Intelligent scheduling method and system for heterogeneous, off-site cloud computing services

Abstract

The application discloses a method and system for intelligent scheduling of heterogeneous and remote cloud computing services, comprising resource abstraction: constructing a unified abstract model of heterogeneous resources to map heterogeneous computing resources into a unified resource description language; step S2: topology modeling: constructing a cross-regional resource pool; step S3: state awareness: collecting real-time state information to construct a state vector; through the unified abstract model of heterogeneous resources and the intelligent adaptive scheduling algorithm, the application can realize efficient utilization of resources, significantly improve resource utilization, and effectively avoid the problems of resource idling and over-allocation in traditional methods; meanwhile, the intelligent adaptive scheduling algorithm can dynamically select the optimal scheduling strategy according to the task characteristics and resource state, significantly reduce the average task completion time, and the cross-regional collaborative mechanism can further reduce the task queuing waiting time by adjusting resource allocation in advance through load prediction, thereby greatly improving the task scheduling efficiency.

Description

Technical Field

[0001] This invention belongs to the field of cloud computing technology and relates to an intelligent scheduling method and system for heterogeneous and geographically dispersed cloud computing services. Background Technology

[0002] With the rapid development of cloud computing technology and the deepening of enterprise digital transformation, cloud computing services have evolved from single data centers to multi-regional, global deployments. Current cloud computing services exhibit the following significant characteristics: widespread application of heterogeneous computing resources (GPUs, FPGAs, ASICs, TPUs, etc.), geographical distribution of multi-regional deployments, and complex and ever-changing business workload requirements.

[0003] However, existing cloud computing scheduling technologies have the following problems: (1) The scheduling strategy is too simple. Traditional scheduling systems often employ static configuration or simple round-robin algorithms, lacking a deep understanding of real-time load and business characteristics. Furthermore, cloud computing data centers generally have low resource utilization rates, with CPU utilization at only 15%-30%, memory utilization at approximately 40%-60%, and GPU utilization even lower, typically only 20%-40%. This low utilization not only leads to resource waste but also increases operating costs.

[0004] (2) Difficulty in coordinating heterogeneous resources In the current cloud computing environment, different types of heterogeneous computing resources (CPU, GPU, FPGA, ASIC, etc.) are typically managed by their own independent resource management systems, lacking a unified resource abstraction model. This results in the inability of these heterogeneous resources to achieve efficient collaboration. Furthermore, there is a lack of effective solutions to problems such as communication overhead and data transmission efficiency between heterogeneous resources.

[0005] (3) Low efficiency of cross-regional scheduling Existing cross-regional scheduling technologies mostly employ simple load balancing strategies, lacking effective predictive models and fault-tolerance mechanisms. Furthermore, data centers in different regions vary in terms of power costs, network bandwidth, and climate conditions, making it difficult for existing scheduling systems to comprehensively consider these factors for global optimization.

[0006] (4) Lack of intelligent decision-making capabilities Existing cloud computing scheduling systems primarily rely on manual rules and experience for scheduling decisions, and updates to scheduling strategies often lag behind changes in business needs. These systems cannot learn from historical scheduling data for optimization, nor can they adapt to complex and ever-changing business scenarios. Summary of the Invention

[0007] To overcome at least one deficiency of existing technologies, this invention provides a unified heterogeneous resource description model for heterogeneous, geographically dispersed cloud computing service scenarios. This model maps resources such as CPUs, GPUs, FPGAs, ASICs, and TPUs to unified resource capability vectors. Simultaneously, it models cross-regional data centers as a weighted graph structure, constructing state vectors based on real-time resource status, task queue status, network status, and historical scheduling data. These vectors are then input into an Actor-Critic-based deep reinforcement learning model, outputting scheduling actions such as task selection, resource placement, task migration, and resource scaling. After the model outputs actions, the system uses an action verification mechanism to determine whether they meet resource capacity, resource compatibility, network latency, and SLA constraints. If verification is successful, scheduling is executed, and the execution results are fed back to an experience replay pool for continuous optimization of the scheduling strategy. When node failures, network anomalies, or data center failures occur, the system triggers task retries, task migration, or data center switching, thus forming a closed-loop scheduling system encompassing heterogeneous resource abstraction, cross-regional scheduling, intelligent decision-making, action verification, execution feedback, and fault tolerance.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent scheduling method for heterogeneous, geographically dispersed cloud computing services, comprising the following steps: Step S1: Resource Abstraction: Construct a unified abstraction model for heterogeneous resources to map heterogeneous computing resources into a unified resource description language; Step S2: Topology Modeling: Constructing a cross-regional resource pool; Step S3: State Awareness: Collect real-time state information to construct a state vector; Step S4: Model Inference: Input the state vector into the deep reinforcement learning model and output the scheduling decision action; Step S5: Action Verification: Verify the feasibility of the actions output by the model to ensure that they do not violate resource constraints and service level agreements; Step S6: Action Execution: Send task allocation instructions or resource scheduling instructions to the target computing node; Step S7: Result Feedback: Record the scheduling execution results, update the environment state, and store the data in the experience replay pool for model training; Step S8: Fault tolerance: Detect whether a fault has occurred. If so, trigger the task migration or data center switching process.

[0009] Furthermore, the method for constructing a unified abstract model for heterogeneous resources is as follows: Step 11: Define the resource metadata structure, including the resource unique identifier, resource type, and computing power metrics; Step 12: Standardize the measurement indicators of heterogeneous computing resources into standardized units of measurement; Step 13: Construct resource capability vectors; Step 14: Construct a four-layer resource abstraction architecture, including a physical resource layer, a virtualization layer, a resource description layer, and a service interface layer; Step 15: Provide heterogeneous resource access interfaces, with a corresponding resource adapter plugin for each resource type.

[0010] Furthermore, step 2 includes Step S21: Construct a cross-regional resource pool topology model, modeling multiple globally distributed data centers as nodes in a graph structure; model the network connections between data centers as weighted edges, with the weight defined as a function of network latency; Weight_ij =α×Norm(Latency_ij)+β×Norm(1 / Bandwidth_ij)+γ×Norm(Cost_ij); Where α, β, and γ are adjustable weight coefficients, Latency_ij represents the network latency from data center i to j, Bandwidth_ij represents the available bandwidth, and Cost_ij represents the data transmission cost. Through a dynamic weight update mechanism, the system can perceive changes in network status in real time. Step S22: Design resource pool grouping strategy; The resource pool adopts a hierarchical grouping strategy: the first layer is the global resource pool, which contains all schedulable heterogeneous computing resources; the second layer is the regional resource pool, which divides the data center into different regions according to the principle of geographical proximity; the third layer is the local resource pool, which consists of all resources within a single data center. When making scheduling decisions, resources are first located at the global level, then load balancing is performed at the regional level, and finally task allocation is performed at the local level.

