Cloud task scheduling method and device based on intelligent agent, equipment and medium

Abstract

The invention provides a cloud task scheduling method and device based on an intelligent agent, equipment and a medium, and can be applied to the technical field of complex cloud task scheduling. The cloud task scheduling method comprises the steps of obtaining task characteristics and user preferences in response to a task request; based on task characteristics and user preferences, an action plan of task scheduling is generated by utilizing an intelligent agent; according to the action plan, utilizing the intelligent agent to call each tool in sequence, and generating a task scheduling scheme, including: generating task description based on task characteristics and user preferences; collecting resource cost, resource use information and link condition information, and generating a network topological graph; predicting a task load fluctuation condition by using a time sequence prediction model; selecting a link node based on the network topological graph and the task load fluctuation condition; selecting a task scheduling method based on task types and user preferences; and generating a task scheduling scheme by using a task scheduling method based on the task description and the link nodes.

Description

Technical Field

[0001] The present invention relates to the technical field of complex cloud task scheduling, and more particularly to an agent-based cloud task scheduling method, apparatus, device and medium. Background Art

[0002] In today's digital age, cloud computing has become the core technological force driving the development of various industries. With the continuous expansion of application scenarios and the increasing complexity of business requirements, the cloud computing environment has gradually evolved into a complex ecosystem where local clouds, public clouds, and edge clouds are intertwined. Local clouds, with their advantages in enterprise internal deployment, can provide highly customized resource services for enterprises, ensuring data security and controllability, and are particularly suitable for industries with strict requirements for data privacy protection, such as finance and healthcare. However, their resource scale is often limited by the enterprise's own hardware investment, and their resource expansion ability is relatively limited when dealing with large-scale sudden tasks. Public clouds, with their powerful resource pools, flexible resource configurations, and convenient pay-as-you-go models, have attracted a large number of small and medium-sized enterprises that are sensitive to costs and have large fluctuations in resource requirements. However, in a public cloud environment, resource competition may occur among multiple tenants sharing resources, resulting in uncertainty in service quality. The emergence of edge clouds effectively makes up for the deficiencies of local clouds and public clouds in real-time processing. It is close to data sources and user terminals, can quickly process data at the network edge, significantly reduce data transmission latency, and plays a key role in fields with extremely high requirements for real-time response, such as the Internet of Things, real-time video processing, and autonomous driving. However, edge cloud node resources are relatively scattered, resource management is difficult, and the resource capacity of a single node is limited.

[0003] Facing such a complex and diverse cloud environment, achieving unified management of resources and full scheduling of tasks has become a key challenge in improving cloud computing efficiency. The resource types of local clouds, public clouds, and edge clouds are diverse, covering computing resources (such as CPUs, GPUs), storage resources (such as hard disks, memory), and network resources (such as bandwidth, IP addresses), etc. These resources not only differ in physical location and performance parameters, but their resource usage status is also in dynamic change. For example, the utilization rate of the CPU will change in real-time with the fluctuation of the task load, and the memory occupancy will also be continuously adjusted due to the operation of different application programs. Traditional resource management models often act independently and lack a unified coordination mechanism across cloud platforms, resulting in coexistence of resource idle and waste phenomena and difficulty in improving resource utilization efficiency.

[0004] Meanwhile, the types of tasks faced in complex cloud environments are rich and diverse, and each task exhibits unique characteristics in terms of resource requirements, task structure, and scheduling objectives. Some tasks are urgent and non-interruptible, and such tasks have extremely high performance requirements for computing resources and network resources, and it is necessary to ensure that sufficient resources are invoked within the shortest time to achieve the task objectives, that is, to complete the task as soon as possible and ensure the smooth progress of the task. Scientific research tasks are characterized by large amounts of computation and long task cycles, and require a large amount of computing resources to be occupied for a long time to process massive data. Scientific research tasks usually allow interruption under certain conditions, and their scheduling objectives focus more on the lowest cost. How to make reasonable use of the resource price differences of different cloud platforms and reduce the overall cost through optimized resource allocation becomes the key. Autonomous driving tasks represent a type of task scenario with extremely high requirements for real-time performance and reliability. Vehicles need to process a large amount of data from sensors in real time to make accurate driving decisions, such as accelerating, braking, and steering. Therefore, autonomous driving tasks need to invoke edge resources close to the vehicle, utilize the low-latency characteristics of the edge cloud to quickly process data, minimize latency, and ensure the high reliability of the system through technical means such as redundant backups, ensuring safe and stable operation in various complex road conditions and environments.

[0005] In summary, the diversity of complex cloud environments and the complexity of task types pose unprecedented challenges to cloud task scheduling methods and systems. There is an urgent need for an innovative solution to achieve efficient utilization of resources and precise scheduling of tasks to meet the growing digital needs of various industries.

