Task hierarchical scheduling method for intelligent computing fusion network

Abstract

The invention discloses a task hierarchical scheduling method for an intelligent computing fusion network, belongs to the technical field of the intelligent computing fusion network, and designs a double-layer task scheduling mechanism (ICCN-DSF) for the intelligent computing fusion network to effectively cope with task scheduling challenges in a large-scale heterogeneous resource environment. Wherein the global task scheduling layer is based on a CIEK-Means algorithm, integrates geographic position constraint and group intelligent optimization, and realizes efficient clustering and dynamic matching of resource ethnic groups; and the local task scheduling layer constructs a three-dimensional state action space and a multi-target reward function through a reinforcement learning algorithm to complete fine-grained resource allocation. According to the method, the scheduling complexity is remarkably reduced, meanwhile, the cross-domain connection requirement is reduced through a layering mechanism, global resource coordination can be achieved only through a small number of WAN links, the operation and maintenance cost and the safety risk are reduced, meanwhile, the cooperation potential of the computing power and the network is fully excavated, and efficient and robust scheduling support is provided for intelligent services in complex scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent computing fusion networks, and in particular relates to a task hierarchical scheduling method for intelligent computing fusion networks. Background Art

[0002] At this critical juncture of technological revolution and industrial transformation, computing power has become a core driver of scientific and technological progress and economic growth. With the widespread adoption of cutting-edge technologies such as artificial intelligence, virtual reality, general-purpose large-scale models, and the metaverse, market demand for computing power is rapidly increasing. The application of these technologies has not only triggered an explosive growth in demand for computing power, but also increasingly diversified, distributed, and complex demand. However, the traditional single-point computing model is no longer able to meet this growing demand. Therefore, how to efficiently utilize distributed computing resources and improve computing power supply has become a critical issue that needs to be addressed.

[0003] Against this backdrop, the deep integration of computing power and networks has gradually become a core technological direction for addressing these challenges. This fusion not only optimizes resource allocation but also improves the efficiency and quality of computing services through intelligent scheduling and efficient management. Driven by national top-level design, industry needs, and technological advancements, the concept of "intelligent computing converged network" has emerged and gradually become a research focus in this field. This concept aims to overcome the limitations of traditional computing power supply, which is constrained by geographical and network boundaries, by deeply integrating computing and network technologies, enabling intelligent computing services across regions and platforms. Leveraging an intelligent drive mechanism, this network architecture enables real-time and flexible scheduling based on dynamic changes in service demand, providing convenient and efficient computing resources.

[0004] However, unlike traditional resources like water and electricity, computing power itself cannot flow freely across regions. In reality, "flow" refers to the transfer of elements such as service requests, data, and functions. Task scheduling is particularly important in this process. Task scheduling not only involves the allocation and optimization of computing power but also comprehensively considers multiple factors, including service demand priorities, dynamic resource changes, and network load. Efficient task scheduling maximizes computing power utilization, ensures intelligent service adaptation, and guarantees system stability and reliability under high load and high concurrency.

[0005] In summary, research on intelligent computing converged networks has promoted the deep integration and innovative development of computing and network technologies, laying a solid foundation for the widespread application of intelligent technologies. By optimizing the scheduling and management of computing power and network resources, intelligent computing converged networks will play a vital role in the rapid development of the digital economy and become critical infrastructure for the development of the information and communications sector. With continuous breakthroughs in related technologies and the continued optimization of the policy environment, intelligent computing converged networks will further promote the efficient allocation of computing power resources and the widespread application of intelligent services, providing a key guarantee for building an intelligent and efficient digital society.

[0006] Task scheduling algorithms are crucial in modern computing systems, especially in complex environments such as large-scale distributed systems, cloud computing, and intelligent computing convergence networks. Their primary goal is to efficiently allocate limited computing, storage, and bandwidth resources to improve task execution efficiency and overall system performance. Traditional task scheduling methods are often based on heuristic strategies, such as optimization-based dynamic programming and meta-heuristic algorithms.

[0007] Traditional heuristic scheduling algorithms often exhibit significant limitations when faced with the dynamic, highly heterogeneous nature of intelligent computing converged networks. First, these algorithms typically have high computational complexity and cannot quickly respond to the demands of large-scale tasks and networks. Second, traditional algorithms fail to adequately consider the interactions of multi-dimensional resources within the network. They often rely on fixed rules and lack adaptability, resulting in low resource utilization and an inability to efficiently respond to rapidly changing network conditions. Summary of the Invention

[0008] In response to the above-mentioned deficiencies in the prior art, the hierarchical task scheduling method for intelligent computing fusion networks provided by the present invention solves the problems of high computational complexity of traditional heuristic task scheduling methods, inability to quickly respond to the needs of large-scale tasks and networks, low resource utilization, and inability to efficiently cope with rapidly changing network conditions.

