Mass heterogeneous task elastic scheduling method based on multi-granularity state perception

Abstract

The invention discloses a massive heterogeneous task elastic scheduling method based on multi-granularity state awareness, which belongs to the field of cloud computing, and adopts the technical scheme that the method comprises the following steps: collecting operation data of a system in a continuous time window; processing the operation data to obtain a fusion state vector; splicing the fusion state vector and the feature vector of the current to-be-scheduled task to obtain a state required by a task scheduling decision; and generating a candidate target node set according to task feature information and system resource constraints to construct an action space, and transmitting the fusion state vector and related features of the current to-be-scheduled task as input to a strategy network based on reinforcement learning to obtain a scheduling result. The elastic scheduling method for the massive heterogeneous tasks has the beneficial effect that the elastic scheduling method for the massive heterogeneous tasks can realize multi-granularity state sensing in a long-term dynamic environment.

Description

Technical Field

[0001] This invention belongs to the field of cloud computing, and specifically relates to a method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness. Background Technology

[0002] With the development of the Industrial Internet, artificial intelligence, and large-scale online services, cloud computing systems have become the core support platform for modern information. Various application tasks need to be executed collaboratively across the cloud, edge, and local devices, with systems handling tens of thousands of heterogeneous computing tasks daily. Due to the sheer number and diverse types of tasks, and the highly random arrival patterns, cloud computing systems are often in a long-term dynamic operating state. For example, during peak business periods, a large number of sudden tasks arriving simultaneously can intensify resource competition and increase task queuing latency; while during low-load periods, some nodes may experience resource idleness, leading to a significant decrease in overall system utilization. To improve overall resource utilization and service quality, cloud platforms typically deploy task scheduling modules to dynamically allocate computing and communication resources based on factors such as node computing resources, task priority, and execution overhead. Currently, mainstream scheduling frameworks (such as Kubernetes and OpenStack) generally achieve resource allocation by monitoring single metrics such as CPU utilization, memory utilization, and bandwidth usage. However, these methods are generally based on the assumption of a relatively static system and stable task arrival, lacking detailed modeling of long-term dynamic changes and heterogeneous resource characteristics. This leads to scheduling decision failures and significant performance degradation in complex, highly dynamic scenarios. Therefore, how to achieve accurate perception of system operating status and adaptive task scheduling in a long-term dynamic environment has become a key direction for the intelligent development of cloud computing systems.

[0003] Reinforcement learning, with its self-learning ability in complex environments, has gradually become an important research direction in cloud computing task scheduling. This method continuously updates the state-action value function through interaction with the environment, enabling the scheduler to autonomously optimize resource allocation strategies in a nonlinear, high-dimensional state space. In cloud computing systems, task scheduling can be modeled as a continuous state transition process, with system states typically composed of indicators such as node load, task queues, and communication bandwidth. Reinforcement learning models utilize neural networks to approximate the value function, achieving end-to-end optimization from state perception to scheduling decision-making, and have achieved some success in small- to medium-scale systems. However, existing models still represent states in a relatively simplistic way, typically considering only instantaneous features at the task or node level, making it difficult to reflect the hierarchical relationships of the system in the temporal and spatial dimensions. As system scale increases and environmental non-stationarity intensifies, models are prone to estimation biases and policy oscillations, leading to scheduling instability, frequent task migrations, and overall performance degradation.

[0004] To enhance the ability of scheduling strategies to perceive and adapt to complex environments, some studies have introduced deep neural networks and temporal modeling techniques into reinforcement learning frameworks. For example, long short-term memory networks or gated recurrent units are used to capture the temporal dependencies between task arrival and resource utilization, thereby enabling dynamic prediction of system state change trends. These methods show certain advantages in short-term sequence modeling, alleviating the lag problem of traditional models relying solely on instantaneous states. However, as system runtime increases and environmental changes accumulate, temporal networks are prone to gradient decay and memory forgetting, making it difficult for the model to maintain long-term state memory. Furthermore, feature modeling of temporal networks is usually limited to single nodes or task levels, making it difficult to simultaneously characterize the topological structure and resource interaction relationships between different nodes. To address this, some studies have begun to explore the use of graph neural networks (GNNs) for cloud computing state modeling, achieving dependency representation between nodes through multi-layer information aggregation between node features and edge features. Although this method has achieved some success in capturing local structural associations, its high computational complexity and sensitivity to topological updates make it prone to structural mismatch in cloud computing environments with frequent dynamic node changes. Therefore, while existing deep learning-based state modeling methods have improved the level of scheduling intelligence to some extent, they still fail to achieve comprehensive fusion and robust perception of multi-granular spatiotemporal features. Summary of the Invention

[0005] The purpose of this invention is to provide a method for elastic scheduling of massive heterogeneous tasks that can achieve multi-granularity state awareness in long-term dynamic environments.

