of the target microservice node; This represents the initial feature vector of the neighboring node v; This represents the initial feature vector of any node r in the set of adjacent nodes; This represents the linear mapping weight matrix used for feature space alignment; This represents the learnable parameter vector in the masked self-attention mechanism; This represents the information aggregation weight assigned by adjacent nodes to the target microservice node; This represents the extracted micro-level resource competition feature vector; This represents a non-linear activation function. This step dynamically quantifies the actual interference level of different physically adjacent containers on the target microservice through a graph attention mechanism, breaking down the analytical gap between surface software metrics and underlying hardware metrics, and effectively extracting micro-resource preemption features that are easily overlooked in high-concurrency scenarios.

[0036] Please see Figure 2 Specifically, the process of calculating the cross-container resource preemption interference index within the physical host and marking abnormal nodes based on the micro-resource competition feature vector is as follows: A multilayer perceptron is used to perform a nonlinear mapping on the micro-resource competition feature vector, outputting a cross-container resource preemption interference index that represents the probability of hardware bus queuing conflicts between multiple data processing microservices within the same physical host; the exponential moving average and standard deviation of the cross-container resource preemption interference index are calculated within a time sliding window to construct a dynamic adaptive interference threshold; the difference between the current cross-container resource preemption interference index and the dynamic adaptive interference threshold is compared, and microservice container entities whose differences exceed the tolerance range multiple times consecutively are extracted, and abnormal node location marking is performed in the cross-layer digital twin topology graph.

[0037] In this implementation scheme, considering that the overall resource load of a data center will exhibit a normal tidal phenomenon of drastic fluctuations when facing sudden high-concurrency traffic, using a static fixed threshold is prone to causing large-scale false alarms. Therefore, the system first uses a multilayer perceptron to perform a high-dimensional nonlinear mapping on the micro-resource contention feature vector, reducing its dimensionality to a one-dimensional cross-container resource preemption interference index. Subsequently, within a continuous time sliding window, the dynamic baseline and standard deviation of the interference index are calculated based on the exponential moving average algorithm, thereby constructing a dynamically adaptive interference threshold that fluctuates elastically with normal business traffic. Specifically, the calculation process is as follows: at the current time step d, the dynamic adaptive interference threshold... The calculation formula is expressed as: ; where the exponential moving average The derivation formula is expressed as: Standard deviation The derivation formula is expressed as: In the above formula, d represents the index of the current discrete time step; d-1 represents the index of the previous historical time step. This represents the dynamic adaptive interference threshold for the current time step. This represents the cross-container resource preemption interference index input at the current time step; This represents the exponentially moving average of the disturbance index at the current time step. This represents the exponentially moving average of the disturbance index at the previous time step. This represents the standard deviation of the disturbance index at the current time step. This represents the standard deviation of the disturbance index at the previous time step; The smoothing weighting coefficient represents the sliding time window; This represents the tolerance sensitivity adjustment parameter. Regarding the method for determining the above key parameters, the smoothing weighting coefficient... Publicly available through formula ;Calculated and obtained, parameter N represents the total number of sampling points within the preset sliding time window; tolerance sensitivity adjustment parameter The parameter is dynamically negatively correlated with the latency sensitivity level in the data processing service level agreement attribute corresponding to the microservice, meaning the higher the service level, the smaller the value of this adjustment parameter. Finally, the current interference index is calculated. With threshold The difference is only flagged as an anomaly in the cross-layer digital twin topology graph when the difference is greater than zero for a specified number of consecutive times. This step upgrades the hard comparison of absolute values ​​to a flexible anti-jitter and fault-tolerant judgment based on probability distribution trends, which can accurately filter out instantaneous performance spikes caused by normal traffic surges and completely solve the problem of the difficulty in accurately locating hidden physical resource contention interference in cloud-native environments.

[0038] Please see Figure 3Specifically, for cross-layer digital twin topology nodes with abnormal node location markers, and combined with real-time high-concurrency data traffic load characteristics, the specific process of sandbox fault propagation evolution along the call dependency edges of the distributed data processing microservice link tracing sequence is as follows: The total number of concurrent connections and the underlying request queue backlog depth in the real-time high-concurrency data traffic load characteristics are mapped to the corresponding node edges of the cross-layer digital twin topology, establishing a microservice dynamic load capacity model; the cross-layer digital twin topology nodes with abnormal node location markers are extracted as directed graph propagation source points, and the asynchronous non-blocking call protocol and exponential backoff retry network parameters in the call dependency edges of the distributed data processing microservice link tracing sequence are analyzed; based on the directed graph propagation source point, the microservice dynamic load capacity model, and the asynchronous non-blocking call protocol and exponential backoff retry network parameters, a node capacity overflow transfer function based on queuing theory is constructed. In the virtual space, the call chain request congestion state transition caused by the exhaustion of underlying hardware resources is iteratively simulated at time steps, generating a link congestion state evolution sequence in the virtual sandbox environment.