[0009] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a task hierarchical scheduling method for intelligent computing fusion network, which adopts a two-layer task scheduling mechanism, including: Global task scheduling layer: Uses the CIEK-Means algorithm to divide resource clusters in the intelligent computing fusion network to form logically independent resource cluster areas; Generate a region ID for each resource group region and generate a region upgrade table; When the task is completed, it will be assigned to the most matching resource group area based on the comprehensive distance between it and each resource group area and the current area upgrade table; Local task scheduling layer: prioritizes tasks assigned to resource clusters, and selects a scheduling algorithm based on task priorities and the current resource cluster scheduling scenario to assign tasks in sequence. In the local task scheduling layer, a scheduling algorithm based on reinforcement learning is adopted to gradually optimize the task allocation strategy through continuous interaction with the actual task load and operating environment.

[0010] Furthermore, the method of using the CIEK-Means algorithm to divide resource clusters in the intelligent computing fusion network to form logically independent resource cluster areas includes the following steps: S1. Normalize the dataset of resource groups; S2. Generate a candidate group based on the normalized resource clusters; each candidate represents a centroid selection scheme for a resource cluster region; S3. Apply the Geo-KMeans algorithm to each candidate and calculate its objective function value; S4. Perform multi-point adaptive mutation operation on each candidate solution to generate several mutation points; S5. Select the optimal solution as the candidate new solution according to the objective function value of each variation point; S6. Generate a random number for each candidate and select a behavior to follow through the roulette wheel method; S7, after each candidate selects a follow-up behavior, update the sampling interval of its solution space; S8. Generate several samples within the updated sampling interval, calculate their objective function values, and select the behavior with the best performance as the new optimal solution; S9. Determine whether the candidate group's behavior is saturated; If yes, go to step S10; If not, go to step S11; S10, restore the sampling interval of each candidate to the initial range, and proceed to step S11; S11, judging whether the CIEK-Means algorithm converges; If so, the current optimal solution is taken as the final solution and divided into several resource group areas; If not, return to step S3.

[0011] Furthermore, in step S2, the candidate group Expressed as: Where, represents the cth candidate, represents the j-th cluster center vector of the c-th candidate, represents the value of the d-th resource dimension of the j-th cluster of the c-th candidate, represents the number of candidates, is the number of cluster centers specified, represents the resource dimension of each centroid; In step S3, the objective function value The calculation formula is: Where, represents the cluster center, Indicates the ethnic group attributes; In step S4, a variation point is generated. The formula is: Where, represents the solution of the i-th candidate, represents the current optimal solution, represents a randomly selected solution, and represents the coefficient of variation during adaptive mutation operation, and Respectively and The corresponding initial coefficient of variation, Indicates the current iteration number, Indicates the maximum number of iterations; In step S5, the formula for selecting the optimal solution is: Where, represents the objective function; In step S7, the updated sampling interval Expressed as: Where, represents the sampling interval before updating, represents the sampling interval reduction factor; In step S9, when the behavior of the candidate group satisfies the following formula, it is determined that saturation is reached; Where, represents the objective function value of the optimal solution in the candidate population at the sth iteration, represents the objective function value of the worst solution in the candidate population at the sth iteration, Indicates the saturation judgment value.

[0012] Furthermore, based on the characteristic values ​​of the resource group region in each resource dimension, a corresponding region identifier is generated; Among them, the characteristic value Expressed as: Where, Indicates the minimum value of the resource group area in the jth resource dimension, represents the cluster center, Indicates the offset ratio; Adjusting the position of the area marker by adjusting the offset ratio; When the offset ratio , the area marker is located at the edge; when the offset ratio , the regional logo is located in the center of the resource group area; when the offset ratio , the region identification is close to the centroid.

[0013] Furthermore, the comprehensive distance between the task and the resource group area for: Where, represents the resource requirement vector of the task, The region identification vector representing the resource group region, and Represents the geographical coordinates of the task and resource group areas, is the weight coefficient of the geographical location.

[0014] Furthermore, the global task scheduling layer also includes: Dynamically monitor the resource utilization of each resource group area. When the resource utilization exceeds the preset threshold, assign tasks to the upper-level resource group area according to the area upgrade table. After the task allocation is completed, the load information of the resource group area is updated. The method for generating the regional upgrade table is: Select resource cluster regions with a value greater than −β% in each resource dimension and a value greater than γ% in at least one dimension; Sort the selected resource group regions by Euclidean distance to obtain a regional upgrade table; Among them, β and γ represent the basic parameters of the regional upgrade table.

[0015] Furthermore, at the local task scheduling layer, within each resource group area, tasks are prioritized according to their urgency; Among them, the priority of task i Expressed as: Where, Indicates the grace of the task, Indicates the current time, Indicates the deadline of the task, Indicates the task execution requirements, represents the computational requirements of task i, represents the computing power of task i assigned to the computing node, represents the data transmission requirement of task i, represents the bandwidth capacity of the communication link allocated to task i, Represents the life cycle of a task, Indicates the arrival time of the task, and Indicates that the priority is being adjusted. and The weight of .

[0016] Furthermore, in the local task scheduling layer, when the current resource group region cannot meet the task requirements, the task is scheduled to the upper-level resource group region according to the region upgrade table in the global task scheduling layer.