[0006] This invention is achieved through the following measures: In a first aspect, this embodiment provides a method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness, characterized in that it includes: S1: Collect system operation data within a continuous time window; S2: Process the running data to obtain a fused state vector; by taking the long-term running data of the system as input, a unified, high-dimensional system state representation vector is finally output, providing a basis for subsequent scheduling decisions.

[0007] S3. Concatenate the fused state vector with the feature vector of the current task to be scheduled to obtain the state required for task scheduling decision. S4. Based on task feature information and system resource constraints, a set of candidate target nodes is generated to construct the action space. Candidate nodes that are unreachable or lack sufficient resources are excluded. The fused state vector, along with the relevant features of the current task to be scheduled, is then passed as input to the reinforcement learning-based policy network to obtain the scheduling result. The task feature information and system resource constraints include the source node of the current task, the task type, computation / communication requirements, and system resource constraints.

[0008] Furthermore, the running data is processed to obtain the fused state vector, including: Semantic abstraction is performed on the runtime data to obtain a global semantic embedding vector; An undirected graph is constructed based on the link relationships between nodes, and the local state and topological dependency of the nodes are modeled to obtain the local structure vector; the global semantic embedding vector and the local structure vector are concatenated to obtain the fused state vector.

[0009] Furthermore, semantic abstraction is performed on the runtime data to obtain a global semantic representation, including: The operational data is structured to form a system state set. This operational data typically consists of multi-source heterogeneous observation metrics over a continuous time window of t, including operational metrics, task load, network topology, resource usage, and historical scheduling information. These metrics are structured to form a comprehensive system state set. This state set includes original node information (hardware configuration, geographical location, etc. of each computing node / edge node), task load (task arrival rate, task type distribution, task requirements for CPU / GPU / memory, etc.), resource distribution characteristics (utilization of CPU, GPU, memory, and storage for each node, etc.), link topology characteristics (link bandwidth and latency between nodes, etc.), and historical operational trajectories (historical change curves of key performance indicators, historical task scheduling records, and execution results, etc.).

[0010] The system state set is mapped to natural language prompts; the prompts include node load distribution, task arrival patterns, cross-node collaboration relationships, resource bottlenecks, and evolution trends. The natural language prompts and system state set are input into the large language model.

[0011] Furthermore, an undirected graph is constructed based on the link relationships between nodes, and the local states and topological dependencies of the nodes are modeled to obtain local structure vectors, including: Based on the physical links and collaborative relationships between nodes in the cloud platform, an undirected graph structure is constructed; G={V,E}, where the node set V represents edge base stations or computing nodes, and the edge set E represents the link connection relationship between nodes.

[0012] Collect local resource usage information and task execution status of each node to form a multi-dimensional node status feature vector, including but not limited to: CPU / GPU utilization, number of queued tasks, number of currently executing tasks, average task waiting time, and average link utilization.

[0013] Multi-level adjacency relationships are introduced, and a multi-level adjacency matrix and a multi-head graph attention mechanism are used to perform weighted aggregation of the state of the node's neighbors; By combining the node's own characteristics with information aggregated from its neighbors, a local structure vector is obtained.

[0014] Furthermore, the fused state vector, along with the relevant features of the current task to be scheduled, is fed as input into a reinforcement learning-based policy network to obtain the scheduling result, including: In each decision step, the fused state vector, along with the relevant features of the current task to be scheduled, is input into all M main Q networks to obtain a set of action value estimates. Calculate the mean and standard deviation of the above valuations; The Q-value after confidence penalty is defined based on the mean and standard deviation; The Q-value is normalized, and the optimal decision action is selected through multinomial sampling.

[0015] Furthermore, the scheduling method also includes optimizing the Q-value update process, including: Select the optimal action based on the Q value; The fused state vector, along with the relevant features of the current task to be scheduled, is input into all target Q networks to extract the estimated value of the corresponding action; Minimize along the network dimension to construct a conservative path, calculate the Q-value of the minimum path, and update the Q-value of the minimum path. After each round of policy updates, the main network and the target network are updated synchronously.