[0039] In this implementation plan, to accurately predict the scope of business avalanche caused by the depletion of underlying resources without interfering with real business operations, the system introduces a virtual space for evolutionary simulation. Modern distributed systems generally employ asynchronous non-blocking call protocols, meaning that after initiating a network request, the caller does not suspend the current thread to wait for a response, but instead processes the returned result by registering a callback function; simultaneously, this is combined with exponential backoff retry network parameters, meaning that when encountering downstream node congestion or denial of service, the caller will initiate retry requests at exponentially increasing time intervals. This mechanism can provide protection during normal system fluctuations, but when micro-congestion occurs due to the depletion of underlying hardware resources, it can lead to an exponential backflow of retry requests into the caller's memory. Therefore, the system first establishes a dynamic microservice capacity water level model for each node in the cross-layer digital twin topology graph by mapping the total number of concurrent connections and the depth of the underlying request queue, to define the upper limit of the request processing capacity of each node. Subsequently, abnormal nodes are extracted as the source points for directed graph propagation, and a node capacity overflow transfer function based on queuing theory is constructed to simulate the congestion state transition over time. The specific calculation process is as follows: under any simulation time step s in the virtual sandbox environment, for any upstream microservice node y in the topology graph, its simulated queue backlog depth is... The state update formula is expressed as: ; where the overflow transfer function This represents the incremental backlog caused by severe congestion at downstream node x, leading to request traffic flowing back to upstream node y. Its calculation formula is as follows: In the above formula, s represents the current discrete time step of the virtual sandbox simulation; s+1 represents the next simulation time step; y represents the upstream microservice node that initiates the network call; and x represents the downstream microservice node that receives the network call. This represents the set of all downstream microservice nodes that the upstream microservice node y depends on; This represents the simulated queue depth of the upstream microservice node y at the current time step; This represents the simulated queue depth of the upstream microservice node y in the next time step; This represents the basic service arrival rate of node y from the external network. This represents the current business processing release rate of node y; Indicates the step size and time interval of the sandbox simulation; This represents the capacity overflow transfer function value applied in reverse from downstream node x to upstream node y; This indicates the current simulated queue backlog depth of downstream microservice node x; This represents the arrival rate of concurrent requests routed from node y to node x. This represents the maximum queue capacity limit of the downstream microservice node x, as determined by the microservice dynamic capacity level model. This represents the retry request traffic amplification factor; This represents the current value of the exponential backoff retry network counter at the current time step. It relates to the retry request traffic amplification factor. The publicly disclosed method for determining the congestion state is to calculate the ratio of the average data volume of retry request packets caused by network timeouts to the average data volume of the first normal request packets in the historical data center operation and maintenance logs. This step, through rigorous queuing theory state transition equations, realistically recreates the physical process of how abnormal traffic backflows and cascades in a complex microservice mesh after a single point of failure caused by underlying micro-resource contention. This generates a link congestion state evolution sequence in a virtual sandbox environment, providing a high-fidelity data extrapolation basis for accurately defining the scope of system disasters.

[0040] Specifically, the process of outputting the affected microservice node set and the cascading failure network topology path is as follows: From the link congestion state evolution sequence, trace the derived fault nodes whose simulated request backlog exceeds the capacity limit of the microservice dynamic load level model; perform deduplication and aggregation operations on the derived fault nodes to generate the affected microservice node set; extract the connected subgraph connecting the directed graph propagation source point and the affected microservice node set from the cross-layer digital twin topology graph; remove normal bypass branches in the connected subgraph that do not exhibit congestion transfer characteristics to generate an abnormal propagation connected subgraph; convert the abnormal propagation connected subgraph into a directed acyclic graph containing fault evolution timestamps to generate the cascading failure network topology path.