[0006] Facing complex cloud environments and diverse task types, efficient task scheduling is a key link in achieving efficient utilization of resources and meeting user needs. As the complexity of cloud tasks continues to increase, traditional task scheduling methods face many challenges. On the one hand, existing cloud scheduling algorithms mostly focus on a single objective, and a single scheduling method is difficult to comprehensively consider the diversity characteristics of tasks and the complex resource environment, while a few multi-objective scheduling algorithms seek a balance between multiple objectives and cannot meet the unique requirements of different tasks; on the other hand, for tasks containing multiple modal information, existing scheduling methods cannot fully utilize this information to make better scheduling decisions. Therefore, it is necessary to design a scheduling method that can meet various tasks in complex cloud environments. Summary of the Invention

[0007] (I) Technical Problems to be Solved

[0008] In view of the above problems, the present invention provides an agent-based cloud task scheduling method, device, equipment and medium, which comprehensively considers complex cloud environments and various types of tasks. Through an innovative agent architecture and multi-model fusion strategy, it effectively realizes the comprehensive consideration and accurate scheduling of multi-type tasks, assisting users or cloud application providers to better schedule tasks according to requirements, reducing the threshold of cloud resource usage, improving the utilization efficiency of cloud resources, and promoting the popularization and application of cloud computing.

[0009] (2) Technical solutions

[0010] In view of the above technical problems, embodiments of the present invention propose an agent-based cloud task scheduling method, device, equipment and medium.

[0011] According to the first aspect of the present invention, an agent-based cloud task scheduling method is provided, including: in response to a task request, obtaining task characteristics and user preferences; based on the task characteristics and user preferences, using an agent to generate a task scheduling action plan, where the agent includes a trained large model, a configuration module, a planning module, a memory module, and a variety of tools, and the tools include a task description generation tool, a node information acquisition tool, a load fluctuation prediction tool, a node selection tool, a task scheduling scheme generation tool, and a task execution tool; according to the action plan, using the agent to sequentially call each tool to generate a task scheduling scheme, where, according to the action plan, using the agent to sequentially call each tool to generate a task scheduling scheme includes: based on the task characteristics and user preferences, using the agent to call the task description generation tool to determine the task type, and integrating the task characteristics, user preferences, and task type, and generating a task description in a unified data format; using the agent to call the node information acquisition tool to collect resource costs, resource usage information, and link condition information, and generating a network topology diagram based on graph theory and visualization technology; using the agent to call the load fluctuation prediction tool to predict the task load fluctuation situation based on the trained time series prediction model; using the agent to call the node selection tool to select link nodes based on the network topology diagram and the task load fluctuation situation; using the agent to call the task scheduling scheme generation tool, based on the task type and user preferences, select a task scheduling method, and based on the task description and the selected link nodes, use the selected task scheduling method to generate a task scheduling scheme, where the task characteristics include the resource requirements of the task, the execution time requirement of the task, the priority of the task, the interruptibility of the task, and the dependency relationship of the task. In some exemplary embodiments, the large model includes a large model suitable for natural language processing and multi-modal information processing, and the method for training the large model includes: obtaining training data, where the training data includes data of various task characteristics; using the training data to train the large model to adjust the parameters of the large model and complete the training of the large model, where the large model can process multi-modal input data, and the input data includes data in text, time series, and image modalities.

[0012] In some exemplary embodiments, the time series prediction model includes a long short-term memory network. The method for training the time series prediction model includes: obtaining historical task load data and annotating the time points of key events, where the task load data includes traffic data and hardware status data, and the key events include traffic peaks and hardware failures; training the time series prediction model using the historical task load data, and adjusting the time series prediction model so that the time series prediction model can predict task load fluctuations.

[0013] In some exemplary embodiments, the task scheduling methods include a reinforcement learning scheduling method, a heuristic algorithm scheduling method, a meta-heuristic algorithm scheduling method, and a rule-based scheduling algorithm. Among them, the method for training the reinforcement learning scheduling method includes: based on the reinforcement learning scheduling method, setting a reward function, where the reward function is used to reflect the quality of the scheduling strategy; designing a state space and an action space, where the state space includes the resource requirements of tasks, whether they can be interrupted, and the resource status of the current cloud environment, and the action space includes different scheduling decisions; based on the reward function, the state space, and the action space, using a reinforcement learning algorithm to train the reinforcement learning scheduling method in a simulated cloud task environment so that the reinforcement learning scheduling method learns the optimal scheduling strategy; and / or, the method for training the heuristic algorithm scheduling method includes: setting heuristic rules based on the characteristics of tasks and the characteristics of the cloud environment; based on actual task data, running the heuristic algorithm using the heuristic algorithm scheduling method to optimize the parameters and rules of the heuristic algorithm; and / or, the meta-heuristic algorithm scheduling method includes a genetic algorithm model, and the training method of the genetic algorithm model includes: defining a chromosome encoding method and encoding the scheduling scheme as a chromosome; setting a fitness function to evaluate the quality of the scheduling scheme corresponding to each chromosome; through genetic operations such as selection, crossover, and mutation, continuously evolving the chromosome to obtain a better scheduling scheme, thereby training a genetic algorithm model applicable to cloud task scheduling; and / or, the method for training the rule-based scheduling algorithm includes: formulating scheduling rules according to domain knowledge and scheduling requirements; applying the scheduling rules in an actual task scenario and continuously adjusting and improving the scheduling rules according to the feedback.

[0014] In some exemplary embodiments, the nodes in the network topology diagram represent cloud resource nodes, and the edges represent the links between the nodes; the network topology diagram includes the bandwidth, latency, and packet loss rate of the links; a dynamic update mechanism is adopted to ensure that the network topology diagram reflects the actual state of the cloud environment in real time.