[0017] Furthermore, when using a scheduling algorithm based on reinforcement learning to optimize the task allocation strategy, the cluster state in the state space and the action output by the agent in the action space are represented by three-dimensional matrices respectively; In the state space, the three-dimensional matrix of the cluster state at any time is expressed as , a three-dimensional matrix Each element in represents a machine Specific resource usage on the resource attribute dimension h; where the height H of the three-dimensional matrix represents the resource attribute dimension, and the length L and width W of the three-dimensional matrix are mapped to the machine ID through two dimensions; In the action space, the three-dimensional matrix of the action output by the agent is represented as , a three-dimensional matrix Each element task Indicates that task i is assigned to machine , I represents the number of tasks, J and K represent the two dimension indexes of the machine respectively.

[0018] Furthermore, when using a scheduling algorithm based on reinforcement learning to optimize the task allocation strategy, the reward function for: Where, represents the reward for task response time, A reward representing the load balancing of the machines, and Respectively and Adjustable weight parameters of ; Among them, the reward for task i’s response time is and machine load balancing bonus Respectively expressed as: Where, represents the execution time of task i in the current round, represents the actual response time of task i in the current round, represents the expected response time of task i in the current round, Indicates the total number of machines, Indicates the load of the machine, Indicates the average load of all machines The beneficial effects of the present invention are: The method of the present invention effectively addresses the challenges of task scheduling in large-scale heterogeneous resource environments by designing a two-tier task scheduling mechanism (ICCN-DSF) for intelligent computing converged networks. The global task scheduling layer integrates geographic location constraints and swarm intelligence optimization based on the CIEK-Means algorithm to achieve efficient clustering and dynamic matching of resource clusters. The local task scheduling layer constructs a three-dimensional state-action space and a multi-objective reward function through a reinforcement learning algorithm to achieve fine-grained resource allocation. This framework significantly reduces scheduling complexity, shortening task completion time by an average of 32% and improving resource utilization by 29%. At the same time, a layered mechanism reduces cross-domain connectivity requirements, enabling global resource coordination with only a small number of WAN links. This framework fully exploits the synergistic potential of computing power and networks while reducing operational costs and security risks, providing efficient and robust scheduling support for intelligent services in complex scenarios.

[0019] Specifically, compared with the traditional resource scheduling framework, the technical solution of the present invention has the following advantages: 1. Reduce scheduling complexity: Through the hierarchical scheduling mechanism, the complex global task scheduling problem is split into multiple smaller local task scheduling problems, significantly reducing the overall scheduling complexity.

[0020] 2. Improve resource utilization: Resource clustering and dynamic matching at the global task scheduling layer, as well as fine-grained task allocation at the local task scheduling layer, effectively improve resource utilization efficiency.

[0021] 3. Enhance system robustness and adaptability: The system can flexibly adjust scheduling strategies according to the dynamic changes of resources and adjustments to the network topology, thereby enhancing the robustness and adaptability of the system in the face of changing environments.

[0022] 4. Optimize task execution performance: Through the optimization of reinforcement learning algorithms, the task response time is significantly shortened, balanced resource utilization is achieved, and the overall performance of task execution is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Flowchart of the task hierarchical scheduling method for intelligent computing fusion network provided by the present invention.

[0024] Figure 2 This is a structural diagram of the task hierarchical scheduling framework provided by the present invention.

[0025] Figure 3 This is a diagram of the global task scheduling layer architecture provided by the present invention.

[0026] Figure 4 This is a schematic diagram of the impact of the offset ratio on matching accuracy provided by the present invention.

[0027] Figure 5 This is a diagram of the local task scheduling layer architecture provided by the present invention.

[0028] Figure 6 Schematic diagram of cluster state and action space design provided by the present invention. DETAILED DESCRIPTION

[0029] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0030] The embodiment of the present invention provides a task hierarchical scheduling method for intelligent computing fusion network, such as Figure 1 As shown, a two-layer task scheduling mechanism is adopted, including: Global task scheduling layer: Uses the CIEK-Means algorithm to divide resource clusters in the intelligent computing fusion network to form logically independent resource cluster areas; Generate a region ID for each resource group region and generate a region upgrade table; When the task is completed, it will be assigned to the most matching resource group area based on the comprehensive distance between it and each resource group area and the current area upgrade table; Local task scheduling layer: prioritizes tasks assigned to resource clusters, and selects a scheduling algorithm based on task priorities and the current resource cluster scheduling scenario to assign tasks in sequence. In the local task scheduling layer, a scheduling algorithm based on reinforcement learning is adopted to gradually optimize the task allocation strategy through continuous interaction with the actual task load and operating environment.

[0031] In the actual intelligent computing fusion network scenario, when the number of user requests and resource groups is extremely large, in order to quickly find the appropriate scheduling group in the huge solution space and realize global task scheduling and local task scheduling, such as Figure 2 As shown, the present invention proposes a scheduling architecture with dual-layer intelligent computing tasks based on the above method, and realizes resource management and task allocation in the intelligent computing fusion network through the design of an efficient dual-layer scheduling framework ICCN-DSF (Intelligent Computing Convergence Network Dual-layerScheduling Framework).