[0041] In this implementation plan, to accurately identify truly threatened business links from massive evolutionary sequence data and avoid ineffective scheduling during subsequent cross-region migration, the system needs to perform topology path extraction and pruning on the generated link congestion state evolution sequence. First, the simulated queue backlog depth of each microservice node in the sequence is continuously tracked. When the backlog of a node exceeds the capacity limit specified by its microservice dynamic load capacity model, it is identified as a derived failure node. A hash deduplication and aggregation operation is performed on all triggered derived failure nodes to eliminate duplicate statistics caused by the intersection of multiple complex call paths, thereby generating a clean set of affected microservice nodes. Next, an initial connected subgraph is defined in the cross-layer digital twin topology graph, connecting the directed graph propagation source point to all nodes within this set. Since some normal bypass branches in the connected subgraph are merely logical interconnections but do not exhibit obvious congestion transfer characteristics in actual simulations, the system must calculate the congestion transfer strength of each call dependency edge in the connected subgraph and remove them. Specifically, the calculation process involves calculating the congestion transfer strength of any directed dependency edge in the initial connected subgraph. Its congestion transfer intensity The calculation formula is expressed as: When congestion transfer intensity Greater than the background flow fluctuation noise threshold If a branch exhibits substantial congestion propagation, the directed dependency edge is retained; otherwise, it is removed. After comprehensive pruning, an anomalous propagation connected subgraph is generated. In the above formula, This represents a directed dependency edge from upstream microservice node y to downstream microservice node x; This represents the cumulative characteristic value of the congestion transfer intensity of the dependent edge throughout the entire sandbox simulation period; This indicates the total number of time steps set in the sandbox simulation; This represents the capacity overflow transfer function value obtained at the simulation time step s; Indicates the time step interval for the simulation; This represents the background flow fluctuation noise threshold used to filter invalid branches. The method for determining the threshold is as follows: during the historical period of normal and stable operation of the statistical data center, the 95th percentile of the cumulative distribution of one-sided retry traffic caused by occasional network jitter among the microservices in the entire grid is used as the filtering boundary value. Finally, the fault evolution timestamp of the first time each retained node exceeds the capacity limit is recorded. Based on the absolute chronological order of the timestamps, the node connection relationships in the abnormal propagation connected subgraph are topologically sorted and converted into a directed acyclic graph structure that does not allow loops. The purpose of this step is to thoroughly eliminate false positives and accurately filter out invalid service nodes that are connected in the topological space but not physically affected by congestion. The output structure is extremely clear and contains a strict fault propagation order of the cascading failure network topology path, enabling the system to accurately grasp the propagation direction of service avalanche and providing decisive path constraints for subsequent security isolation and hot migration strategies.

[0042] Please see Figure 4 Specifically, the process of extracting the data processing service level agreement (SSPA) attributes of the affected microservice node set, combining the micro-level computing power margin of the standby physical machine cluster and the cross-node migration bandwidth cost of the cascading failure network topology path, and generating a dynamic management script through multi-objective collaborative game analysis is as follows: Analyze the data processing SSPA attributes corresponding to the affected microservice node set, extract the tolerance latency cap and business degradation priority weights for core financial data processing tasks; collect the available last-level cache capacity and idle network interface controller throughput from the micro-level computing power margin of the standby physical machine cluster, and evaluate the cross-node migration bandwidth cost between the source host and the target standby node in the cascading failure network topology path; construct a mixed-integer programming scheduling model with the optimization objectives of minimizing the tolerance latency cap default rate and minimizing the cross-node migration bandwidth cost, inputting the business degradation priority weights, available last-level cache capacity, and idle network interface controller throughput, and using a heuristic constraint solver to calculate the optimal resource allocation solution that satisfies the hard constraint of the micro-level computing power margin, and compile and generate a dynamic management script based on the optimal resource allocation solution.