[0011] Furthermore, the deep reinforcement learning model in step S4 employs an adaptive scheduling algorithm based on deep reinforcement learning, and the model is constructed using the following method: Step S41: Construct a deep reinforcement learning framework; Its state space includes resource state, task queue state, network state and historical scheduling data; its action space includes task selection, resource allocation, task migration and resource scaling actions; and its reward function includes resource utilization efficiency reward, service quality reward, cost reward and fairness reward. Step S42: Training and optimizing the scheduling strategy; Step S43: Design the real-time scheduling decision-making process.

[0012] Furthermore, step S41 includes Step S411: Design the state space of the scheduling system; The state space of a scheduling system is defined as a tuple: S = {S_resource, S_task, S_network, S_history} Here, S_resource represents the current resource status set, including metrics such as CPU utilization, memory usage, GPU utilization, FPGA usage status, and ASIC load for each computing node; S_task represents the status of the queue of tasks to be scheduled, including attributes such as task priority, resource requirement type, expected execution time, and deadline; S_network represents the network status, including parameters such as inter-datacenter latency, bandwidth utilization, and packet loss rate; S_history represents historical scheduling data, including task completion status and resource utilization trends over the past N scheduling cycles; the dimension of the state vector is M × N, where M is the number of resource nodes and N is the feature dimension of each node; Step S412: Design the action space of the scheduling system; The action space is defined as the set of scheduling decisions. A = {a_select, a_place, a_migrate, a_scale} a_select represents selecting the next task to be scheduled from the task queue, using a priority sorting strategy; a_place represents the probability distribution of candidate compute nodes and resource types output to the policy network, with heuristic rules used for action filtering or verification; a_migrate represents making migration decisions for running tasks to achieve load balancing; a_scale represents making dynamic scaling decisions for virtual resources; the action space adopts a discretized design, with each action dimension containing multiple optional values; Step S413: Design the reward function R; The reward function R is expressed as R = w1 × R_efficiency + w2 × R_quality + w3 × R_cost + w4 × R_fairness Wherein, R_efficiency represents the resource utilization efficiency reward, defined as the negative deviation between the actual resource utilization rate and the target utilization rate; R_quality represents the service quality reward, defined as the comparison between the task completion time and the expected completion time; R_cost represents the cost reward, defined as the comparison between the computational resource usage cost and the budget; R_fairness represents the fairness reward, defined as the degree of balance in the waiting time of each task; the weight coefficients w1, w2, w3, and w4 are obtained through offline training optimization; Step S414: Construct a deep neural network architecture; Deep reinforcement learning adopts the Actor-Critic architecture, in which the Actor network uses an attention-enhanced feature extractor, including an input layer that receives the state vector; the first layer is a fully connected layer with 1024 nodes and uses the ReLU activation function; the second layer is a self-attention layer with 8 attention heads; the third layer is a fully connected layer with 512 nodes; and the output layer is the policy output, which generates the action probability distribution.

[0013] Furthermore, step S42 includes Step S421: Collect training data; The training data acquisition module is deployed in each data center, and the data collected includes task submission logs, resource usage logs, and scheduling decision logs; data acquisition is conducted asynchronously. Step S422: Design an experience playback mechanism; A priority experience replay mechanism is designed to store training samples; each experience sample contains a quintuple; the priority calculation formula is: P(i) = |δ_i|^α + ε Where δ_i is the timing difference error, α is the priority exponent (usually taken as 0.6), and ε is the minimum priority constant; a hierarchical sampling strategy is adopted during sampling. Step S423: Design a distributed training architecture; The design parameter server architecture is used for distributed training, and the training adopts an asynchronous update strategy. The model parameters are trained using mixed precision.

[0014] Intelligent scheduling system for heterogeneous, geographically dispersed cloud computing services, including Heterogeneous Resource Abstraction Layer: Used to build a unified abstraction model for heterogeneous resources, mapping heterogeneous computing resources to a unified resource description language; Intelligent scheduling engine layer: Collects real-time status information and outputs scheduling decisions based on an adaptive scheduling algorithm using deep reinforcement learning; Cross-regional collaboration layer: used to coordinate resource scheduling between multiple regional data centers; Application Interface Layer: Provides application interfaces to support business systems in submitting scheduling requests conveniently, and provides a visual monitoring interface to facilitate operations and maintenance personnel in monitoring the system's operating status in real time.

[0015] Furthermore, the heterogeneous resource unified abstraction model adopts a hierarchical structure design, including a physical resource layer, a virtualization abstraction layer, a resource description layer, and a service interface layer. Physical Resource Layer: Defines the physical resources of various heterogeneous computing resources and collects real-time status data of the underlying hardware; The virtualization abstraction layer maps physical resources to virtual computing units, and uses container technology and virtualization technology to achieve resource isolation and sharing; each virtual computing unit includes the number of computing cores, memory capacity, and storage bandwidth attributes; this layer introduces a resource compatibility matrix to describe the resource requirements of different types of tasks and the degree of matching of various resources; The resource description layer converts the output of the virtualization abstraction layer into a unified resource description language format; the logical attributes of the resource description language include resource type identifier, capability level, current load status, and availability metrics. The service interface layer provides two interface forms: RESTful API and gRPC, for the upper-layer scheduling system to call.

[0016] Furthermore, the intelligent scheduling engine layer includes a real-time status perception module, a strategy decision module, and an execution feedback module. The real-time status perception module is responsible for collecting the resource usage status of data centers in various regions; the strategy decision module outputs scheduling decisions based on a deep reinforcement learning network; and the execution feedback module feeds the scheduling results back to the training module to achieve continuous optimization of the strategy.

[0017] Furthermore, the adaptive scheduling algorithm adopts an Actor-Critic architecture, which includes a state representation network component, a policy network component, and a value network component. The state representation network component transforms the system state into a feature vector that the algorithm can process; the state representation network uses a multilayer perceptron and attention mechanism to extract features and capture key information in the state space. The policy network component outputs scheduling decisions. The policy network uses a Softmax output layer to give the output scheduling decisions probabilistic distribution characteristics. The value network component evaluates the merits of the current scheduling strategy. The inputs to the value network are the outputs of the state representation network and the policy network, and the output is the value estimate of the state-action pair.