[0016] Secondly, this embodiment provides a massive heterogeneous task elastic scheduling system based on multi-granularity state awareness, characterized in that it includes a multi-granularity state awareness model and a real-time scheduling decision module. The multi-granularity state-aware model includes a global semantic branch module, a structural topology branch module, and a multi-granularity feature fusion module. The global semantic branch module is used to perform semantic abstraction on the running data to obtain a global semantic embedding vector. The structural topology branching module constructs an undirected graph based on the link relationships between nodes and models the local state and topological dependencies of nodes to obtain local structural vectors. The multi-granularity feature fusion module is used to concatenate the global semantic embedding vector and the local structure vector, and to standardize the format and align the dimensions of the state representations from different branches to obtain a fused state vector. Through this multi-granularity feature fusion module, the effective integration of system-level global trends and node-level local dynamics is achieved, enhancing the semantic expressiveness and robustness of state modeling.

[0017] Furthermore, the real-time scheduling decision module comprises: The fusion state input module is used to receive the fusion state vector and the relevant features of the current task to be scheduled; The action space construction module is used to generate a set of candidate target nodes based on task feature information and system resource constraints to construct the action space. The multi-Q valuation generation module is used to obtain action value valuations by utilizing M master Q networks; The confidence adjustment strategy generation module is used to calculate the mean and standard deviation of the above estimates and define the Q value after confidence penalty based on the mean and standard deviation; The strategy sampling and execution module normalizes the Q-value and selects the best decision action through multinomial sampling.

[0018] Thirdly, this embodiment provides an electronic device, characterized in that it includes: a processor and a memory, wherein the processor is used to execute a program stored in the memory for a method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness, so as to realize the method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness.

[0019] Fourthly, this embodiment provides a storage medium, characterized in that the storage medium stores one or more programs, which can be executed by one or more processors to realize the elastic scheduling method for massive heterogeneous tasks based on multi-granularity state awareness.

[0020] The beneficial effects of the technical solution provided by the embodiments of the present invention are as follows: Through a multi-granularity state perception mechanism that combines a large language model with a policy network of reinforcement learning, the dynamics of tasks and resources are perceived from both semantic and topological dimensions, realizing collaborative modeling of cross-time slot trend prediction and inter-node feedback adjustment; based on a conservative estimation path and variance penalty mechanism, overestimation bias and policy oscillation are effectively avoided, enabling the policy to maintain stable convergence and robust decision-making in dynamic load environments; the two work together to construct an intelligent scheduling system for heterogeneous resources and asynchronous tasks, significantly improving global optimization capabilities and task execution efficiency. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings listed below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 The flowchart shows a method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness provided in the implementation of this invention. Figure 2 This is a flowchart illustrating the state perception and real-time scheduling decision-making process in the implementation of this invention. Figure 3This is a schematic diagram illustrating the specific structure of the global semantic branching module and the structural topology branching module in this invention. Figure 4 This is a structural diagram of the real-time scheduling decision module based on multi-Q network and confidence adjustment mechanism in the implementation of this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] Example 1: See Figures 1-4 This embodiment provides a method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness, characterized by including: S1: Collect system operation data within a continuous time window; S2: Process the running data to obtain the fused state vector; by processing the long-term running data of the system... As input, the final output is a unified, high-dimensional system state representation vector, providing a foundation for subsequent scheduling decisions. This step specifically includes: semantically abstracting the runtime data to obtain a global semantic embedding vector; constructing an undirected graph based on inter-node link relationships and modeling the local states and topological dependencies of nodes to obtain a local structure vector; and concatenating the global semantic embedding vector and the local structure vector to obtain a fused state vector.

[0025] Specifically, semantic abstraction is performed on the runtime data to obtain a global semantic representation, including: The operational data is structured to form a system state set. This operational data typically consists of multi-source heterogeneous observation metrics over a continuous time window of t, including operational metrics, task load, network topology, resource usage, and historical scheduling information. These metrics are structured to form a comprehensive system state set. This state set includes original node information (hardware configuration, geographical location, etc. of each computing node / edge node), task load (task arrival rate, task type distribution, task requirements for CPU / GPU / memory, etc.), resource distribution characteristics (utilization of CPU, GPU, memory, and storage for each node, etc.), link topology characteristics (link bandwidth and latency between nodes, etc.), and historical operational trajectories (historical change curves of key performance indicators, historical task scheduling records, and execution results, etc.).