[0043] In this implementation scheme, to find the optimal scheduling solution that simultaneously considers service level agreement (SLA) constraints and underlying hardware resource limitations in a complex physical environment, the system first analyzes the SLA attributes of the affected microservice node set, extracting the tolerable latency cap set for core financial data processing tasks, and the priority weight for allowing the system to perform degraded services with compromised guarantees in extreme cases. Simultaneously, by calling physical environment probes to collect the micro-state of the backup physical machine cluster, the available last-level cache capacity and idle network interface controller throughput are obtained to accurately assess the bandwidth cost required for cross-node state migration in a cascading failure network topology. Based on the extracted multi-dimensional hardware and software indicators, the system constructs a mixed-integer programming scheduling model with the joint optimization objective of minimizing the default rate of the tolerable latency cap and minimizing the cross-node migration bandwidth cost. Specifically, the joint optimization objective function... The derivation formula is expressed as: Meanwhile, to ensure that the new node is not overwhelmed, a hard constraint condition for micro-computing power margin is established, the formula of which is expressed as: In the above formula, This represents the calculated value of the joint optimization objective function; This indicates the set of microservice nodes affected. represents the set of available nodes in the standby physical machine cluster; p represents any microservice node to be migrated in the set of affected microservice nodes; q represents any candidate target node in the standby physical machine cluster. This represents a discrete Boolean decision variable, which takes the value of one if and only if node p is scheduled and assigned to node q by the system, and otherwise takes the value of zero. This represents the objective preference weight coefficient in a multi-objective cooperative game; This indicates the priority weight of the service degradation corresponding to node p; This represents the predicted service execution latency based on the current network topology after migrating node p to node q. This represents the upper limit of the tolerable delay for node p; This represents the size of the memory snapshot data required to represent the running state of node p; This represents the idle network interface controller throughput of candidate target node q; This represents the minimum last-level cache capacity required for node p to operate normally and smoothly. This represents the available last-level cache capacity monitored in real time for candidate target node q. The target preference weight coefficients in the above formula... The method for determining this is disclosed as follows: A dynamic linear mapping is performed by calculating the ratio of the total number of real-time concurrent transaction requests received by the data center gateway at the current scheduling moment to the peak request volume during the same period over the past thirty days. This ensures that during peak financial transaction periods, the system can adaptively allocate computing resources to minimize business latency and default rates. Subsequently, the system uses a heuristic constraint solver to iteratively search within a complex discrete solution space to find a solution that satisfies the objective function... The Boolean decision variable combination matrix that minimizes and satisfies hard resource capacity constraints is then compiled into a machine-executable dynamic management script based on the matrix result. This step perfectly transforms the macroscopically vague business continuity and security requirements with the microscopically rigid physical hardware limitations into a rigorous mathematical programming problem, fundamentally avoiding secondary congestion and avalanche of backup clusters caused by blind scheduling based on traditional manual experience.

[0044] Specifically, the process of distributing the dynamic management script to the physical data center to perform cross-region hot migration of the target microservice and secure isolation of abnormal containers is as follows: Parse the abnormal container security isolation instructions in the dynamic management script, call the control group kernel interface of the host operating system in the physical data center, and implement hard limits on last-level cache allocation and CPU instruction issuance quota compression for microservice container entities with abnormal node location markers; Parse the target microservice cross-region hot migration instructions in the dynamic management script, perform incremental memory page synchronization and checkpoint snapshot capture for high-priority microservice containers in the affected microservice node set, and transmit the checkpoint snapshot to the backup physical machine cluster for container state recovery; After the backup physical machine cluster completes container state recovery, update the virtual LAN routing table and distributed microservice registration and discovery center of the physical data center to complete the cross-region hot migration of the target microservice.