[0032] In this embodiment of the present invention, the ICCN-DSF's global task scheduling layer is responsible for scheduling task allocation and computing resource coordination across the entire intelligent computing convergence network. It is a key link in achieving efficient task distribution and maximizing resource utilization. Its primary function is to allocate tasks to appropriate resource clusters based on their resource requirements, thereby optimizing resource allocation and improving overall scheduling performance. This stage comprehensively considers factors such as the geographic distribution, type, and capacity of resources to ensure that resources are utilized rationally while meeting task requirements.

[0033] In this embodiment, Figure 3 In the structure of the global task scheduling layer shown, the CIEK-Means algorithm is first used to partition the resource clusters within the intelligent computing fusion network, forming logically independent resource cluster regions, laying the foundation for subsequent task scheduling. These resulting resource cluster regions are not only geographically concentrated but also relatively balanced in terms of resource dimensions such as computing, storage, and bandwidth. Through this initial division, the system establishes a clear resource hierarchy, which helps improve the accuracy of task allocation and reduces the range of resources that need to be traversed during scheduling, thereby narrowing the resource space and improving scheduling efficiency. After the resource cluster division is completed, the global task scheduling layer generates a unique identifier for each cluster region for subsequent task matching. When a task arrives at the task pool, the system calculates the combined distance between it and the identifiers of each cluster region based on the task's resource requirements and geographic location, ensuring that the task is assigned to the resource cluster region with the best matching location and resource characteristics.

[0034] Specifically, in this embodiment, a cross-domain, large-scale, geographically-scaled intelligent computing fusion network is developed, which integrates multi-cloud, multi-edge, and multi-terminal devices to meet the needs of tasks of different types and scales. The resources of this network are not only widely distributed, but also show obvious multi-dimensional and highly heterogeneous characteristics. Various resources, including computing, communication, and storage resources, vary greatly in different regions and physical locations. This results in a huge solution space for both task scheduling and resource allocation, which faces the challenge of spatial explosion in the decision-making process.

[0035] In order to solve the scheduling problem in a hierarchical manner under the intelligent computing fusion network, the present invention proposes an innovative computing resource clustering algorithm - Cohort Intelligence Enhanced K-Means (CIEK-Means) algorithm. CIEK-Means combines the characteristics and advantages of the K-means algorithm and the Enhanced CI algorithm to form a new hybrid algorithm - Cohort Intelligence Enhanced K-Means (CIEK-Means) algorithm. This algorithm combines the spatial constraint clustering capability of Geo-KMeans and the optimization mechanism of the EnhancedCI algorithm, and improves the efficiency and accuracy of resource allocation by considering multi-factor clustering of geographical location and resource attributes. At the same time, the multi-point adaptive mutation operation of the Enhanced CI algorithm enables the CIEK-Means algorithm to not only overcome the problem that the K-means algorithm is prone to falling into local optimal solutions, but also significantly improves the convergence speed of the algorithm.

[0036] In this embodiment, the method for dividing resource groups using the CIEK-Means algorithm includes the following steps: S1. Normalize the dataset of resource groups to remove the dimensional differences between different attributes; Assume that the dataset of a resource group is ,in , indicating the ethnic group attributes; the normalization formula is: in, Represents the normalized attribute value, Indicates the This normalization process ensures that resources in each dimension have equal influence in the clustering process, laying the foundation for subsequent weighted Euclidean distance calculation and geographic distance constraints. S2. Generate a candidate group based on the normalized resource clusters; each candidate represents a centroid selection scheme for a resource cluster region; Among them, the candidate group Expressed as: Where, represents the cth candidate, represents the j-th cluster center vector of the c-th candidate, represents the value of the d-th resource dimension of the j-th cluster of the c-th candidate, represents the number of candidates, is the number of cluster centers specified, represents the resource dimension of each centroid; S3. Apply the Geo-KMeans algorithm to each candidate and calculate its objective function value; Among them, the objective function value The calculation formula is: Where, represents the cluster center, Indicates the ethnic group attributes; S4. Perform multi-point adaptive mutation operation on each candidate solution to generate several mutation points; Among them, the generated mutation point The formula is: Where, represents the solution of the i-th candidate, represents the current optimal solution, represents a randomly selected solution, and represents the coefficient of variation during adaptive mutation operation, and Respectively and The corresponding initial coefficient of variation, Indicates the current iteration number, Indicates the maximum number of iterations; S5. According to each variation point The objective function value of , select the optimal solution as the candidate new solution; Among them, the formula for selecting the optimal solution is: Where, represents the objective function; S6. Generate a random number for each candidate and select a behavior to follow through the roulette wheel method; Behavior refers to the candidate's search strategy in the solution space (such as mutation, crossover, and other operation modes). Roulette is used to select high-quality behaviors (search modes with high fitness) according to probability, simulating group evolution, guiding the solution to update towards the optimal solution, and improving clustering efficiency. S7. After each candidate selects a follow-up behavior, update the sampling interval of its solution space; Among them, the sampling interval is the candidate's exploration range in the solution space (such as the range of the center of mass coordinates); Among them, the updated sampling interval Expressed as: Where, represents the sampling interval before updating, represents the sampling interval reduction factor; S8. Generate several samples within the updated sampling interval, calculate their objective function values, and select the best-performing behavior and set it as the new optimal solution. ; Here, sample refers to the data in the cluster dataset; S9. Determine whether the candidate group's behavior is saturated; If yes, go to step S10; If not, go to step S11; Among them, when the behavior of the candidate group satisfies the following formula, the judgment reaches saturation; Where, represents the objective function value of the optimal solution in the candidate population at the sth iteration, represents the objective function value of the worst solution in the candidate population at the sth iteration, Indicates the saturation judgment value.