[0015] In some exemplary embodiments, selecting link nodes based on a network topology map and task load fluctuations includes: based on the network topology map, using a trained large model to analyze the bandwidth and latency information of each link; based on the bandwidth and latency information, identifying the key information of the link, including reliability and transmission efficiency; based on resource costs and task load fluctuations, obtaining initial link nodes; and further screening the initial link nodes based on task characteristics to obtain final link nodes.

[0016] In some exemplary embodiments, it further includes: allocating the task scheduling plan to the selected cloud nodes for execution; continuously monitoring the execution status of the task, where when an abnormal situation occurs during the task execution, corresponding countermeasures are taken, including: recording the abnormal information; selecting a corresponding processing method according to the severity and type of the abnormality; and feeding back the abnormal information to the task scheduling agent to adjust and optimize the task scheduling plan.

[0017] According to a second aspect of the present invention, there is provided an agent-based cloud task scheduling device, including: an acquisition module for obtaining task characteristics and user preferences in response to a task request; a planning module for generating a task scheduling action plan using an agent based on task characteristics and user preferences, where the agent includes a trained large model, a configuration module, a planning module, a memory module, and a variety of tools, and the tools include a task description generation tool, a node information acquisition tool, a load fluctuation prediction tool, a node selection tool, a task scheduling plan generation tool, and a task execution tool; a generation module for generating a task scheduling plan by sequentially invoking each tool using the agent according to the action plan, where generating a task scheduling plan by sequentially invoking each tool using the agent according to the action plan includes: based on task characteristics and user preferences, using the agent to invoke the task description generation tool to determine the task type, and integrating the task characteristics, user preferences, and task type, and generating a task description in a unified data format; using the agent to invoke the node information acquisition tool to collect resource costs, resource usage information, and link condition information, and generating a network topology map based on graph theory and visualization technology; using the agent to invoke the load fluctuation prediction tool to predict task load fluctuations based on a trained time series prediction model; using the agent to invoke the node selection tool to select link nodes based on the network topology map and task load fluctuations; using the agent to invoke the task scheduling plan generation tool to select a task scheduling method based on the task type and user preferences, and generating a task scheduling plan using the selected task scheduling method based on the task description and the selected link nodes, where the task characteristics include the resource requirements of the task, the execution time requirement of the task, the priority of the task, the interruptibility of the task, and the dependency relationship of the task.

[0018] According to a third aspect of the present invention, there is provided an electronic device, including: one or more processors; a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the above method.

[0019] A fourth aspect of the present invention further provides a computer-readable storage medium, on which a computer program or instruction is stored, and when the computer program or instruction is executed by a processor, the steps of the above method are implemented.

[0020] (III) Beneficial effects

[0021] As can be seen from the above technical solutions, a method, apparatus, device and medium for cloud task scheduling based on an agent provided by an embodiment of the present invention at least have the following beneficial effects:

[0022] (1) By fine-tuning a large model to construct a task scheduling agent, through the flexible decision-making of the agent and the collaborative work among various tools, the full process automation from data collection, task scheduling to task execution is realized, reducing manual intervention and improving scheduling efficiency.

[0023] (2) Based on the large model, an appropriate scheduling method is automatically selected according to the task type, enabling it to adapt to the task rules in different fields, improving the versatility and flexibility of the scheduling system, being able to handle diverse and complex task scenarios, and having strong compatibility with complex cloud scenarios such as heterogeneous clouds, distributed clouds, federated clouds, hybrid clouds, and edge clouds, and being able to flexibly integrate the resource pools of different cloud service providers to meet the complex IT architecture requirements of enterprises.

[0024] (3) It supports multiple modal inputs such as text, time-series data, and images, can comprehensively capture task features and environmental information, provides richer context information for scheduling decisions, and improves the comprehensiveness and accuracy of decisions.

[0025] (4) By combining a large model, a time-series prediction model, a reinforcement learning scheduling method, a heuristic scheduling method, a meta-heuristic scheduling method, and a rule-based scheduling method, the advantages of different scheduling strategies are complemented to meet the requirements of different types of tasks, ensuring the diversity and efficiency of scheduling decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Through the following description of the embodiments of the present invention with reference to the drawings, the above content and other objects, features and advantages of the present invention will become clearer. In the drawings:

[0027] Figure 1 Schematically shows a flowchart of a method for cloud task scheduling based on an agent according to an embodiment of the present invention.

[0028] Figure 2Schematically shows a flowchart of a task execution and monitoring method according to an embodiment of the present invention.

[0029] Figure 3 Schematically shows an agent-based cloud task scheduling device according to an embodiment of the present invention.

[0030] Figure 4 Schematically shows a block diagram of an electronic device for an agent-based cloud task scheduling method according to an embodiment of the present invention. Detailed implementation manners

[0031] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In the following detailed description, for the sake of explanation, many specific details are set forth in order to provide a thorough understanding of the embodiments of the present invention. However, it is obvious that one or more embodiments can be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily obscuring the concepts of the present invention.

[0032] The terms used herein are merely for describing specific embodiments and are not intended to limit the present invention. The terms "including", "comprising", etc. used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0033] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0034] In the case of using expressions such as "at least one of A, B, and C", generally, it should be interpreted according to the meaning commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include, but not be limited to, a system having only A, only B, only C, having A and B, having A and C, having B and C, and / or having A, B, and C).

[0035] Figure 1 Schematically shows a flowchart of an agent-based cloud task scheduling method according to an embodiment of the present invention.