[0045] In this implementation plan, after generating the dynamic management script, the system must seamlessly translate it into actual operational instructions for the underlying physical data center to complete fault isolation and service recovery. First, the system parses the abnormal container security isolation instructions in the dynamic management script. By directly calling the host operating system's control group kernel low-level interface, it implements precise hardware-level confinement on the microservice container entities that were previously marked as abnormal nodes. Specifically, the system not only imposes strict hard address space restrictions on their final-level cache allocation but also executes a CPU instruction issuance quota compression algorithm to reduce the instruction throughput quota. The recalculated formula is expressed as follows: In the above formula, This indicates the CPU instruction emission quota for the abnormal container that has been re-approved after compression. This indicates the original launch quota prior to the implementation of security isolation intervention; This represents the isolation penalty severity factor configured globally within the system. This indicates the severity value of cross-container resource preemption assessed in the twin spatiotemporal analysis of the abnormal microservice container entity. Through this underlying constraint, the system can instantly block excessive predation of critical shared hardware buses without directly terminating the container process, playing a crucial role in controlling system resource bleeding points. Simultaneously, the system parses the cross-region hot migration instructions for the target microservice in the script, initiating seamless cross-physical node business flow for high-priority microservice containers within the affected microservice node set. To ensure zero-interruption awareness for financial business clients, the system rapidly captures checkpoint snapshots on the source host side and calculates the dynamic migration bandwidth allocation for performing incremental memory page synchronization. The calculation formula is expressed as: In the above formula, This represents the incremental synchronization network bandwidth of memory pages specifically allocated for this high-performance microservice container; This represents the total amount of dirty memory page data generated during the continuous operation of the content container in the hot migration preparation and iterative transmission phases; This indicates the maximum unresponsive service interruption time allowed in the service level agreement for core financial data processing tasks; This represents the fixed physical time consumed by the source host machine to capture checkpoint snapshots and initialize network handshake connections. Based on this rigorous bandwidth allocation strategy, the system safely and smoothly transfers the operational state to the backup physical machine cluster for container state recovery without triggering a network storm at the bottom layer of the entire data center. Finally, after the backup physical machine successfully takes over the memory state and process handles, the system uses a software-defined networking protocol to send routing update signaling to the control plane of the physical data center, simultaneously refreshing the virtual LAN routing table and node addressing records in the distributed microservice registration and discovery center. This accurately directs new business requests from the external network to the healthy backup physical machine cluster, completing the entire fault self-healing process. The purpose of this step is to perfectly implement the virtual sandbox's deductive decision-making at the physical execution layer, achieving hard-core isolation of high-risk containers and hot-swapping of core microservices across regions without the business being aware of it, completely eliminating the risk of data center cascading failures caused by micro-resource contention.

[0046] Example 2; please refer to Figure 5A data center intelligent operation and maintenance system based on digital twins, used to execute the data center intelligent operation and maintenance method based on digital twins described in the embodiments, includes: a twin mapping module, used to obtain the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice link tracing sequence, and map the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice link tracing sequence to a pre-constructed cross-layer digital twin topology graph to generate hardware and software operation status slice features; an anomaly localization module, used to perform graph attention network aggregation on the data processing microservice response latency gradient and the hardware last-level cache miss rate in the hardware and software operation status slice features, extract micro-resource contention feature vectors, and calculate cross-container resource contention within the physical host based on the micro-resource contention feature vectors. The system includes: an interference index and anomaly node location marking; a sandbox evolution module, used for cross-layer digital twin topology graph nodes with anomaly node location markings, combined with real-time high-concurrency data traffic load characteristics, to trace the call dependency edges along the distributed data processing microservice link sequence to perform sandbox fault propagation evolution, outputting the set of affected microservice nodes and cascading failure network topology paths; and a dynamic management module, used to extract the data processing service level protocol attributes of the affected microservice node set, combined with the micro-computing power margin of the backup physical machine cluster and the cross-node migration bandwidth cost of the cascading failure network topology path, to generate dynamic management scripts through multi-objective collaborative game analysis, and to distribute the dynamic management scripts to the physical data center to execute cross-region hot migration of target microservices and secure isolation of abnormal containers.

[0047] In this implementation plan, the twin mapping module serves as the data foundation and digital space construction basis for the entire intelligent operation and maintenance system. Its main function is to break down data silos between the underlying hardware devices and upper-layer software services in the data center. This module collects and integrates cross-dimensional telemetry time-series matrices and link tracing sequences to accurately align and map the micro-hardware status of physical entities with the business call relationships of the logical architecture in the digital twin space. This provides the system with a high-fidelity, spatiotemporally synchronized panoramic hardware and software map, providing necessary multimodal data input for subsequent in-depth anomaly analysis and fault inference. The anomaly localization module plays a crucial role in the intelligent diagnosis of micro-level faults within the system. Unlike traditional monitoring components that rely on passive alarms based on macro-level resource utilization, this module focuses on the extremely hidden cross-container resource contention issues in cloud-native environments. It utilizes a graph network model to perform deep feature mining on the fused hardware and software slices, enabling it to keenly detect performance degradation in upper-layer microservices caused by competition for underlying shared physical resources. Before widespread congestion and outages occur, it can proactively identify and precisely mark latent micro-anomaly nodes, achieving very early proactive fault detection. The sandbox evolution module plays a crucial role in zero-risk fault simulation and damage boundary prediction. After the anomaly localization module discovers the source of micro-congestion, this module fully leverages the isolation characteristics of the digital twin environment, combining real-time concurrent traffic load and actual business call dependencies to accelerate the simulation of the chain reaction that may be triggered by the depletion of underlying hardware resources in a virtual sandbox. Through this forward-looking virtual simulation, the system can accurately quantify the future propagation path of the fault and the set of business nodes that will be affected without interfering with any real physical business operations, defining precise physical and logical boundaries for subsequent loss mitigation and control strategies. The dynamic management module is the final execution hub for realizing intelligent self-healing scheduling of data center resources. This module receives the disaster boundaries defined in the preliminary simulation, comprehensively considers the service level requirements of threatened services as well as the hard physical limitations of backup node computing power and network bandwidth, and finds the globally optimal resource reallocation strategy through multi-dimensional game analysis. It is responsible for directly translating complex computational decisions into underlying physical control commands, precisely executing the security isolation of abnormal containers and the cross-regional hot migration of core microservices. This ensures the continuity of core services while completely cutting off the source of physical failures, completing the automated closed loop of the entire intelligent operation and maintenance management system.