[0018] In summary, the advantages of this invention are: This invention achieves efficient resource utilization and significantly improves resource efficiency through a unified abstract model of heterogeneous resources and an intelligent adaptive scheduling algorithm, effectively avoiding resource idleness and over-allocation problems in traditional methods. Simultaneously, the intelligent adaptive scheduling algorithm dynamically selects the optimal scheduling strategy based on task characteristics and resource status, significantly reducing the average task completion time. The cross-regional collaboration mechanism adjusts resource allocation in advance through load prediction, further reducing task queuing time and greatly improving task scheduling efficiency. The intelligent adaptive scheduling algorithm optimizes resource allocation and reduces the operating costs of cloud computing services.

[0019] The three-level fault tolerance mechanism of this invention can effectively cope with various failure scenarios and ensure service continuity. When a single node fails, the task is automatically retried or migrated; when a data center-level failure occurs, traffic is automatically switched to a backup center. The introduction of the fault tolerance mechanism reduces the failure rate of task execution in cross-regional scheduling, significantly improves service quality, and greatly enhances the reliability of cross-regional services.

[0020] The task distribution controller of this invention comprehensively considers energy efficiency ratio, guiding tasks to be preferentially allocated to computing nodes with high energy efficiency ratio. This not only reduces data center energy consumption but also reduces carbon emissions, aligning with the development trend of green cloud computing.

[0021] The hierarchical architecture and standardized interface design proposed in this invention give the system excellent scalability. When new heterogeneous resource types need to be supported, only the corresponding description template needs to be added to the resource abstraction layer, without modifying the upper-layer scheduling logic. This greatly reduces the system's maintenance costs and the difficulty of secondary development, and improves the system's scalability and maintainability. Attached Figure Description

[0022] Figure 1 This is a flowchart of the intelligent scheduling method of the present invention.

[0023] Figure 2 This is a diagram of the intelligent scheduling system architecture of the present invention.

[0024] Figure 3 This is a comparison chart of resource utilization rates for different scheduling algorithms of this invention.

[0025] Figure 4 This is a comparison chart of the execution times of different scheduling algorithms of the present invention.

[0026] Figure 5 This is a cost comparison chart of different scheduling algorithms of the present invention.

[0027] Figure 6 This is a network architecture diagram of the Deep Reinforcement Learning-based Adaptive Scheduling Algorithm (DRL-ASA) of the present invention.

[0028] Figure 7 This is a diagram illustrating the implementation architecture of the cross-regional three-level fault tolerance mechanism of the present invention. Detailed Implementation

[0029] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0030] Example: like Figure 1 As shown, an intelligent scheduling method for heterogeneous, geographically dispersed cloud computing services includes the following steps: Step S1: Resource Abstraction: Construct a unified abstraction model for heterogeneous resources to map heterogeneous computing resources into a unified resource description language; The method for constructing a unified abstract model for heterogeneous resources is as follows: Step 11: Define the resource metadata structure, including the resource unique identifier, resource type, and computing power metrics; Step 12: Unify the measurement indicators of heterogeneous computing resources into standardized units of measurement, with CPU measured in MIPS, GPU in GFLOPS, FPGA in the product of logic gate count and clock frequency, and ASIC in the execution efficiency of dedicated instruction set. Step 13: Construct a resource capability vector, which includes standardized values ​​for six dimensions: computing power, memory bandwidth, network communication capability, storage I / O capability, power efficiency, and task response latency. Step 14: Construct a four-layer resource abstraction architecture, including a physical resource layer, a virtualization layer, a resource description layer, and a service interface layer; Step 15: Provide heterogeneous resource access interfaces, with a corresponding resource adapter plugin for each resource type.

[0031] The resource capability vectors are corrected for nonlinearity using a Max-Min linear normalization method combined with a Sigmoid function to ensure that the attribute values ​​of different types of resources are comparable.

[0032] Step S2: Topology Modeling: Constructing a cross-regional resource pool; Step S21: Construct a cross-regional resource pool topology model by modeling multiple globally distributed data centers as nodes in a graph structure; Each data center node contains the following attributes: data center identifier (DC_ID), geographic location coordinates (Latitude, Longitude), network topology location (Network_Level), resource capacity vector (Capacity_Vector), and current load state (Load_State).

[0033] Network connections between data centers are modeled as weighted edges, with the weights defined as a function of network latency. Weight_ij =α×Norm(Latency_ij)+β×Norm(1 / Bandwidth_ij)+γ×Norm(Cost_ij); Where α, β, and γ are adjustable weight coefficients, Latency_ij represents the network latency from data center i to j, Bandwidth_ij represents the available bandwidth, and Cost_ij represents the data transmission cost. Through a dynamic weight update mechanism, the system can perceive changes in network status in real time. Step S22: Design resource pool grouping strategy; The resource pool employs a hierarchical grouping strategy: the first layer is the global resource pool, containing all schedulable heterogeneous computing resources; the second layer is the regional resource pool, dividing data centers into regions such as Asia Pacific, Europe, North America, and South America based on geographical proximity; and the third layer is the local resource pool, composed of all resources within a single data center. Scheduling decisions are made by first locating resources at the global level, then performing load balancing at the regional level, and finally allocating tasks at the local level.

[0034] Step S3: State Awareness: Collect real-time state information to construct a state vector; Status information includes resource status, task queue status, network status, etc. Step S4: Model Inference: Input the state vector into the deep reinforcement learning model and output the scheduling decision action; The deep reinforcement learning model in step S4 adopts an adaptive scheduling algorithm based on deep reinforcement learning, and the model is constructed as follows; Step S41: Construct a deep reinforcement learning framework; its state space includes resource state, task queue state, network state and historical scheduling data, its action space includes task selection, resource allocation, task migration and resource scaling actions, and its reward function includes resource utilization efficiency reward, service quality reward, cost reward and fairness reward.

[0035] Step S411: Design the state space of the scheduling system; The state space of a scheduling system is defined as a tuple: S = {S_resource, S_task, S_network, S_history} Here, S_resource represents the current resource status set, including metrics such as CPU utilization, memory usage, GPU utilization, FPGA usage status, and ASIC load for each computing node; S_task represents the status of the queue of tasks to be scheduled, including attributes such as task priority, resource requirement type, expected execution time, and deadline; S_network represents the network status, including parameters such as inter-data center latency, bandwidth utilization, and packet loss rate; and S_history represents historical scheduling data, including task completion status and resource utilization trends over the past N scheduling cycles.