[0036] As Figure 1 shown, an agent-based cloud task scheduling method according to an embodiment of the present invention includes step S110-step S130.

[0037] In step S110, in response to a task request, task characteristics and user preferences are obtained.

[0038] After the task submission information, detailed task characteristics are extracted therefrom, including the resource requirements of the task (such as the number of CPU cores, memory size, storage capacity, network bandwidth requirements, etc.), the execution time requirements of the task (such as the deadline, the expected execution duration, etc.), the priority of the task, the interruptibility of the task, the dependency relationship of the task, etc.

[0039] Embodiments of the present invention can provide diverse user preference setting methods, such as graphical interfaces, command-line inputs, configuration file uploads, etc. Users can set preferences according to their own needs, such as processing tasks at the fastest speed, processing tasks at the lowest cost, ensuring high reliability of task execution, preferentially using specific types of cloud resources, etc.

[0040] In step S120, based on the task characteristics and user preferences, an intelligent agent is used to generate a task scheduling action plan. The intelligent agent includes a trained large model, a configuration module, a planning module, a memory module, and a variety of tools. The tools include a task description generation tool, a node information acquisition tool, a load fluctuation prediction tool, a node selection tool, a task scheduling scheme generation tool, and a task execution tool.

[0041] When the task arrives, the intelligent agent (Agent) plans the action plan in real time and calls the corresponding tools as needed according to the planned plan to complete task scheduling and execution. During this process, the intelligent agent will dynamically adjust the action plan according to the situation in real time to ensure the completion of the task. For example, if a node is found to be unavailable during the task execution, an available node will be reselected and task scheduling and execution will be performed.

[0042] An intelligent agent is a computer program based on a large language model, which has the abilities of planning and thinking, memory, and using tools, and can autonomously complete a given task. It has autonomy, adaptability, and interaction capabilities, can perceive the environment, make autonomous decisions, and execute actions to achieve specific goals. It has the following characteristics: (1) Autonomy: The intelligent agent can control its own behavior based on its internal state and perception of the environment without external direct intervention. (2) Perception ability: It can perceive information about the surrounding environment through sensors or data interfaces, etc. (3) Decision-making ability: Based on the perceived environmental information and its own goals, the intelligent agent can make reasonable decisions using complex algorithms or logical systems. (4) Execution ability: The intelligent agent can convert decisions into actual actions, thereby affecting the environment or its own state.

[0043] In the embodiments of the present invention, the intelligent agent is composed of the following parts:

[0044] The trained large model, as the core foundation of the intelligent agent, is mainly responsible for analyzing and processing data, making decision plans, generating output results, etc., and uses its powerful language understanding and generation capabilities to handle various tasks. The large model includes large models applicable to natural language processing and multi-modal information processing. The methods for training the large model include: obtaining training data, where the training data includes data with various task characteristics. For example, medical task data is collected, emphasizing high reliability requirements and related indicator data; using the training data to train the large model to adjust the parameters of the large model, complete the training of the large model, and make it better adapt to the rules of various domain tasks during the learning process, so as to support the subsequent processing of multiple tasks. Among them, the large model can process multi-modal input data, and the input data includes text, time series data, and image modal data to improve the comprehensiveness of scheduling decisions.

[0045] The configuration module is used to define the role, specialty, skills, behavior patterns, etc. of the intelligent agent, and helps the intelligent agent show specific behaviors and capabilities in different scenarios.

[0046] The planning module is used to decompose complex tasks into executable small tasks, select and formulate an efficient workflow to achieve the goal, and guide the action process of the intelligent agent through reasoning and planning.

[0047] The memory module is a knowledge base for storing various types of data required by the intelligent agent. Its content comes from the information and data generated during the interaction between the intelligent agent and the environment, or from external knowledge bases, involving information acquisition, storage, management, and retrieval, etc., to help the intelligent agent make better decisions using historical information.

[0048] The tool module is used to provide the intelligent agent with richer skills than the large model itself, so as to expand the skill library of the intelligent agent, thereby improving the ability of the intelligent agent. It can receive input parameters, execute corresponding functions, and return results. The tools include: task description generation tool, node information acquisition tool, load fluctuation prediction tool, node selection tool, task scheduling plan generation tool, task execution tool.

[0049] In step S130, according to the action plan, the intelligent agent is used to sequentially call each tool to generate a task scheduling plan.

[0050] In some exemplary embodiments, step S130 includes steps S131 - S135.

[0051] In step S131, based on the task characteristics and user preferences, the intelligent agent is used to call the task description generation tool to judge the task type, and integrate the task characteristics, user preferences, and task type, and generate a task description in a unified data format.

[0052] The agent calls the task description generation tool on demand. Relying on a specially designed information collection interface, it efficiently and accurately obtains various detailed information related to the task. At the same time, it comprehensively collects the specific preferences of the user for task processing. According to the established data standards and specifications, it systematically integrates the obtained task information and user preferences to generate a unified and standardized task description, which serves as an important input for the subsequent task scheduling decision-making process. For example, information such as resource requirements and interruptibility is parsed from the task submission information.

[0053] The task description generation tool designs a standardized and flexible task information collection interface that supports multiple data formats and transmission protocols. This interface can be docked with task submission systems from different sources, such as in-house business systems and third-party task publishing platforms.