[0037] S10, restore the sampling interval of each candidate to the initial range to re-explore the potential optimal solution, and proceed to step S11; S11, judging whether the CIEK-Means algorithm converges; If so, the current optimal solution is taken as the final solution and divided into several resource group areas; If not, return to step S3; Specifically, at the end of each iteration, the CIEK-Means algorithm checks convergence. If the population stabilizes within the specified number of iterations and the target value range narrows to within the threshold, or if there is no significant improvement after the maximum number of iterations, the algorithm converges and the current solution is accepted as the final solution. Otherwise, the iteration continues.

[0038] In this embodiment, the CIEK-Means algorithm demonstrates significant advantages in resource allocation in intelligent computing fusion networks. During algorithm execution, Geo-KMeans clustering fully considers the geographical distribution properties of resources, making the algorithm more reasonable and flexible when dealing with the complex resource topology of intelligent computing fusion networks; the introduction of Enhanced CI makes the clustering results tend to the global optimum after each iteration, improving the accuracy of resource scheduling. Overall, the CIEK-Means algorithm takes into account both spatial constraints and resource heterogeneity, has high practical value in intelligent computing fusion networks, and helps to complete task scheduling efficiently and stably in large-scale complex resource environments in the future.

[0039] In this embodiment, after the resource group division is completed, a unique region identifier is generated for each group area; specifically, the corresponding region identifier is generated based on the characteristic values ​​of the resource group area in each resource dimension; Among them, the characteristic value Expressed as: Where, Indicates the minimum value of the resource group area in the jth resource dimension, represents the cluster center, Indicates the offset ratio.

[0040] In this embodiment, for the offset ratio: When the offset ratio ,The regional identifier is located at the edge, which results in resource capacity being much larger than the task requirements, and inaccurate matching; When the offset ratio ,The regional identifier is located in the center of the resource group region, but the task requirements exceed the group resource range, resulting in ,sufficient resources; When the offset ratio , the regional identification is close to the centroid, and the resource capacity is better matched with the task requirements, achieving the best matching effect.

[0041] In this embodiment, if Figure 4 As shown, the influence of the offset ratio on the accuracy of task allocation is demonstrated, so the accuracy and efficiency of resource matching can be improved by adjusting the offset ratio and the position of the area identifier.

[0042] When a task arrives at the task pool, the system will calculate the comprehensive distance between it and the ethnic area identifiers based on the task's resource requirements and geographical location to ensure that the task is assigned to the ethnic area with the best match in geographical location and resource characteristics.

[0043] In this embodiment, the comprehensive distance between the task and the resource group area for: Where, represents the resource requirement vector of the task, The region identification vector representing the resource group region, and Represents the geographical coordinates of the task and resource group areas, This is the geographic weight coefficient, used to adjust the importance of location in distance calculations. By incorporating the geographic weight into the comprehensive distance calculation, we can prioritize geographically close communities for task allocation while meeting task resource requirements, thus avoiding the high latency and resource waste that can result from cross-region scheduling.

[0044] In an embodiment of the present invention, the global task scheduling layer further includes: Dynamically monitor the resource utilization of each resource group area. When the resource utilization exceeds the preset threshold, assign tasks to the upper-level resource group area according to the area upgrade table. After the task allocation is completed, the load information of the resource group area is updated. Among them, resource utilization rate for: in, Indicates the Resource groups in the The usage of each resource dimension, represents the capacity on this dimension, Indicates the number of resource groups in the group area, Indicates the number of resource dimensions.

[0045] In this embodiment, by monitoring resource utilization, the system can monitor resource usage in real time and, when necessary, limit the utilization of cluster regions to ensure load balancing. If the resource utilization of a cluster region exceeds a preset threshold, the system will assign tasks to the next higher cluster region according to the preset region upgrade table.

[0046] In this embodiment, the method for generating the regional upgrade table is: Select resource cluster regions with a value greater than −β% in each resource dimension and a value greater than γ% in at least one dimension; Sort the selected resource group regions by Euclidean distance to obtain a regional upgrade table; Among them, β and γ represent the basic parameters of the regional upgrade table, which are respectively expressed as: In this embodiment, the above-mentioned regional upgrade table enables the system to quickly find a suitable alternative ethnic group region to take over the task when the resources of the current ethnic group region are insufficient, thereby ensuring the continuity and reliability of scheduling.

[0047] After the task allocation is completed, the global task scheduling layer will update the load information of the group area, including resource usage, remaining capacity, etc., to ensure the system's real-time control of resource status and facilitate dynamic adjustments in subsequent task scheduling.

[0048] In this embodiment, the above-mentioned "divide first, then schedule" strategy is adopted at the global task scheduling layer, which not only improves resource utilization efficiency, but also ensures that the intelligent computing fusion network maintains efficient task scheduling and load balancing in a complex and changing environment.