[0036] The state space design includes resource state, task queue state, network state, and historical scheduling data. The state vector has a dimension of M × N, where M is the number of resource nodes and N is the feature dimension of each node. For a medium-sized system with 1000 computing nodes, the state vector dimension is approximately 50,000.

[0037] Step S412: Design the action space of the scheduling system; Action space is defined as the set of scheduling decisions. Action space design includes task selection actions, resource allocation actions, task migration actions, and resource scaling actions. A = {a_select, a_place, a_migrate, a_scale} `a_select` selects the next task to be scheduled from the task queue, using a priority sorting strategy; `a_place` represents the probability distribution of candidate compute nodes and resource types output by the policy network, with heuristic rules used for action filtering or verification; `a_migrate` makes migration decisions for running tasks to achieve load balancing; and `a_scale` makes dynamic scaling decisions for virtual resources. The action space is discretized, with each action dimension containing multiple selectable values.

[0038] Step S413: Design the reward function R; The reward function design adopts a multi-objective composite reward function, which includes resource utilization efficiency reward, service quality reward, cost reward and fairness reward; The reward function R is expressed as R = w1 × R_efficiency + w2 × R_quality + w3 × R_cost + w4 × R_fairness Here, R_efficiency represents the resource utilization efficiency reward, defined as the negative deviation between the actual resource utilization rate and the target utilization rate; R_quality represents the service quality reward, defined as the comparison between the task completion time and the expected completion time; R_cost represents the cost reward, defined as the comparison between the computational resource usage cost and the budget; and R_fairness represents the fairness reward, defined as the degree of balance in waiting times among tasks. The weight coefficients w1, w2, w3, and w4 were obtained through offline training optimization.

[0039] Step S414: Construct a deep neural network architecture; Deep reinforcement learning employs the Actor-Critic architecture. The Actor network uses an attention mechanism-enhanced feature extractor, specifically comprising: an input layer receiving the state vector; a first fully connected layer with 1024 nodes and ReLU activation; a second self-attention layer containing 8 attention heads to capture correlations between different resource nodes; a third fully connected layer with 512 nodes; and an output layer that outputs the policy, generating the action probability distribution. The Critic network structure is similar to the Actor network, outputting a value estimate of the state-action pair.

[0040] Step S42: Training and optimizing the scheduling strategy; Step S421: Collect training data; The training data acquisition module is deployed in each data center, collecting data including: task submission logs (containing task ID, submission time, resource requirements, execution duration, etc.); resource usage logs (containing the CPU, GPU, memory, network, and other resource usage of each computing node); and scheduling decision logs (containing the decision input, decision output, and execution results for each scheduling cycle). Data acquisition is asynchronous, with a 1-second interval, and the daily data volume collected by a single data center is approximately 10GB.

[0041] Step S422: Design an experience playback mechanism; The Prioritized Experience Replay (PER) mechanism is designed to store training samples. Each experience sample contains a 5-tuple (state, action, reward, next_state, done). The experience replay buffer size is 1 million samples, and the priority calculation formula is as follows: P(i) = |δ_i|^α + ε Where δ_i is the timing difference error, α is the priority exponent (usually taken as 0.6), and ε is the minimum priority constant. A tiered sampling strategy is used during sampling to ensure that high-priority samples are selected first.

[0042] Step S423: Design a distributed training architecture; The design employs a parameter server architecture for distributed training, with asynchronous update strategies and mixed-precision training for model parameters.

[0043] The training cluster consists of multiple worker nodes and several parameter server nodes. Each worker node is responsible for calculating gradients and uploading them to the parameter server; the parameter server is responsible for aggregating gradients and updating model parameters. Training employs an asynchronous update strategy, with a gradient synchronization period of 100ms. Model parameters are trained using mixed precision training, converting some parameters to FP16 format to accelerate the training process.

[0044] Step S43: Design the real-time scheduling decision-making process; Step S431: Design a scheduling cycle triggering mechanism; The scheduling decision-making process employs a combination of event-driven and timed triggering mechanisms. The timed triggering cycle is 10 seconds. Event-driven triggering conditions include: new task submission events, task completion events, resource failure events, and load alarm events. The 10-second timed triggering cycle ensures that the system can respond promptly to load changes.

[0045] Step S432: Design the scheduling decision execution process; The scheduling decision execution process includes the following steps: The first step is state awareness, which involves collecting information such as the current resource state, task queue state, and network state to construct a state vector; The second step is model inference, which involves inputting the state vector into the trained deep reinforcement learning model and outputting scheduling decision actions. The third step is action verification, which verifies the feasibility of the actions output by the model to ensure that they do not violate resource constraints and service level agreements (SLAs). The fourth step is action execution, which involves sending task allocation instructions or resource scheduling instructions to the target computing node; The fifth step is to provide feedback on the results, record the scheduling execution results, and update the environment status.

[0046] Step S5: Action Verification: Verify the feasibility of the actions output by the model to ensure that they do not violate resource constraints and service level agreements; Step S6: Action Execution: Send task allocation instructions or resource scheduling instructions to the target computing node; Step S7: Result Feedback: Record the scheduling execution results, update the environment state, and store the data in the experience replay pool for model training; Step S8: Fault tolerance: Detect whether a fault has occurred. If so, trigger the task migration or data center switching process.

[0047] Step S8 performs fault tolerance processing based on cross-regional fault tolerance and load balancing mechanisms; Cross-regional fault tolerance and load balancing mechanisms include Step S81: Design a fault tolerance mechanism; Step S811: Design a multi-level fault detection mechanism; Multi-level fault detection mechanisms include: The first layer is heartbeat detection. Each computing node sends a heartbeat packet to the monitoring center every 5 seconds. If no heartbeat is received for 3 consecutive times, the node is judged to be faulty. The second layer is resource anomaly detection, which identifies potential faults by analyzing indicators such as sudden changes in resource utilization and increased task failure rate. The third layer is network anomaly detection, which identifies network problems by monitoring indicators such as changes in network latency and increases in packet loss rate.

[0048] A fault detection module is designed based on a multi-level fault detection mechanism. The fault detection module is deployed in each data center and adopts a primary and backup dual-machine hot standby architecture.

[0049] Step S812: Design a task migration mechanism; When a compute node failure is detected, the system automatically triggers the task migration process. The task migration mechanism includes: for stateless compute tasks, they are directly re-executed on the new compute node; for stateful compute tasks, the execution state is restored based on the task checkpoint record; for tasks with high real-time requirements, they are preferentially migrated to a backup node within the same data center; for tasks that can tolerate latency, they are preferentially migrated to other data centers with lower load.