[0026] The set of states is represented as , .

[0027] in, It represents the physical information of each node in the base station, such as its GPU / CPU computing power, storage / memory scale, etc. This indicates the current average system load rate (only currently executing tasks are counted to avoid redundancy with other metrics); This indicates whether the distribution of available resources is balanced and whether there are bottleneck base station nodes. These represent the task load and distribution, respectively. These represent (link reachability, average path distance) and (link utilization distribution, hotspot identification), respectively. It represents the resource usage trajectory and task statistics summary of the past Δt steps, used to construct the evolution trend.

[0028] The system state set is mapped to natural language prompts; the prompts include node load distribution, task arrival patterns, cross-node collaboration relationships, resource bottlenecks, and evolution trends; the mapping prompt word function is: This function converts the structured system state into natural language input containing the following semantic elements: (i) the distribution and imbalance of resource loads on each node; (ii) the current system's task load and arrival patterns; (iii) cross-node collaboration patterns or bottleneck characteristics; and (iv) recent task growth trends and resource stress trends.

[0029] prompt word function and system status Input pre-trained large language model The large language model can be a large language model based on the Transformer structure.

[0030] Extract global semantic embedding vectors from the multi-layer semantic features output by the large language model. By extracting the most representative parts from the multi-layered Transformer features output by the large language model through pooling or attention mechanisms, a final global semantic embedding vector is formed. This vector concisely summarizes the overall health, resource pressure, task mode, and potential risks of the system at time t.

[0031] Specifically, an undirected graph is constructed based on the link relationships between nodes, and the local states and topological dependencies of the nodes are modeled to obtain local structure vectors, including: Based on the physical links and collaborative relationships between nodes in the cloud platform, an undirected graph structure is constructed; G={V,E}, where This represents the set of edge base station nodes (N base stations). This represents the connection links between edge nodes.

[0032] Collect local resource usage information and task execution status of each node to form a multi-dimensional node state feature vector. , ,in, These represent total resources, CPU utilization, and GPU utilization, respectively. These represent the node's task queue length and the number of currently running tasks, respectively. This represents the average queuing time for tasks at the node. Represents a node The average link utilization rate between the link and its neighbors is calculated as follows:

[0033] in Indicates the traffic on the link. This indicates the channel capacity between two nodes. It is a small positive number to prevent division by zero errors.

[0034] Introducing multi-order adjacency relations, defining adjacency matrix It also aggregates neighborhood state information using a multi-head graph attention mechanism, where the first... Output of the k-th attention head in layer Represented as:

[0035] in, For the Sigmod activation function, It is the weight matrix for each attention head. The residual mapping matrix is... This indicates a splicing operation. This indicates the first layer. Individual attention weight.

[0036] By combining the node's own features with information aggregated from its neighbors, a local structure vector is obtained. .

[0037] Specifically, for global semantic embedding vectors With local structure vector The vectors are concatenated to obtain the fused state vector. , Projection and normalization can be performed during splicing to meet the input format of the policy network.

[0038] S3, merging the state vector The feature vector of the currently scheduled task By concatenating the data, we can obtain the states required for task scheduling decisions. , .

[0039] S4. Based on task feature information and system resource constraints, a set of candidate target nodes is generated to construct the action space. Candidate nodes that are unreachable or lack sufficient resources are excluded. The fused state vector, along with the relevant features of the current task to be scheduled, is then passed as input to the reinforcement learning-based policy network to obtain the scheduling result. The task feature information and system resource constraints include the source node of the current task, the task type, computation / communication requirements, and system resource constraints.

[0040] Specifically, the fused state vector, along with the relevant features of the current task to be scheduled, is fed as input into a reinforcement learning-based policy network to obtain the scheduling result, including: In each decision step, the system considers the current state. Input all M main Q networks to obtain a set of action value estimates. In each decision step, the state required for task scheduling decision-making is determined. Input all M master Q networks to obtain a set of action value estimates. ,in This represents the parameters of the m-th master Q-network, and each network... It will output Each value is used to represent the state. The following are the scores for the corresponding actions, among which... This represents the set of selectable actions. Represents a set of actions The dimensions of selectable actions.

[0041] The average of the above valuations was calculated. with standard deviation , specifically,

[0042] The Q-value is defined based on the mean and standard deviation after confidence penalty. ,in This is the uncertainty penalty coefficient, used to balance the mean and the degree of confidence penalty.