[0054] After collecting the task submission information, detailed task characteristics are extracted from it, including the resource requirements of the task (such as the number of CPU cores, memory size, storage capacity, network bandwidth requirements, etc.), the execution time requirements of the task (such as the deadline, estimated execution duration, etc.), the priority of the task, the interruptibility of the task, the dependency relationship of the task, etc.

[0055] Diversified user preference setting methods are provided, such as graphical interfaces, command-line inputs, configuration file uploads, etc. Users can set preferences according to their own needs, such as processing tasks at the fastest speed, processing tasks at the lowest cost, ensuring high reliability of task execution, and preferentially using specific types of cloud resources.

[0056] For example, according to the task characteristics and user preferences, the task type is judged, such as judging whether it is a compute-intensive task, a data transmission task, or other types of tasks.

[0057] The unified data format includes JSON format, XML format, etc. The task characteristics and user preference fields are clearly distinguished in the task description, which is convenient for subsequent processing and analysis.

[0058] In step S132, the agent calls the node information acquisition tool to collect resource costs, resource usage information, and link condition information, and generates a network topology diagram based on graph theory and visualization technology.

[0059] Resource costs and resource usage information are collected, such as CPU usage rate, memory occupancy rate, remaining storage resources, etc., as well as link condition information, such as network bandwidth, latency, packet loss rate, jitter, etc. For different types of cloud platforms, adapted communication methods are adopted, such as using the Representational State Transfer Application Programming Interface (RESTful API) to interact with public clouds for data, and communicating efficiently with local clouds through internal network protocols.

[0060] Clean and preprocess the massive amount of collected information to remove duplicate, incorrect, and invalid data. Based on the processed information, use graph theory and visualization techniques to generate an intuitive network topology map. In the network topology map, nodes represent cloud resource nodes, and edges represent the links between nodes; at the same time, key attributes such as the bandwidth, latency, and packet loss rate of the links are marked. Adopt a dynamic update mechanism to ensure that the network topology map reflects the actual state of the cloud environment in real time. Adopt a dynamic update mechanism to ensure that the network topology map reflects the actual state of the cloud environment in real time.

[0061] In step S133, use the agent to call the load fluctuation prediction tool, and based on the trained time series prediction model, predict the task load fluctuation.

[0062] Pass the relevant information of the current task, such as the task type, historical execution data, etc., to the time series prediction model. The time series prediction model predicts the task load fluctuation according to the received information and combines its own trained model parameters, such as predicting whether there will be a traffic peak in the future period of time, the possibility of hardware failure, etc.

[0063] In some exemplary embodiments, the time series prediction model includes a long short-term memory network, and the method for training the time series prediction model includes: obtaining historical task load data and marking the time points of key events, where the task load data includes traffic data and hardware status data, and the key events include traffic peaks and hardware failures; using the historical task load data to train the time series prediction model and adjusting the time series prediction model so that the time series prediction model can predict the task load fluctuation.

[0064] In step S134, use the agent to call the node selection tool, and based on the network topology map and the task load fluctuation, select link nodes.

[0065] In some exemplary embodiments, step S134 includes steps S1341 - S1344.

[0066] In step S1341, based on the network topology map, use the trained large model to analyze the bandwidth and latency information of each link.

[0067] In step S1342, based on the bandwidth and latency information, identify the key information of the link, including reliability and transmission efficiency.

[0068] In step S1343, based on the resource cost and the task load fluctuation, obtain the initial link nodes. If it is predicted that a traffic peak is about to occur, preferentially select the nodes with high-bandwidth and low-latency links to cope with the possible high-load situation; if the risk of hardware failure is predicted to be relatively high, avoid selecting the nodes that may be affected.

[0069] In step S1344, the initial link nodes are further screened based on the task characteristics to obtain the final link nodes. For example, for tasks with large resource requirements, nodes with abundant resources are selected; for tasks that cannot be interrupted, nodes with high reliability are selected.

[0070] In step S135, the intelligent agent is used to call the task scheduling solution generation tool, and the task scheduling method is selected based on the task type and user preference. Based on the task description and the selected link nodes, the task scheduling solution is generated using the selected task scheduling method.

[0071] Task scheduling methods can be targeted at different types of tasks, including parallel tasks and workflow tasks. Task scheduling methods can use different scheduling methods such as reinforcement learning, heuristic algorithms, meta-heuristic algorithms, and rule-based scheduling algorithms, which are suitable for different types of tasks. Task scheduling methods include reinforcement learning scheduling methods, heuristic algorithm scheduling methods, meta-heuristic algorithm scheduling methods, and rule-based scheduling algorithm methods.

[0072] Optionally, the method for training the reinforcement learning scheduling method includes: based on the reinforcement learning scheduling method, setting a reward function, for example, giving high rewards to scheduling strategies that can complete tasks quickly and have high resource utilization, and the reward function is used to reflect the pros and cons of the scheduling strategy; designing a state space and an action space, the state space includes the resource requirements of the task, whether it can be interrupted, and the resource status of the current cloud environment, and the action space includes different scheduling decisions, such as selecting different cloud nodes, allocating different amounts of resources, etc.; based on the reward function, state space and action space, using a reinforcement learning algorithm, such as a deep Q network algorithm, to train the reinforcement learning scheduling method in a simulated cloud task environment, so that the reinforcement learning scheduling method learns the optimal scheduling strategy.