[0050] Step S813: Construct a data consistency model to control data consistency; Cross-regional data transmission employs a data consistency model. Measures to control data consistency under this model include: using a distributed transaction protocol (TCC mode) to ensure the atomicity of cross-regional task calls; using an asynchronous replication mechanism to ensure the consistency of data replicas, with a replication latency target of less than 100ms; and using a conflict detection and resolution mechanism to handle concurrent update conflicts.

[0051] Step S82: Design a load balancing mechanism; Step S821: Deploy a global load balancer; A global load balancer is deployed at the network ingress layer, routing requests to the nearest data center based on the user's geographical location. When a data center becomes overloaded, the global load balancer automatically redirects traffic to other data centers. Global load balancing uses a weighted round-robin algorithm, with weights dynamically adjusted based on the real-time load rate of the data centers.

[0052] Step S822: Deploy a local load balancer; Within a single data center, a local load balancer distributes tasks to appropriate compute nodes. The local load balancer employs a heuristic algorithm based on resource matching: first, it calculates the matching degree between the task's resource requirements and the resource capabilities of each compute node; then, it selects the node with the highest matching degree as the target node; when multiple nodes have similar matching degrees, it selects the node with the lowest current load.

[0053] like Figure 2 Figure 7 shows an intelligent scheduling system for heterogeneous, geographically dispersed cloud computing services, including... Heterogeneous Resource Abstraction Layer: Used to build a unified abstraction model for heterogeneous resources, mapping heterogeneous computing resources to a unified resource description language; Intelligent scheduling engine layer: Collects real-time status information and outputs scheduling decisions based on an adaptive scheduling algorithm using deep reinforcement learning; Cross-regional collaboration layer: used to coordinate resource scheduling between multiple regional data centers; Application Interface Layer: Provides a unified API interface, enabling business systems to easily submit scheduling requests, and provides a visual monitoring interface to facilitate operations and maintenance personnel in monitoring the system's operating status in real time.

[0054] The heterogeneous resource abstraction layer, as the system's underlying infrastructure, includes resource abstraction middleware and a heterogeneous resource description language module. It constructs a unified heterogeneous resource description language (UHRDL) model to uniformly abstract heterogeneous computing resources such as CPUs, GPUs, FPGAs, ASICs, and TPUs, outputting a standardized resource description language. Through the heterogeneous resource abstraction layer, various heterogeneous resources are mapped to unified resource identifiers, providing a consistent resource view for upper-layer scheduling.

[0055] The Heterogeneous Resource Unified Abstraction Model maps different types of heterogeneous computing resources such as CPU, GPU, FPGA, ASIC, and TPU to a unified resource description language. The Heterogeneous Resource Unified Abstraction Model adopts a hierarchical structure design, including a physical resource layer, a virtualization abstraction layer, a resource description layer, and a service interface layer.

[0056] Physical Resource Layer: Defines the physical resources of various heterogeneous computing resources and collects real-time status data of the underlying hardware. For CPU resources, it defines parameters such as the number of cores, frequency, and cache hierarchy; for GPU resources, it defines parameters such as the number of stream processors, video memory capacity, and bandwidth; for FPGA resources, it defines parameters such as the number of logic units, the number of DSP blocks, and memory resources; for ASIC resources, it defines parameters such as dedicated computing capabilities and power consumption characteristics. Real-time status data includes, but is not limited to, CPU utilization, GPU video memory usage, FPGA configuration status, and ASIC operating mode.

[0057] The virtualization abstraction layer maps physical resources to virtual computing units, using container and virtualization technologies to achieve resource isolation and sharing. Each virtual computing unit includes attributes such as the number of computing cores, memory capacity, and storage bandwidth. This layer introduces a resource compatibility matrix to describe the resource requirements of different types of tasks and the degree of matching between various resources.

[0058] The resource description layer converts the output of the virtualization abstraction layer into a unified resource description language format. The logical attributes of the resource description language include resource type identifier, capability level, current load status, availability metrics, etc. The UHRDL language is defined in JSON format, which facilitates interaction and parsing between systems.

[0059] The service interface layer provides two interface forms: RESTful API and gRPC, for the upper-layer scheduling system to call.

[0060] The resource description language defines the resource metadata structure, which includes a unique resource identifier (Resource_ID), resource type (Resource_Type), computing capacity index (Compute_Capacity), and resource capability vector (RCV).

[0061] Resource Unique Identifier: Adopts 128-bit UUID format to ensure uniqueness in global cross-regional scheduling scenarios; Resource types: defined using an enumeration type, including six basic types such as CPU, GPU, FPGA, ASIC, TPU, and DSP, as well as custom extended types; Computing power metrics are represented using standardized floating-point numbers and are defined as follows: For CPUs, the standardized unit is MIPS (Million Instructions Per Second); for GPUs, the standardized unit is GFLOPS (Floating-point Operations Per Second); for FPGAs, the standardized unit is the product of the number of logic gates and the clock frequency; and for ASICs, the standardized unit is the execution efficiency of the dedicated instruction set and the resource metadata structure. Resource capability vectors enable unified quantitative comparison of heterogeneous resources, including standardized values ​​for computing power, memory bandwidth, network communication capability, storage I / O capability, power efficiency, and task response latency. The mathematical expression of resource capability vectors is as follows: RCV = [C_com, C_mem, C_net, C_storage, P_power, T_latency] Wherein, C_com represents the standardized value of computing power, C_mem represents the standardized value of memory bandwidth, C_net represents the standardized value of network communication capability, C_storage represents the standardized value of storage I / O capability, P_power represents the standardized value of power efficiency, and T_latency represents the standardized value of task response latency. The value of each dimension is mapped to the [0,1] interval through a normalization algorithm. The normalization algorithm uses the Max-Min linear normalization method combined with the Sigmoid function for non-linear correction to ensure that the attribute values ​​of different types of resources are comparable.

[0062] The intelligent scheduling engine layer is the core decision-making unit of the system, integrating a deep reinforcement learning-based adaptive scheduling algorithm (DRL-ASA). The intelligent scheduling engine layer comprises three key modules: a real-time state awareness module, a policy decision module, and an execution feedback module. The real-time state awareness module is responsible for collecting resource usage status data from various regional data centers; the policy decision module outputs scheduling decisions based on the deep reinforcement learning network; and the execution feedback module feeds the scheduling results back to the training module, enabling continuous policy optimization.