[0049] In the embodiment of the present invention, Figure 5 Figure 1 shows the overall architecture of the local task scheduling layer. After assigning tasks to cluster regions, they are first prioritized based on their urgency. After the task prioritization is completed, the local task scheduler selects the most appropriate scheduling algorithm for task allocation based on the scheduling context within the current region.

[0050] In this embodiment, tasks are prioritized within each resource group based on their urgency. Specifically, task urgency is measured using two factors: the task's grace period and the task's lifecycle. The task's grace period represents the maximum amount of time a task can be delayed without violating its deadline. It is calculated as the sum of the task's deadline minus the current time and the task's estimated execution time. The task's lifecycle reflects the time a task has existed in the system and is often used to measure task urgency.

[0051] Based on this, in this embodiment, the priority of task i is Expressed as: Where, Indicates the grace of the task, Indicates the current time, Indicates the deadline of the task, Indicates the task execution requirements, represents the computational requirements of task i, represents the computing power of task i assigned to the computing node, represents the data transmission requirement of task i, represents the bandwidth capacity of the communication link allocated to task i, Represents the life cycle of a task, Indicates the arrival time of the task, and Indicates that the priority is being adjusted. and The weight of .

[0052] In this embodiment, task prioritization is performed using both the grace and lifecycle of tasks, aiming to comprehensively assess the urgency and importance of tasks. The grace measures the maximum amount of time a task can be delayed without exceeding its deadline, providing scheduling flexibility. The lifecycle reflects the waiting time of a task in the system, ensuring that long-outstanding tasks are prioritized and avoid long backlogs.

[0053] In this embodiment, when selecting a scheduling algorithm for task allocation, due to the high compatibility of the framework, the scheduling system can flexibly adopt different algorithms, such as heuristic algorithms, metaheuristic algorithms or reinforcement learning, to cope with the heterogeneity of resources and the dynamic changes in task requirements, thereby optimizing the overall system performance.

[0054] In this embodiment, in the local task scheduling layer, when the current resource group area cannot meet the task requirements, the task is scheduled to the resource group area of ​​the upper level according to the area upgrade table in the global task scheduling layer to flexibly deal with the situation of insufficient resources and ensure timely processing of tasks.

[0055] In embodiments of the present invention, at the local task scheduling layer, traditional scheduling algorithms typically rely on static assumptions and predefined rules, making them difficult to adapt to complex systems with dynamic changes and heterogeneous resources. This invention utilizes a scheduling algorithm based on reinforcement learning to optimize task allocation strategies. In contrast, reinforcement learning algorithms can gradually optimize scheduling strategies through continuous interaction with the actual task load and operating environment, avoiding reliance on inaccurate prior assumptions and exhibiting greater adaptability and generalization capabilities.

[0056] In this embodiment, in the scheduling algorithm based on reinforcement learning, Figure 6 As shown in Figure 2, the cluster state in the state space and the action output by the agent in the action space are represented by three-dimensional matrices respectively.

[0057] Specifically, in the state space, the three-dimensional matrix of the cluster state at any time is expressed as , a three-dimensional matrix Each element in represents a machine Specific resource usage on the resource attribute dimension h; where the height H of the three-dimensional matrix represents the resource attribute dimension, and the length L and width W of the three-dimensional matrix are mapped to the machine ID through two dimensions; This design effectively reduces the single-dimensional representation of machine IDs to a two-dimensional representation, significantly reducing the complexity of the state space and creating a more regular matrix shape, thus avoiding the problem of feature representation imbalance caused by a large aspect ratio. This compact and efficient state representation not only reduces the dimensionality of the input features, further reducing the size of the policy network, but also significantly improves the efficiency and convergence speed of model training. Overall, the three-dimensional matrix representation method effectively optimizes the representation of cluster state, providing more comprehensive and easy-to-process information for the agent's subsequent scheduling decisions.

[0058] In the action space, in order to solve the problem of action space explosion and take into account the correlation between tasks, the three-dimensional matrix of the action output by the agent is represented as , a three-dimensional matrix Each element task Indicates that task i is assigned to machine , I represents the number of tasks, J and K represent the two dimension indexes of the machine respectively.

[0059] The agent generates this probability matrix through a policy network and then determines the final mapping between tasks and machines based on the probability values ​​of each cell in the matrix. Compared to traditional one-dimensional action space designs, this approach reduces the complexity of the action space from nm to n×m by adding a dimension, significantly alleviating the problem of action space expansion. This design not only effectively reduces the size of the action space and eases model training, but also significantly enhances the flexibility and efficiency of parallel scheduling.

[0060] In this embodiment, when using the scheduling algorithm based on reinforcement learning to optimize the task allocation strategy, in order to effectively guide the agent to optimize the task scheduling strategy, a reward function is designed. for: Where, represents the reward for task response time, A reward representing the load balancing of the machines, and Respectively and The adjustable weight parameter of It is used to measure the response efficiency of the current round of scheduled tasks. The core idea is to compare the actual response time of the task with the expected response time, while taking into account the length of the task execution time.