[0048] In summary, this application has at least the following effects: The intelligent operation and maintenance method and system for data centers based on digital twins completely breaks down the data barriers of traditional data center hardware and software monitoring by constructing a cross-layer digital twin topology graph. It utilizes graph attention networks to aggregate microservice latency gradients and underlying hardware cache miss rates, accurately capturing and locating easily overlooked cross-container micro-resource preemption risks in high-concurrency scenarios, effectively filtering system false alarms caused by normal business traffic surges. Simultaneously, relying on a virtual sandbox environment and a capacity overflow propagation model based on queuing theory, it proactively deduces cascading failure paths and business impact boundaries of faults under zero physical risk. Finally, combining core business service level agreement requirements and underlying computing power bandwidth constraints, it automatically generates and distributes dynamic control strategies through multi-objective collaborative game analysis. The overall technical solution not only achieves early proactive perception of latent micro-faults and accurate quantification of their destructive impact boundaries, but also fundamentally avoids secondary congestion disasters easily caused by blind scheduling based on traditional manual experience. While ensuring the seamless and continuous operation of core data processing services, it achieves intelligent adaptive reorganization of underlying computing resources and secure physical isolation of abnormally high-risk containers.

[0049] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A data center intelligent operation and maintenance method based on digital twins, characterized in that, Includes the following steps: S1. Obtain the telemetry timing matrix of the underlying server microarchitecture and the link tracing sequence of the distributed data processing microservice. Map the telemetry timing matrix of the underlying server microarchitecture and the link tracing sequence of the distributed data processing microservice to the pre-built cross-layer digital twin topology map to generate hardware and software running status slice features. S2. Perform graph attention network aggregation on the microservice response latency gradient and hardware last-level cache miss rate in the data processing microservice response latency gradient and hardware last-level cache miss rate in the hardware and software operation status slice features, extract micro-resource competition feature vectors, calculate the cross-container resource preemption interference index in the physical host machine based on the micro-resource competition feature vectors, and mark abnormal nodes. S3. For cross-layer digital twin topology nodes with abnormal node location markers, combined with the real-time high-concurrency data traffic load characteristics, the sandbox fault propagation evolution is carried out by tracing the call dependency edges along the distributed data processing microservice link, and the set of affected microservice nodes and cascading failure network topology paths are output. S4. Extract the data processing service level protocol attributes of the affected microservice node set, combine the micro-computing power margin of the backup physical machine cluster and the cross-node migration bandwidth cost of the cascaded failed network topology path, generate dynamic management scripts through multi-objective collaborative game analysis, and send the dynamic management scripts to the physical data center to execute the cross-region hot migration of the target microservices and the security isolation of abnormal containers.

2. The intelligent data center operation and maintenance method based on digital twins according to claim 1, characterized in that: The specific process for obtaining the underlying server microarchitecture telemetry timing matrix and the distributed data processing microservice tracing sequence is as follows: By collecting the hardware micro-instruction retirement rate, bus bandwidth usage, and cache hit allocation status of each level of the central processing unit through kernel-level probes pre-installed in the physical server operating system, the hardware micro-instruction retirement rate, bus bandwidth usage, and cache hit allocation status of each level are aligned and spliced ​​according to timestamps to construct the underlying server microarchitecture telemetry timing matrix. A non-intrusive tracing proxy is injected into the distributed data processing microservice mesh to capture the request enqueue timestamps and execution end timestamps of cross-container remote procedure calls. The request enqueue timestamps and execution end timestamps are then assembled into a topology according to the microservice call dependency hierarchy to generate a distributed data processing microservice tracing sequence.