[0063] The Adaptive Scheduling Algorithm (DRL-ASA) adopts an Actor-Critic architecture, which includes three core components: a state representation network, a policy network, and a value network.

[0064] The state representation network component transforms the system state into a feature vector that the algorithm can process. Input states include: CPU utilization, memory utilization, network bandwidth utilization, utilization of heterogeneous resources such as GPUs / FPGAs / ASICs, historical task completion times, and current queue lengths for each data center. The state representation network uses a multilayer perceptron (MLP) and attention mechanism to extract features, effectively capturing key information in the state space.

[0065] The policy network (Actor network) component is responsible for outputting scheduling decisions. The output of the policy network is a task-resource matching vector for each task to be scheduled, specifying the specific computing node and resource type the task should be allocated to. The policy network uses a Softmax output layer to ensure that the output scheduling decisions have probabilistic distribution characteristics.

[0066] The value network (Critic network) component is responsible for evaluating the merits of the current scheduling policy. The value network takes as input the outputs of the state representation network and the policy network, and outputs a value estimate of the state-action pair. The presence of the value network makes the training of the policy network more stable.

[0067] The adaptive scheduling algorithm employs an experience replay mechanism, a target network mechanism, and a reward function for training. The experience replay pool stores historical scheduling experience, and gradient updates are performed by randomly sampling from the replay pool during training. The target network periodically synchronizes parameters from the main network to ensure the stability of the training process. The reward function comprehensively considers multiple objectives, including task completion time, resource utilization, energy efficiency, and load balancing.

[0068] The cross-regional collaboration layer coordinates resource scheduling across multiple regional data centers. This layer deploys a global load prediction module, a task distribution controller, and a fault tolerance management module to predict business load trends in each region and automatically trigger task migration in the event of a failure.

[0069] The global load forecasting module employs a hybrid forecasting model combining Long Short-Term Memory (LSTM) networks and Transformer. LSTM captures long-term dependencies in time series data, while Transformer extracts spatial correlation features between different regions. The global load forecasting module outputs predicted load values ​​for each region within a future time window, providing data support for scheduling decisions.

[0070] Based on the prediction results and real-time status of the global load prediction module, the task distribution controller uses a mixed-integer programming algorithm to solve for the optimal task allocation scheme. The optimization objectives of the task distribution controller include: minimizing the average task completion time, maximizing resource utilization, maximizing energy efficiency, and minimizing cross-regional communication overhead. Constraints include: resource capacity constraints, latency constraints, and fault tolerance constraints.

[0071] The fault tolerance management module has a three-level fault tolerance mechanism. The first level is a task-level retry mechanism, which automatically reschedules tasks when they fail. The second level is a virtual machine / container-level migration mechanism, which automatically migrates tasks to healthy nodes when a node fails. The third level is a data center-level failover mechanism, which switches traffic to a backup data center when the entire data center is down.

[0072] The application interface layer includes user interfaces and heterogeneous resource access interfaces: The user interface offers two options: a web management console and a RESTful API. The web console, developed using the Vue.js framework, provides functionalities such as resource monitoring, task management, and policy configuration. The RESTful API conforms to the OpenAPI 3.0 specification and supports integration with third-party systems.

[0073] Heterogeneous resource access adopts a plug-in design, with one resource adapter plug-in corresponding to each resource type. The heterogeneous resource access interface definitions include: resource registration interface (register_resource), resource deregistration interface (unregister_resource), status reporting interface (report_status), task execution interface (execute_task), and task cancellation interface (cancel_task).

[0074] The scheduling system also includes a distributed collaborative computing architecture, which supports a heterogeneous resource abstraction layer, an intelligent scheduling engine layer, a cross-regional collaboration layer, and an application interface layer. Specifically, it includes... Scheduling control nodes: Employ a high-performance server cluster with a deployed intelligent scheduling engine. The scheduling control nodes use a multi-replica deployment approach, ensuring consistency in scheduling decisions through the Raft consensus protocol. Each control node is configured with a high-performance CPU (such as an Intel Xeon Gold series) and a large capacity of memory (256GB or more) to ensure the real-time requirements of the scheduling algorithm.

[0075] Resource awareness nodes: Deployed in data centers across various regions, these nodes are responsible for collecting local resource status data and reporting it to the scheduling and control nodes. They employ an edge computing architecture, performing data preprocessing locally to reduce network transmission overhead.

[0076] High-speed interconnection network: Utilizing InfiniBand or 100Gbps Ethernet to connect data centers, a low-latency, high-bandwidth cross-regional network is constructed. The network supports the RDMA protocol, enabling direct cross-regional memory access and further reducing task migration latency.

[0077] Intelligent Accelerator Cards: Provide hardware acceleration for deep reinforcement learning scheduling algorithms. Utilizing GPUs, NPUs, TPUs, FPGAs, or other AI accelerator cards, they accelerate the inference and training processes of policy and value networks.

[0078] The scheduling system also includes hardware architecture and software system architecture. The hardware architecture includes scheduling control plane hardware, computing resource plane hardware, hardware network architecture, software system architecture, and integration interfaces. The scheduling and control plane hardware is deployed in a three-node cluster, with each node configured as follows: CPU: Intel Xeon Gold 6248R (24 cores, 48 ​​threads) × 2; Memory: DDR4 2933MHz 64GB × 8 (total 512GB); Storage: NVMe SSD 2TB × 4 (RAID10); Network: 100Gbps InfiniBand network cards × 2; Operating system: Ubuntu 20.04 LTS. The cluster uses the Raft consensus algorithm to ensure high availability of the control plane.

[0079] The computing resource plane hardware includes heterogeneous computing nodes, with typical configurations as follows: CPU computing nodes use Intel XeonGold 6248R + 256GB memory + 10TB NVMe SSD; GPU computing nodes use NVIDIA A100 40GB × 8 + 512GB memory + 10TB NVMe SSD; FPGA computing nodes use Xilinx Alveo U250 × 2 + 256GB memory + 10TB NVMe SSD; and ASIC computing nodes use custom accelerator cards + 128GB memory + 10TB NVMe SSD.

[0080] The hardware network architecture adopts a two-layer Spine-Leaf network architecture: the Leaf layer switches use 100Gbps ports, and each compute node is connected to the Leaf switch via a 100Gbps link; the Spine layer switches use 400Gbps ports, and the Leaf switches are connected to the Spine switches via 400Gbps links. Cross-data center networks are connected via leased lines or VPN tunnels, with bandwidth flexibly scalable according to business needs.