[0061] In this embodiment, the reward for task i’s response time is and machine load balancing bonus Respectively expressed as: Where, represents the execution time of task i in the current round, represents the actual response time of task i in the current round, represents the expected response time of task i in the current round, Indicates the total number of machines, Indicates the load of the machine, Indicates the average load of all machines.

[0062] In this embodiment, in the above-mentioned reward function, the effects of both task response time and machine load balancing are comprehensively considered. By adjusting the values ​​of ω1 and ω2, the reward function can achieve a dynamic balance between task response efficiency and load balancing to meet the optimization requirements in different scenarios. This design enables the reward function to focus on the scheduling results of the current round and can evaluate the effectiveness of the scheduling strategy in real time.

[0063] In this embodiment, in addition to the above-mentioned reinforcement learning method, the Advantage Actor-Critic (A2C) algorithm can also be used to directly optimize the strategy and use the value network to reduce variance, thereby improving sample efficiency and stability while addressing the limitations of traditional methods.

[0064] In an embodiment of the present invention, a specific experimental example of the above-mentioned task hierarchical scheduling method is provided.

[0065] In this experiment, detailed resource clusters and tasks were configured to simulate the diverse resource distribution and task requirements in intelligent computing convergence networks. The experiment established 10,000 resource clusters, each containing three types of resources: computing power, storage, and transmission. Geographic information was assigned to reflect their cross-regional distribution. Task generation used a Poisson distribution to simulate arrival frequency, combining the set arrival rate and resource requirements to generate task sets of varying sizes. The computing power, storage, and transmission requirements of tasks were randomly set within a preset range to reflect the resource heterogeneity in real-world scenarios. Task timeliness requirements were also met by setting duration and reserve time. Specific parameter information is provided in Table 1.

[0066] Table 1: Some parameter information of resource groups and tasks Furthermore, parameter configuration within the ICCN-DSF framework is crucial for optimizing task scheduling and resource allocation. This parameter system encompasses key metrics such as task grace, lifecycle, group-region identifier offset ratio, and geolocation weight coefficient. By controlling task latency, quantifying geolocation impact, and enabling dynamic resource group upgrades, it aims to ensure efficient task completion under diverse conditions, while balancing regional loads and supporting adaptive resource allocation. Detailed parameter settings are detailed in Table 2.

[0067] Table 2: Some parameter information of ICCN-DSF framework Table 2 shows the clustering performance of CIEK-Means and other algorithms on various datasets. Although CIEK-Means requires more iterations, its clustering performance and stability are significantly better than those of K-Means, K-Means++, Improved K-Means, and MHTSASM, particularly in terms of standard deviation. A smaller standard deviation means more stable clustering results, avoiding the risk of prematurely falling into a local optimum. Particularly on large-scale datasets (such as Blobs1 and Blobs2), although CIEK-Means requires a slightly higher number of iterations, it is able to stably converge to a high-quality solution in fewer iterations, demonstrating strong clustering performance and a lower standard deviation, maintaining high clustering quality. In summary, CIEK-Means far outperforms other algorithms in terms of clustering performance, stability, and robustness, making it particularly suitable for clustering GB-like tasks on high-dimensional and large-scale datasets.

[0068] Specific embodiments are used in the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

[0069] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A task hierarchical scheduling method for intelligent computing fusion network, characterized in that: A two-tier task scheduling mechanism is adopted, including: Global task scheduling layer: Uses the CIEK-Means algorithm to divide resource clusters in the intelligent computing fusion network to form logically independent resource cluster areas; Generate a region ID for each resource group region and generate a region upgrade table; When the task is completed, it will be assigned to the most matching resource group area based on the comprehensive distance between it and each resource group area and the current area upgrade table; Local task scheduling layer: prioritizes tasks assigned to resource clusters, and selects a scheduling algorithm based on task priorities and the current resource cluster scheduling scenario to assign tasks in sequence. In the local task scheduling layer, a scheduling algorithm based on reinforcement learning is adopted to gradually optimize the task allocation strategy through continuous interaction with the actual task load and operating environment.

2. The task hierarchical scheduling method for intelligent computing fusion network according to claim 1 is characterized in that: The method for using the CIEK-Means algorithm to divide resource clusters in an intelligent computing fusion network to form logically independent resource cluster areas includes the following steps: S1. Normalize the dataset of resource groups; S2. Generate a candidate group based on the normalized resource clusters; each candidate represents a centroid selection scheme for a resource cluster region; S3. Apply the Geo-KMeans algorithm to each candidate and calculate its objective function value; S4. Perform multi-point adaptive mutation operation on each candidate solution to generate several mutation points; S5. Select the optimal solution as the candidate new solution according to the objective function value of each variation point; S6. Generate a random number for each candidate and select a behavior to follow through the roulette wheel method; S7. After each candidate selects a follow-up behavior, update the sampling interval of its solution space; S8. Generate several samples within the updated sampling interval, calculate their objective function values, and select the behavior with the best performance as the new optimal solution; S9. Determine whether the candidate group's behavior is saturated; If yes, go to step S10; If not, go to step S11; S10, restore the sampling interval of each candidate to the initial range, and proceed to step S11; S11, judging whether the CIEK-Means algorithm converges; If so, the current optimal solution is taken as the final solution and divided into several resource group areas; If not, return to step S3.