3. The intelligent data center operation and maintenance method based on digital twins according to claim 2, characterized in that: The specific process of mapping the underlying server microarchitecture telemetry time series matrix and the distributed data processing microservice link tracing sequence to a pre-constructed cross-layer digital twin topology map to generate hardware and software runtime status slice features is as follows: Parse the physical data center's infrastructure-as-code configuration file to construct a multi-level entity graph containing physical host nodes, container nodes, and microservice application nodes; Extract the hardware host identifier from the telemetry time series matrix of the underlying server microarchitecture and the container runtime environment identifier from the distributed data processing microservice link tracing sequence. Based on the hardware host identifier and the container runtime environment identifier, establish cross-level resource binding edges in the multi-level entity graph and generate a cross-level digital twin topology graph. A sliding window is used to extract the cross-layer digital twin topology map along the time axis at a set step size. The telemetry time series matrix features of the underlying server microarchitecture and the link tracing sequence features of the distributed data processing microservice are tensor-concatenated along the node feature dimension to generate hardware and software running status slice features.

4. The intelligent data center operation and maintenance method based on digital twins according to claim 1, characterized in that: The specific process of using a graph attention network to aggregate the microservice response latency gradient and the hardware last-level cache miss rate from the hardware and software runtime status slice features to extract the micro-resource contention feature vector is as follows: The distributed data processing microservice link tracing sequence features in the hardware and software operation status slice features are subjected to first-order difference operation along the time dimension to extract the dynamic response time change rate of each microservice application node and construct the response latency gradient of the data processing microservice. The cache line eviction event sequence and multi-core parallel access conflict frequency of physical host nodes are separated from the telemetry time sequence matrix features of the underlying server microarchitecture in the hardware and software operation status slice features. Combined with the latency penalty coefficient of the micro-instruction retirement cycle, the hardware last-level cache miss rate is constructed. The response latency gradient of the data processing microservice and the missing rate of the final-level hardware cache are used as the initial features of the nodes and input to the graph attention network layer. The information aggregation weights are assigned to the adjacent nodes in the cross-layer digital twin topology graph through the masked self-attention mechanism. The features of the adjacent nodes are weighted and activated according to the information aggregation weights, and the micro-resource competition feature vector is output.

5. The intelligent data center operation and maintenance method based on digital twins according to claim 4, characterized in that: The specific process of calculating the cross-container resource preemption interference index within the physical host machine and marking abnormal nodes based on the micro-resource competition feature vector is as follows: By performing nonlinear mapping on the micro-resource competition feature vector through a multilayer perceptron, a cross-container resource preemption interference index is output, which represents the probability of hardware bus queuing conflicts between multiple data processing microservices within the same physical host. Within a time sliding window, calculate the exponential moving average and standard deviation of the cross-container resource preemption interference index to construct a dynamic adaptive interference threshold; The current cross-container resource preemption interference index is compared with the dynamic adaptive interference threshold. Microservice container entities whose differences exceed the tolerance range multiple times are extracted, and abnormal node location and marking are performed in the cross-layer digital twin topology graph.

6. The intelligent data center operation and maintenance method based on digital twins according to claim 1, characterized in that: For cross-layer digital twin topology graph nodes with abnormal node location markers, and considering the real-time high-concurrency data traffic load characteristics, the specific process of sandbox fault propagation evolution along the call dependency edges of the distributed data processing microservice link is as follows: The total number of concurrent connections and the depth of the underlying request queue in the real-time high-concurrency data traffic load characteristics are mapped to the corresponding node edges of the cross-layer digital twin topology map to establish a dynamic microservice load level model. Extract cross-layer digital twin topology graph nodes with abnormal node location markers as directed graph propagation source points, and analyze the asynchronous non-blocking call protocol and exponential backoff retry network parameters in the call dependency edges of the distributed data processing microservice link tracing sequence. Based on the directed graph propagation source point, the microservice dynamic capacity level model, and the asynchronous non-blocking call protocol and exponential backoff retry network parameters, a node capacity overflow transfer function based on queuing theory is constructed. In the virtual space, the call chain request congestion state transition caused by the exhaustion of underlying hardware resources is simulated iteratively according to the time step, generating a link congestion state evolution sequence in the virtual sandbox environment.