[0073] Optionally, the method for training the heuristic algorithm scheduling method includes: setting heuristic rules based on task characteristics and characteristics of the cloud environment, such as allocating resources to urgent tasks with small resource requirements based on the urgency of the task and resource requirements; based on actual task data, running the heuristic algorithm using the heuristic algorithm scheduling method to optimize the parameters and rules of the heuristic algorithm.

[0074] Optionally, the meta-heuristic algorithm scheduling method includes a genetic algorithm model, and the training method of the genetic algorithm model includes: defining a chromosome encoding method and encoding the scheduling plan as a chromosome; setting a fitness function to evaluate the pros and cons of the scheduling plan corresponding to each chromosome; and continuously evolving the chromosomes through genetic operations such as selection, crossover, and mutation to obtain a better scheduling plan, thereby training a genetic algorithm model suitable for cloud task scheduling.

[0075] Optionally, the method for training a rule-based scheduling algorithm includes: formulating scheduling rules according to domain knowledge and scheduling requirements, such as preferentially allocating reliable cloud node resources for non-interruptible tasks; applying the scheduling rules in actual task scenarios and continuously adjusting and improving the scheduling rules according to feedback.

[0076] Transfer the task description, available nodes, and relevant information to the corresponding model. For example, for a compute-intensive task with the user preference of processing the task at the fastest speed, call a scheduling method based on reinforcement learning and optimized for computing resources; for a data transfer task, call a heuristic algorithm scheduling method considering network link conditions.

[0077] After receiving the input information, the scheduling method performs task planning respectively according to its own algorithm and parameters to generate a task scheduling plan. For example, the reinforcement learning scheduling method selects appropriate actions from the optimal strategies it has learned according to the current state to generate a specific scheduling plan, including information such as which cloud nodes the tasks are assigned to, how much resources are allocated to each node, and the execution order of the tasks. The heuristic algorithm scheduling method sorts and allocates resources to the tasks according to the designed heuristic rules to generate a scheduling plan. The meta-heuristic algorithm scheduling method generates a new scheduling plan through genetic operations and evaluates its pros and cons, and selects a better scheduling plan. The rule-based scheduling algorithm generates a scheduling plan according to the formulated scheduling rules.

[0078] On the Figure 1 basis of the method shown, it may further include using an agent to call a task execution tool for task execution and monitoring. The execution and monitoring method includes steps S210 - step S220, refer to Figure 2 .

[0079] In step S210, allocate the task scheduling plan to the selected cloud nodes for execution.

[0080] After receiving the scheduling plan generated by the task scheduling plan generation tool, the agent allocates the tasks to the selected cloud nodes by calling the task execution tool and orderly starts the task execution process. When allocating tasks, interact with the task management system of the cloud node and use a standardized interface protocol to transfer the relevant information of the tasks (such as task code, input data, configuration parameters, etc.) to the target node.

[0081] In step S220, continuously monitor the execution status of the task.

[0082] During the task execution process, closely monitor the execution progress of the task and the resource usage situation through a real-time monitoring mechanism. Once abnormal conditions occur during the task execution, such as resource shortage, link interruption and other problems, quickly feedback to the agent to ensure the continuity and stability of the task execution.

[0083] Specifically, during the task execution process, continuously monitor the execution status of the task. By communicating with the monitoring interface of the cloud node, real-time obtain the progress information of the task (such as the amount of work completed, the remaining amount of work, etc.), resource usage (such as CPU usage rate, memory occupancy rate, storage read / write rate, etc.). Use visualization technology to display the task execution status in an intuitive way, facilitating real-time monitoring and management by the administrator.

[0084] When abnormal situations occur during the task execution process, such as insufficient resources, link interruption, task failure, etc., take corresponding countermeasures in a timely manner. Record the abnormal information, including the time, type, and scope of influence of the abnormality; according to the severity and type of the abnormality, select appropriate processing methods, such as rescheduling the task, adjusting resource allocation, retrying the task, etc. At the same time, feedback the abnormal information to the task scheduling agent to adjust and optimize the action plan.

[0085] When the task execution is completed, verify and confirm the execution result of the task. Check whether the output result of the task meets the expectations and compare it with the input requirements and goals of the task. Feedback the task execution result to the relevant users and systems, and at the same time update the task status information and resource usage records.

[0086] Figure 3 Schematically shows an agent-based cloud task scheduling device according to an embodiment of the present invention.

[0087] As Figure 3 shown, the agent-based cloud task scheduling device 800 of this embodiment includes an acquisition module 810, a planning module 820, and a generation module 830.

[0088] The acquisition module 810 is used to obtain task characteristics and user preferences in response to a task request.

[0089] The planning module 820 is used to generate a task scheduling action plan using an agent based on task characteristics and user preferences, where the agent includes a trained large model, a configuration module, a planning module, a memory module, and a variety of tools, and the tools include a task description generation tool, a node information acquisition tool, a load fluctuation prediction tool, a node selection tool, a task scheduling scheme generation tool, and a task execution tool.