[0081] The software system architecture includes a microservice architecture, a message queue system, and a deployment module; The microservice architecture includes a resource management service module (Resource Manager), a task scheduling service module (Task Scheduler), a policy management service module (Policy Manager), a monitoring and alerting service module (Monitor & Alert), and a data storage service module (Data Store). Resource Manager module: Used for registering, monitoring, and deregistering heterogeneous computing resources; Task Scheduler module: Used for generating and executing scheduling decisions; Policy Manager: Used for configuring and updating scheduling policies; The Monitoring & Alert service module is used for system status monitoring and anomaly alerts. Data Storage Service Module (Data Store): Used for persistent storage of scheduling data, log data, and model parameters.

[0082] The message queue system uses a Kafka cluster as the message middleware.

[0083] The message queue system defines four themes: resource status, scheduling tasks, scheduling decisions, and alarm notifications.

[0084] The message queue system includes the following topics: resource status (resource_status), used to deliver resource status change messages; scheduling task (scheduling_task), used to deliver tasks to be scheduled; scheduling decision (scheduling_decision), used to deliver scheduling decision messages; and alert notification (alert_notification), used to deliver alert notification messages. The Kafka cluster uses a multi-replica deployment to ensure high message reliability.

[0085] The deployment module uses Docker, and all service components are deployed using Docker containers. Container orchestration uses Kubernetes. The Kubernetes cluster adopts a master-slave architecture: the master node is responsible for core functions such as cluster management, API services, and the scheduler; the node nodes are responsible for running Pod instances. The system uses Helm Chart for application deployment management, supporting rolling upgrades and rollbacks.

[0086] Based on the intelligent scheduling system of this application, comparative verification experiments were conducted. By comparing it with random schedulers and round-robin schedulers, the overall performance of the intelligent scheduling system of this application was comprehensively evaluated, and the results are as follows: Figures 3-5 As shown.

[0087] Figure 3 The chart is a bar graph comparing different scheduling algorithms. The horizontal axis represents different scheduling algorithms, and the vertical axis represents resource utilization (%). In the graph, black bars represent the experimental results of the DRL scheduler proposed in this invention, white bars represent the round-robin scheduler, and gray diagonal bars represent the experimental results of the random scheduler. The graph clearly shows that the resource utilization of the DRL scheduler is 94.7%, the round-robin scheduler is 71.3%, and the random scheduler is 62.5%. This invention represents a 23.4% improvement over the round-robin scheduler and a 32.2% improvement over the random scheduler.

[0088] Figure 4 The chart is a bar graph comparing different scheduling algorithms. The horizontal axis represents different scheduling algorithms, and the vertical axis represents the average execution time (seconds). The average execution time of the DRL scheduler is 2.85 seconds, the round-robin scheduler is 4.21 seconds, and the random scheduler is 5.67 seconds. This invention reduces execution time by 32.31% compared to the round-robin scheduler and by 49.73% compared to the random scheduler.

[0089] Figure 5 The chart is a bar graph comparing different scheduling algorithms. The horizontal axis represents different scheduling algorithms, and the vertical axis represents the total cost (in US dollars). The total cost of the DRL scheduler is $285.42, the round-robin scheduler is $412.58, and the random scheduler is $523.17. This invention reduces the cost by 30.82% compared to the round-robin scheduler and by 45.44% compared to the random scheduler.

[0090] Error bars are marked on each experimental graph, representing the standard deviation of multiple experiments to ensure the reliability and repeatability of the results. Experimental conditions: The test environment included data centers in 6 regions, 528 heterogeneous computing resources, and 500 test tasks.

[0091] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. An intelligent scheduling method for heterogeneous, geographically dispersed cloud computing services, characterized by: Includes the following steps: Step S1: Resource Abstraction: Construct a unified abstraction model for heterogeneous resources to map heterogeneous computing resources into a unified resource description language; Step S2: Topology Modeling: Constructing a cross-regional resource pool; Step S3: State Awareness: Collect real-time state information to construct a state vector; Step S4: Model Inference: Input the state vector into the deep reinforcement learning model and output the scheduling decision action; Step S5: Action Verification: Verify the feasibility of the actions output by the model to ensure that they do not violate resource constraints and service level agreements; Step S6: Action Execution: Send task allocation instructions or resource scheduling instructions to the target computing node; Step S7: Result Feedback: Record the scheduling execution results, update the environment state, and store the data in the experience replay pool for model training; Step S8: Fault tolerance: Detect whether a fault has occurred. If so, trigger the task migration or data center switching process.

2. The intelligent scheduling method for heterogeneous, geographically dispersed cloud computing services according to claim 1, characterized in that: The method for constructing a unified abstract model for heterogeneous resources is as follows: Step 11: Define the resource metadata structure, including the resource unique identifier, resource type, and computing power metrics; Step 12: Standardize the measurement indicators of heterogeneous computing resources into standardized units of measurement; Step 13: Construct resource capability vectors; Step 14: Construct a four-layer resource abstraction architecture, including a physical resource layer, a virtualization layer, a resource description layer, and a service interface layer; Step 15: Provide heterogeneous resource access interfaces, with a corresponding resource adapter plugin for each resource type.

3. The intelligent scheduling method for heterogeneous, geographically dispersed cloud computing services according to claim 1, characterized in that: Step 2 includes Step S21: Construct a cross-regional resource pool topology model, modeling multiple globally distributed data centers as nodes in a graph structure; model the network connections between data centers as weighted edges, with the weight defined as a function of network latency; Weight_ij =α×Norm(Latency_ij)+β×Norm(1 / Bandwidth_ij)+γ×Norm(Cost_ij); Where α, β, and γ are adjustable weight coefficients, Latency_ij represents the network latency from data center i to j, Bandwidth_ij represents the available bandwidth, and Cost_ij represents the data transmission cost. Through the dynamic weight update mechanism, the system can perceive changes in network status in real time. Step S22: Design resource pool grouping strategy; The resource pool adopts a hierarchical grouping strategy: the first layer is the global resource pool, which contains all schedulable heterogeneous computing resources; the second layer is the regional resource pool, which divides the data center into different regions according to the principle of geographical proximity; the third layer is the local resource pool, which consists of all resources within a single data center. When making scheduling decisions, resources are first located at the global level, then load balancing is performed at the regional level, and finally task allocation is performed at the local level.