7. The intelligent operation and maintenance method for data centers based on digital twins according to claim 6, characterized in that: The specific process by which the output affects the set of microservice nodes and the cascading failure network topology path is as follows: From the link congestion state evolution sequence, trace the derived failure nodes where the simulated request backlog exceeds the capacity limit of the microservice dynamic load water level model, perform deduplication and aggregation operations on the derived failure nodes, and generate a set of affected microservice nodes. In the cross-layer digital twin topology graph, a connected subgraph is extracted that connects the source point of the directed graph propagation and the set of affected microservice nodes. Normal bypass branches that do not exhibit congestion transfer characteristics in the connected subgraph are removed to generate an abnormal propagation connected subgraph. The abnormal propagation connected subgraph is transformed into a directed acyclic graph containing fault evolution timestamps, generating a cascading failure network topology path.

8. The intelligent data center operation and maintenance method based on digital twins according to claim 1, characterized in that: The specific process of generating dynamic management scripts through multi-objective collaborative game analysis is as follows: Extracting the data processing service level protocol attributes of the affected microservice node set, combining the micro-computing power margin of the backup physical machine cluster and the cross-node migration bandwidth cost of the cascading failed network topology path, and then analyzing these attributes. Analyze the data processing service level protocol attributes corresponding to the affected microservice node set, and extract the tolerance latency limit and business degradation priority weight of the core financial data processing tasks; Collect the available last-level cache capacity and idle network interface controller throughput in the micro-computing power margin of the standby physical machine cluster, and evaluate the cross-node migration bandwidth cost between the source host and the target standby node in the cascaded failure network topology path. A mixed-integer programming scheduling model is constructed with the optimization objectives of minimizing the default rate of the tolerance latency upper limit and minimizing the bandwidth cost of cross-node migration. The inputs are the service degradation priority weight, the available last-level cache capacity, and the throughput of the idle network interface controller. A heuristic constraint solver is used to calculate the optimal solution of resource allocation that satisfies the hard constraint of micro-computing power margin. Dynamic management scripts are generated based on the optimal solution of resource allocation.

9. The intelligent data center operation and maintenance method based on digital twins according to claim 8, characterized in that: The specific process of distributing dynamic management scripts to the physical data center for execution, including cross-region hot migration of target microservices and secure isolation of abnormal containers, is as follows: Parse the abnormal container security isolation instructions in the dynamic management script, call the control group kernel interface of the physical data center host operating system, and implement hard limits on the last-level cache allocation and CPU instruction emission quota compression on the microservice container entities with abnormal node location markers; The target microservice cross-region hot migration instruction in the dynamic management script is parsed, and the memory page incremental synchronization and checkpoint snapshot capture are performed on the high-priority microservice containers in the affected microservice node set. The checkpoint snapshot is then transmitted to the standby physical machine cluster for container state recovery. After the standby physical machine cluster completes container state recovery, the virtual LAN routing table and distributed microservice registration and discovery center of the physical data center are updated to complete the cross-region hot migration of the target microservice.

10. A data center intelligent operation and maintenance system based on digital twins, used to execute the data center intelligent operation and maintenance method based on digital twins as described in any one of claims 1-9, characterized in that, include: The twin mapping module is used to obtain the telemetry time series matrix of the underlying server microarchitecture and the link tracing sequence of the distributed data processing microservice, and map the underlying server microarchitecture telemetry time series matrix and the distributed data processing microservice link tracing sequence to the pre-built cross-layer digital twin topology map to generate hardware and software running status slice features. The anomaly localization module is used to aggregate data from the slice features of software and hardware running status, such as microservice response latency gradient and hardware last-level cache miss rate, using graph attention network to extract micro-resource contention feature vectors. Based on the micro-resource contention feature vectors, it calculates the cross-container resource preemption interference index within the physical host and marks the location of anomaly nodes. The sandbox evolution module is used to perform sandbox fault propagation evolution for cross-layer digital twin topology graph nodes with abnormal node location markers, combined with the real-time high-concurrency data traffic load characteristics, and to trace the call dependency edges along the distributed data processing microservice link sequence to output the set of affected microservice nodes and the cascading failure network topology path. The dynamic management module is used to extract the data processing service level protocol attributes of the affected microservice node set, combine the micro-computing power margin of the backup physical machine cluster and the cross-node migration bandwidth cost of the cascaded failed network topology path, generate dynamic management scripts through multi-objective collaborative game analysis, and distribute the dynamic management scripts to the physical data center to execute the cross-region hot migration of the target microservices and the security isolation of abnormal containers.

Images