[0090] A generation module 830, configured to generate a task scheduling plan by sequentially invoking each tool using an agent according to an action plan. Here, generating a task scheduling plan by sequentially invoking each tool using an agent according to an action plan includes: based on the task characteristics and user preferences, using the agent to invoke a task description generation tool to determine the task type, and integrating the task characteristics, user preferences, and task type, and generating a task description in a unified data format; using the agent to invoke a node information acquisition tool to collect resource cost, resource usage information, and link condition information, and generating a network topology diagram based on graph theory and visualization technology; using the agent to invoke a load fluctuation prediction tool to predict the task load fluctuation situation based on a trained time series prediction model; using the agent to invoke a node selection tool to select link nodes based on the network topology diagram and the task load fluctuation situation; using the agent to invoke a task scheduling plan generation tool to select a task scheduling method based on the task type and user preferences, and generating a task scheduling plan based on the task description and the selected link nodes using the selected task scheduling method. Here, the task characteristics include the resource requirements of the task, the execution time requirement of the task, the priority of the task, the interruptibility of the task, and the dependency relationship of the task. According to an embodiment of the present invention, any multiple of the acquisition module 810, the planning module 820, and the generation module 830 may be combined and implemented in one module, or any one of them may be split into multiple modules. Or, at least part of the functions of one or more of these modules may be combined with at least part of the functions of other modules and implemented in one module. According to an embodiment of the present invention, at least one of the acquisition module 810, the planning module 820, and the generation module 830 may be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or any other reasonable way of integrating or packaging circuits, etc., implemented by hardware or firmware, or implemented in any one of the three implementation manners of software, hardware, and firmware, or in an appropriate combination of any several of them. Or, at least one of the acquisition module 810, the planning module 820, and the generation module 830 may be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding functions may be executed.

[0091] Figure 4 Schematically shows a block diagram of an electronic device for an agent-based cloud task scheduling method according to an embodiment of the present invention.

[0092] As Figure 4As shown, the electronic device 900 according to an embodiment of the present invention includes a processor 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage section 908 into a random access memory (RAM) 903. The processor 901 may include, for example, a general-purpose microprocessor (e.g., CPU), an instruction set processor, and / or a related chipset, and / or a dedicated microprocessor (e.g., an application specific integrated circuit (ASIC)), etc. The processor 901 may also include on-board memory for caching purposes. The processor 901 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0093] In the RAM 903, various programs and data required for the operation of the electronic device 900 are stored. The processor 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. The processor 901 performs various operations of the method flow according to an embodiment of the present invention by executing the program in the ROM 902 and / or the RAM 903. It should be noted that the program may also be stored in one or more memories other than the ROM 902 and the RAM 903. The processor 901 may also perform various operations of the method flow according to an embodiment of the present invention by executing the program stored in one or more memories.

[0094] According to an embodiment of the present invention, the electronic device 900 may further include an input / output (I / O) interface 905, and the input / output (I / O) interface 905 is also connected to the bus 904. The electronic device 900 may further include one or more of the following components connected to the input / output (I / O) interface 905: an input section 906 including a keyboard, a mouse, etc.; an output section 907 including, for example, a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, a modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the input / output (I / O) interface 905 as needed. A removable medium 911, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 910 as needed so that a computer program read from it can be installed into the storage section 908 as needed.

[0095] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or may exist alone without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, and when the one or more programs are executed, the method according to the embodiments of the present invention is implemented.

[0096] According to an embodiment of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium, and may include, for example, but not limited to: portable computer disks, hard disks, random access memories (RAMs), read-only memories (ROMs), erasable programmable read-only memories (EPROMs or flash memories), portable compact disk read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present invention, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to an embodiment of the present invention, the computer-readable storage medium may include the ROM 902 and / or the RAM 903 described above and / or one or more memories other than the ROM 902 and the RAM 903.

[0097] Those skilled in the art can understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, without departing from the spirit and teachings of the present invention, the features described in the various embodiments of the present invention can be combined and / or combined in various ways. All such combinations and / or combinations fall within the scope of the present invention.

Claims

1. An agent-based cloud task scheduling method, characterized in that, Including: In response to a task request, obtain task characteristics and user preferences; Based on the task characteristics and the user preferences, use an agent to generate a task scheduling action plan, where the agent includes a trained large model, a configuration module, a planning module, a memory module, and a variety of tools, and the tools include a task description generation tool, a node information acquisition tool, a load fluctuation prediction tool, a node selection tool, a task scheduling scheme generation tool, and a task execution tool; According to the action plan, use the agent to sequentially call each tool to generate a task scheduling scheme. Among them, the step of using the agent to sequentially call each tool according to the action plan to generate a task scheduling scheme includes: Based on the task characteristics and the user preferences, use the agent to call the task description generation tool to judge the task type, and integrate the task characteristics, the user preferences, and the task type, and generate a task description in a unified data format; Use the agent to call the node information acquisition tool to collect resource costs, resource usage information, and link condition information, and generate a network topology graph based on graph theory and visualization technology; Use the agent to call the load fluctuation prediction tool to predict the task load fluctuation situation based on the trained time series prediction model; Use the agent to call the node selection tool to select link nodes based on the network topology graph and the task load fluctuation situation; Use the agent to call the task scheduling scheme generation tool, select a task scheduling method based on the task type and user preferences, and generate a task scheduling scheme based on the task description and the selected link nodes using the selected task scheduling method. Among them, the task characteristics include the resource requirements of the task, the execution time requirements of the task, the priority of the task, the interruptibility of the task, and the dependency relationship of the task.

2. The method according to claim 1, characterized in that, The large model includes a large model applicable to natural language processing and multi-modal information processing. The method for training the large model includes: Obtain training data, where the training data includes data of various task characteristics; Use the training data to train the large model to adjust the parameters of the large model and complete the training of the large model. Among them, the large model can process multi-modal input data, and the input data includes data in text, time series, and image modalities.