[0043] The Q-value is normalized, and the optimal decision action is selected through multinomial sampling. , specifically, ; In the formula, Indicates the state Select action below Time Network Output rating, This represents the average value of the main network output. This indicates that the network parameters in the main network set are The m-th network, Indicates the best decision action. Indicates equipment The state that can be observed at time t.

[0044] Furthermore, the scheduling method also includes optimizing the Q-value update process, including: pass Select the optimal action :

[0045] The state required for task scheduling decisions Input all target Q networks Extract the valuation of the optimal action:

[0046] Construct a conservative path by taking the minimum value along the network dimension, calculate the Q-value of the minimum path, and update the Q-value of the minimum path: ,

[0047] In the formula, R represents the reward value. For instant rewards, This represents the discount factor.

[0048] After each round of policy updates, the main network and the target network are updated synchronously.

[0049] Example 2: This embodiment provides a massive heterogeneous task elastic scheduling system based on multi-granularity state awareness, characterized in that it includes a multi-granularity state awareness model and a real-time scheduling decision module; The multi-granularity state-aware model includes a global semantic branch module, a structural topology branch module, and a multi-granularity feature fusion module. The global semantic branch module is used to perform semantic abstraction on the running data to obtain a global semantic embedding vector. The structural topology branching module constructs an undirected graph based on the link relationships between nodes and models the local state and topological dependencies of nodes to obtain local structural vectors. The multi-granularity feature fusion module is used to concatenate the global semantic embedding vector and the local structure vector, and to standardize the format and align the dimensions of the state representations from different branches to obtain a fused state vector. Through this multi-granularity feature fusion module, the effective integration of system-level global trends and node-level local dynamics is achieved, enhancing the semantic expressiveness and robustness of state modeling.

[0050] Furthermore, the real-time scheduling decision module comprises: The fusion state input module is used to receive the fusion state vector and the relevant features of the current task to be scheduled; The action space construction module is used to generate a set of candidate target nodes based on task feature information and system resource constraints to construct the action space. The multi-Q valuation generation module is used to obtain action value valuations by utilizing M master Q networks; The confidence adjustment strategy generation module is used to calculate the mean and standard deviation of the above estimates and define the Q value after confidence penalty based on the mean and standard deviation; The strategy sampling and execution module normalizes the Q-value and selects the best decision action through multinomial sampling.

[0051] For the specific functions of each module, please refer to the relevant descriptions in the above method embodiments, which will not be repeated here.

[0052] Example 3: Based on Example 1 or 2, the network structure of the policy network for reinforcement learning includes M main Q-network submodules. Each of the M target Q networks is initialized independently and evaluated in parallel. This is a delayed update copy of the main network mentioned above; a confidence fusion module is used to construct the confidence penalty Q value based on the valuation mean and variance; and a minimum path selector is used to select the most conservative valuation path in the target valuation.

[0053] Each main Q-network structure is a three-layer fully connected feedforward neural network using ReLU activation. .

[0054] in, These represent the i-th fully connected layer.

[0055] Example 4: This embodiment provides an electronic device, characterized in that it includes: a processor and a memory, wherein the processor is used to execute a program stored in the memory for a method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness, so as to implement the method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness.

[0056] An electronic device includes at least one processor, memory, at least one network interface, and other user interfaces. The various components of the electronic device are coupled together via a bus system. It is understood that the bus system is used to enable communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0057] The user interface may include a display, keyboard, or clicking device (e.g., mouse, trackball, touchpad, or touchscreen). It is understood that the memory in this embodiment may be volatile memory or non-volatile memory, or may include both.

[0058] In this embodiment of the invention, the processor executes the method steps provided in each method embodiment by calling a program or instruction stored in the memory, specifically a program or instruction stored in an application program.

[0059] In some implementations, the memory stores elements such as executable units or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0060] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. The program implementing the method of this invention can be included in the application programs.

[0061] Example 5: This embodiment provides a storage medium, characterized in that the storage medium stores one or more programs, which can be executed by one or more processors to realize the elastic scheduling method for massive heterogeneous tasks based on multi-granularity state awareness.