3. The task hierarchical scheduling method for intelligent computing fusion network according to claim 2 is characterized in that: In step S2, the candidate group Expressed as: Where, represents the cth candidate, represents the j-th cluster center vector of the c-th candidate, represents the value of the d-th resource dimension of the j-th cluster of the c-th candidate, represents the number of candidates, is the number of cluster centers specified, represents the resource dimension of each centroid; In step S3, the objective function value The calculation formula is: Where, represents the cluster center, Indicates the ethnic group Attributes; In step S4, a variation point is generated. The formula is: Where, represents the solution of the i-th candidate, represents the current optimal solution, represents a randomly selected solution, and represents the coefficient of variation during adaptive mutation operation, and Respectively and The corresponding initial coefficient of variation, Indicates the current iteration number, Indicates the maximum number of iterations; In step S5, the formula for selecting the optimal solution is: Where, represents the objective function; In step S7, the updated sampling interval Expressed as: Where, represents the sampling interval before updating, represents the sampling interval reduction factor; In step S9, when the behavior of the candidate group satisfies the following formula, it is determined that saturation is reached; Where, represents the objective function value of the optimal solution in the candidate population at the sth iteration, represents the objective function value of the worst solution in the candidate population at the sth iteration, Indicates the saturation judgment value.

4. The task hierarchical scheduling method for intelligent computing fusion network according to claim 1 is characterized in that: Generate corresponding regional identifiers based on the characteristic values ​​of resource group regions in each resource dimension; Among them, the characteristic value Expressed as: Where, Indicates the minimum value of the resource group area in the jth resource dimension, represents the cluster center, Indicates the offset ratio; Adjusting the position of the area marker by adjusting the offset ratio; When the offset ratio , the area marker is located at the edge; when the offset ratio , the regional logo is located in the center of the resource group area; when the offset ratio , the region identification is close to the centroid.

5. The task hierarchical scheduling method for intelligent computing fusion network according to claim 1 is characterized in that: The comprehensive distance between the task and the resource group area for: Where, represents the resource requirement vector of the task, The region identification vector representing the resource group region, and Represents the geographical coordinates of the task and resource group areas, is the weight coefficient of the geographical location.

6. The task hierarchical scheduling method for intelligent computing fusion network according to claim 1 is characterized in that: The global task scheduling layer also includes: Dynamically monitor the resource utilization of each resource group area. When the resource utilization exceeds the preset threshold, assign tasks to the upper-level resource group area according to the area upgrade table. After the task allocation is completed, the load information of the resource group area is updated. The method for generating the regional upgrade table is: Select resource cluster regions with a value greater than −β% in each resource dimension and a value greater than γ% in at least one dimension; Sort the selected resource group regions by Euclidean distance to obtain a regional upgrade table; Among them, β and γ represent the basic parameters of the regional upgrade table.

7. The task hierarchical scheduling method for intelligent computing fusion network according to claim 1 is characterized in that: At the local task scheduling layer, within each resource group area, tasks are prioritized based on their urgency; Among them, the priority of task i Expressed as: Where, Indicates the grace of the task, Indicates the current time, Indicates the deadline of the task, Indicates the task execution requirements, represents the computational requirements of task i, represents the computing power of task i assigned to the computing node, represents the data transmission requirement of task i, represents the bandwidth capacity of the communication link allocated to task i, Represents the life cycle of a task, Indicates the arrival time of the task, and Indicates that the priority is being adjusted. and The weight of .

8. The task hierarchical scheduling method for intelligent computing fusion network according to claim 1 is characterized in that: In the local task scheduling layer, when the current resource group region cannot meet the task requirements, the task is scheduled to the upper-level resource group region according to the region upgrade table in the global task scheduling layer.

9. The task hierarchical scheduling method for intelligent computing fusion network according to claim 1 is characterized in that: When optimizing the task allocation strategy using a scheduling algorithm based on reinforcement learning, the cluster state in the state space and the action output by the agent in the action space are represented by three-dimensional matrices respectively; In the state space, the three-dimensional matrix of the cluster state at any time is expressed as , a three-dimensional matrix Each element in represents a machine Specific resource usage on the resource attribute dimension h; where the height H of the three-dimensional matrix represents the resource attribute dimension, and the length L and width W of the three-dimensional matrix are mapped to the machine ID through two dimensions; In the action space, the three-dimensional matrix of the action output by the agent is represented as , a three-dimensional matrix Each element task Indicates that task i is assigned to machine , I represents the number of tasks, J and K represent the two dimension indexes of the machine respectively.

10. The task hierarchical scheduling method for intelligent computing fusion network according to claim 9, characterized in that: When using a scheduling algorithm based on reinforcement learning to optimize the task allocation strategy, the reward function for: Where, represents the reward for task response time, A reward representing the load balancing of the machines, and Respectively and Adjustable weight parameters of ; Among them, the reward for task i’s response time is and machine load balancing bonus Respectively expressed as: Where, represents the execution time of task i in the current round, represents the actual response time of task i in the current round, represents the expected response time of task i in the current round, Indicates the total number of machines, Indicates the load of the machine, Indicates the average load of all machines.

Images