3. The method according to claim 1, characterized in that The time series prediction model includes a long short-term memory network. The method for training the time series prediction model includes: Obtain historical task load data and mark the time points of key events, where the task load data includes traffic data and hardware status data, and the key events include traffic peaks and hardware failures; Use the historical task load data to train the time series prediction model and adjust the time series prediction model so that the time series prediction model can predict the task load fluctuation situation.

4. The method according to claim 1, wherein The task scheduling methods include a reinforcement learning scheduling method, a heuristic algorithm scheduling method, a meta-heuristic algorithm scheduling method, and a rule-based scheduling algorithm. Among them, the method for training the reinforcement learning scheduling method includes: Based on the reinforcement learning scheduling method, a reward function is set, and the reward function is used to reflect the quality of the scheduling strategy; A state space and an action space are designed. The state space includes the resource requirements of tasks, whether they can be interrupted, and the resource status of the current cloud environment. The action space includes different scheduling decisions; Based on the reward function, the state space, and the action space, using a reinforcement learning algorithm, the reinforcement learning scheduling method is trained in a simulated cloud task environment so that the reinforcement learning scheduling method learns the optimal scheduling strategy; and / or, The method for training the heuristic algorithm scheduling method includes: Based on the characteristics of the tasks and the characteristics of the cloud environment, heuristic rules are set; Based on actual task data, the heuristic algorithm is run using the heuristic algorithm scheduling method to optimize the parameters and rules of the heuristic algorithm; and / or, The meta-heuristic algorithm scheduling method includes a genetic algorithm model, and the training method of the genetic algorithm model includes: Defining a chromosome encoding method and encoding the scheduling scheme as a chromosome; Setting a fitness function to evaluate the quality of the scheduling scheme corresponding to each chromosome; Through genetic operations such as selection, crossover, and mutation, the chromosomes are continuously evolved to obtain a better scheduling scheme, thereby training a genetic algorithm model applicable to cloud task scheduling; and / or, The method for training the rule-based scheduling algorithm includes: Formulating scheduling rules according to domain knowledge and scheduling requirements; Applying the scheduling rules in an actual task scenario and continuously adjusting and improving the scheduling rules according to the feedback.

5. The method according to claim 1, characterized in that, The nodes in the network topology diagram represent cloud resource nodes, and the edges represent the links between the nodes; The network topology diagram includes the bandwidth, delay, and packet loss rate of the links; Adopting a dynamic update mechanism to ensure that the network topology diagram reflects the actual state of the cloud environment in real time.

6. The method according to claim 1, wherein Based on the network topology diagram and the task load fluctuation situation, the selection of link nodes includes: Based on the network topology diagram, using the trained large model to analyze the bandwidth and delay information of each link; Based on the bandwidth and delay information, identifying the key information of the link, including reliability and transmission efficiency; Based on the resource cost and the task load fluctuation situation, obtaining the initial link nodes; Based on the characteristics of the tasks, further screening the initial link nodes to obtain the final link nodes.

7. The method according to claim 1, characterized in that It also includes: Allocating the task scheduling scheme to the selected cloud nodes for execution; Continuously monitoring the execution status of the task. Among them, when an abnormal situation occurs during the task execution, corresponding countermeasures are taken, including: recording the abnormal information; selecting a corresponding processing method according to the severity and type of the abnormality; and feeding back the abnormal information to the agent to adjust and optimize the task scheduling scheme.

8. An agent-based cloud task scheduling device, characterized in that, The device includes: An acquisition module for acquiring task characteristics and user preferences in response to a task request; A planning module, configured to generate a task scheduling action plan by using an agent based on the task characteristics and the user preferences, wherein the agent includes a trained large model, a configuration module, a planning module, a memory module, and multiple tools, and the tools include a task description generation tool, a node information acquisition tool, a load fluctuation prediction tool, a node selection tool, a task scheduling scheme generation tool, and a task execution tool; A generation module, configured to generate a task scheduling scheme by using the agent to sequentially call each tool according to the action plan; wherein, the generating a task scheduling scheme by using the agent to sequentially call each tool according to the action plan includes: Based on the task characteristics and the user preferences, using the agent to call the task description generation tool to determine the task type, and integrating the task characteristics, the user preferences, and the task type, and generating a task description in a unified data format; Using the agent to call the node information acquisition tool to collect resource cost, resource usage information, and link condition information, and generating a network topology graph based on graph theory and visualization technology; Using the agent to call the load fluctuation prediction tool to predict the task load fluctuation condition based on a trained time series prediction model; Using the agent to call the node selection tool to select link nodes based on the network topology graph and the task load fluctuation condition; Using the agent to call the task scheduling scheme generation tool to select a task scheduling method based on the task type and the user preferences, and generating a task scheduling scheme by using the selected task scheduling method based on the task description and the selected link nodes; wherein, the task characteristics include the resource requirements of the task, the execution time requirement of the task, the priority of the task, the interruptibility of the task, and the dependency relationship of the task.

9. An electronic device, characterized in that, including: One or more processors; A memory, configured to store one or more computer programs; The one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, The computer program or instruction, when executed by the processor, implements the steps of the method according to any one of claims 1 to 7.

Images