[0062] The method steps described in conjunction with Embodiment 1 disclosed herein can be implemented using hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness, characterized in that, include: Collect system operation data within a continuous time window; The fused state vector is obtained by processing the runtime data; The fused state vector is concatenated with the feature vector of the current task to be scheduled to obtain the state required for task scheduling decision. A set of candidate target nodes is generated based on task feature information and system resource constraints to construct the action space. The fused state vector, together with the relevant features of the current task to be scheduled, is then passed as input to the policy network based on reinforcement learning to obtain the scheduling result.

2. The scheduling method according to claim 1, characterized in that, The fused state vector is obtained by processing the runtime data, including: Semantic abstraction is performed on the runtime data to obtain a global semantic embedding vector; An undirected graph is constructed based on the link relationships between nodes, and the local state and topological dependency of the nodes are modeled to obtain the local structure vector; the global semantic embedding vector and the local structure vector are concatenated to obtain the fused state vector.

3. The scheduling method according to claim 2, characterized in that, Semantic abstraction is performed on the runtime data to obtain a global semantic representation, including: Structure the runtime data to form a system state set; Map the system state set to natural language prompts; The natural language prompts and system state set are input into the large language model; Global semantic embedding vectors are extracted from the multi-layer semantic features output by the large language model.

4. The scheduling method according to claim 2, characterized in that, An undirected graph is constructed based on the link relationships between nodes, and the local states and topological dependencies of the nodes are modeled to obtain local structure vectors, including: Based on the physical links and collaborative relationships between nodes in the cloud platform, an undirected graph structure is constructed. Collect local resource usage information and task execution status of each node to form a multi-dimensional node status feature vector; Multi-level adjacency relationships are introduced, and a multi-level adjacency matrix and a multi-head graph attention mechanism are used to perform weighted aggregation of the state of the node's neighbors; By combining the node's own characteristics with information aggregated from its neighbors, a local structure vector is obtained.

5. The scheduling method according to claim 2, characterized in that, The fused state vector, along with the relevant features of the task to be scheduled, is fed as input into a reinforcement learning-based policy network to obtain the scheduling result, including: In each decision step, the fused state vector, along with the relevant features of the current task to be scheduled, is input into all M main Q networks to obtain a set of action value estimates. Calculate the mean and standard deviation of the above valuations; The Q-value after confidence penalty is defined based on the mean and standard deviation; The Q-value is normalized, and the optimal decision action is selected through multinomial sampling.

6. The scheduling method according to claim 5, characterized in that, This also includes optimizing the Q-value update process, including: Select the optimal action based on the Q value; The fused state vector, along with the relevant features of the current task to be scheduled, is input into all target Q networks to extract the estimated value of the corresponding action; Minimize along the network dimension to construct a conservative path, calculate the Q-value of the minimum path, and update the Q-value of the minimum path. After each round of policy updates, the main network and the target network are updated synchronously.

7. A massive heterogeneous task elastic scheduling system based on multi-granularity state awareness, characterized in that, This includes a multi-granularity state-aware model and a real-time scheduling decision module; The multi-granularity state-aware model includes a global semantic branch module, a structural topology branch module, and a multi-granularity feature fusion module. The global semantic branch module is used to perform semantic abstraction on the running data to obtain a global semantic embedding vector. The structural topology branching module constructs an undirected graph based on the link relationships between nodes and models the local state and topological dependencies of nodes to obtain local structural vectors. The multi-granularity feature fusion module is used to concatenate the global semantic embedding vector and the local structure vector to obtain the fused state vector.

8. The scheduling system according to claim 7, characterized in that, The real-time scheduling decision module includes: The fusion state input module is used to receive the fusion state vector and the relevant features of the current task to be scheduled; The action space construction module is used to generate a set of candidate target nodes based on task feature information and system resource constraints to construct the action space. The multi-Q valuation generation module is used to obtain action value valuations by utilizing M master Q networks; The confidence adjustment strategy generation module is used to calculate the mean and standard deviation of the above estimates and define the Q value after confidence penalty based on the mean and standard deviation; The strategy sampling and execution module normalizes the Q-value and selects the best decision action through multinomial sampling.

9. An electronic device, characterized in that, include: A processor and a memory, the processor being configured to execute a program stored in the memory for a method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness, to implement the method for elastic scheduling of massive heterogeneous tasks based on multi-granularity state awareness as described in any one of claims 1 to 6.

10. A storage medium, characterized in that, The storage medium stores one or more programs, which can be executed by one or more processors to implement the elastic scheduling method for massive heterogeneous tasks based on multi-granularity state awareness as described in any one of claims 1 to 6.

Images