4. The intelligent scheduling method for heterogeneous, geographically dispersed cloud computing services according to claim 1, characterized in that: The deep reinforcement learning model in step S4 employs an adaptive scheduling algorithm based on deep reinforcement learning, and the model is constructed as follows: Step S41: Construct a deep reinforcement learning framework; Its state space includes resource state, task queue state, network state and historical scheduling data; its action space includes task selection, resource allocation, task migration and resource scaling actions; and its reward function includes resource utilization efficiency reward, service quality reward, cost reward and fairness reward. Step S42: Training and optimizing the scheduling strategy; Step S43: Design the real-time scheduling decision-making process.

5. The intelligent scheduling method for heterogeneous, geographically dispersed cloud computing services according to claim 4, characterized in that: Step S41 includes Step S411: Design the state space of the scheduling system; The state space of a scheduling system is defined as a tuple: S = {S_resource, S_task, S_network, S_history} Here, S_resource represents the current resource status set, including metrics such as CPU utilization, memory usage, GPU utilization, FPGA usage status, and ASIC load for each computing node; S_task represents the status of the queue of tasks to be scheduled, including attributes such as task priority, resource requirement type, expected execution time, and deadline; S_network represents the network status, including parameters such as inter-datacenter latency, bandwidth utilization, and packet loss rate; S_history represents historical scheduling data, including task completion status and resource utilization trends over the past N scheduling cycles; the dimension of the state vector is M × N, where M is the number of resource nodes and N is the feature dimension of each node; Step S412: Design the action space of the scheduling system; The action space is defined as the set of scheduling decisions. A = {a_select, a_place, a_migrate, a_scale} a_select represents selecting the next task to be scheduled from the task queue, using a priority sorting strategy; a_place represents the probability distribution of candidate compute nodes and resource types output to the policy network, with heuristic rules used for action filtering or verification; a_migrate represents making migration decisions for running tasks to achieve load balancing; a_scale represents making dynamic scaling decisions for virtual resources; the action space adopts a discretized design, with each action dimension containing multiple optional values; Step S413: Design the reward function R; The reward function R is expressed as R = w1 × R_efficiency + w2 × R_quality + w3 × R_cost + w4 × R_fairness Wherein, R_efficiency represents the resource utilization efficiency reward, defined as the negative deviation between the actual resource utilization rate and the target utilization rate; R_quality represents the service quality reward, defined as the comparison between the task completion time and the expected completion time; R_cost represents the cost reward, defined as the comparison between the computational resource usage cost and the budget; R_fairness represents the fairness reward, defined as the degree of balance in the waiting time of each task; the weight coefficients w1, w2, w3, and w4 are obtained through offline training optimization; Step S414: Construct a deep neural network architecture; Deep reinforcement learning adopts the Actor-Critic architecture, in which the Actor network uses an attention-enhanced feature extractor, including an input layer that receives the state vector; the first layer is a fully connected layer with 1024 nodes and uses the ReLU activation function; the second layer is a self-attention layer with 8 attention heads; the third layer is a fully connected layer with 512 nodes; and the output layer is the policy output, which generates the action probability distribution.

6. The intelligent scheduling method for heterogeneous, geographically dispersed cloud computing services according to claim 4, characterized in that: Step S42 includes Step S421: Collect training data; The training data acquisition module is deployed in each data center, and the data collected includes task submission logs, resource usage logs, and scheduling decision logs; data acquisition is conducted asynchronously. Step S422: Design an experience playback mechanism; A priority experience replay mechanism is designed to store training samples; each experience sample contains a quintuple; the priority calculation formula is: P(i) = |δ_i|^α + ε Where δ_i is the timing difference error, α is the priority exponent (usually taken as 0.6), and ε is the minimum priority constant; a hierarchical sampling strategy is adopted during sampling. Step S423: Design a distributed training architecture; The design parameter server architecture is used for distributed training, and the training adopts an asynchronous update strategy. The model parameters are trained using mixed precision.

7. An intelligent scheduling system for heterogeneous, geographically dispersed cloud computing services, characterized by: include Heterogeneous Resource Abstraction Layer: Used to build a unified abstraction model for heterogeneous resources, mapping heterogeneous computing resources to a unified resource description language; Intelligent scheduling engine layer: Collects real-time status information and outputs scheduling decisions based on an adaptive scheduling algorithm using deep reinforcement learning; Cross-regional collaboration layer: used to coordinate resource scheduling between multiple regional data centers; Application Interface Layer: Provides application interfaces to support business systems in submitting scheduling requests conveniently, and provides a visual monitoring interface to facilitate operations and maintenance personnel in monitoring the system's operating status in real time.

8. The intelligent scheduling system for heterogeneous, geographically dispersed cloud computing services according to claim 7, characterized in that: The heterogeneous resource unified abstraction model adopts a hierarchical structure design, including a physical resource layer, a virtualization abstraction layer, a resource description layer, and a service interface layer. Physical Resource Layer: Defines the physical resources of various heterogeneous computing resources and collects real-time status data of the underlying hardware; The virtualization abstraction layer maps physical resources to virtual computing units, and uses container technology and virtualization technology to achieve resource isolation and sharing; each virtual computing unit includes the number of computing cores, memory capacity, and storage bandwidth attributes; this layer introduces a resource compatibility matrix to describe the resource requirements of different types of tasks and the degree of matching of various resources; The resource description layer converts the output of the virtualization abstraction layer into a unified resource description language format; the logical attributes of the resource description language include resource type identifier, capability level, current load status, and availability metrics. The service interface layer provides two interface forms: RESTful API and gRPC, for the upper-layer scheduling system to call.

9. The intelligent scheduling system for heterogeneous, geographically dispersed cloud computing services according to claim 7, characterized in that: The intelligent scheduling engine layer includes a real-time status perception module, a strategy decision module, and an execution feedback module. The real-time status perception module is responsible for collecting the resource usage status of data centers in various regions; the strategy decision module outputs scheduling decisions based on a deep reinforcement learning network; and the execution feedback module feeds the scheduling results back to the training module to achieve continuous optimization of the strategy.

10. The intelligent scheduling system for heterogeneous, geographically dispersed cloud computing services according to claim 7, characterized in that: The adaptive scheduling algorithm adopts an Actor-Critic architecture, which includes a state representation network component, a policy network component, and a value network component. The state representation network component transforms the system state into a feature vector that the algorithm can process; the state representation network uses a multilayer perceptron and attention mechanism to extract features and capture key information in the state space. The policy network component outputs scheduling decisions. The policy network uses a Softmax output layer to give the output scheduling decisions probabilistic distribution characteristics. The value network component evaluates the merits of the current scheduling strategy. The inputs to the value network are the outputs of the state representation network and the policy network, and the output is the value estimate of the state-action pair.

Images