Computing power resource scheduling method and system based on cloud network fusion

Abstract

The invention provides a computing power resource scheduling method and system based on cloud network integration. Selecting a computing power task to be scheduled as a current scheduling task, and generating a node resource adaptation matrix based on the real-time load data, the cloud network topological relation between the nodes and the dynamic bandwidth data; then calculating the lowest scheduling cost of each node according to the matrix, and determining a dynamic adjustment factor when a resource allocation conflict occurs in the current scheduling task; based on the lowest scheduling cost, the computing power demand scale and the data transmission estimated overhead, calculating the final scheduling overhead for scheduling the current scheduling task to each node; and finally, allocating tasks to a target node according to the final scheduling overhead, updating a running task queue, if a conflict occurs, calling a dynamic adjustment factor to execute resource reallocation, and updating the queue after the reallocation succeeds. And circulating the process until all tasks are scheduled. According to the scheme, optimal scheduling of computing power resources in a cloud network convergence environment can be realized.

Description

Technical Field

[0001] This application relates to the field of computing resource scheduling technology, and more specifically, to a computing resource scheduling method and system based on cloud-network convergence. Background Technology

[0002] Cloud-network convergence technology is a new computing paradigm that deeply integrates cloud computing, network, and computing resources. By achieving synergistic optimization of computing power, network, and data, it can significantly improve resource utilization efficiency and task response speed. In a cloud-network convergence environment, computing resource scheduling is a core component, involving the dynamic allocation of distributed nodes to meet diverse computing task requirements, such as AI training, edge computing, and real-time data processing.

[0003] Traditional computing resource scheduling methods primarily rely on static resource allocation or single-dimensional optimization, such as load balancing-based scheduling algorithms or network-aware routing. However, these methods have significant limitations: in cloud-network convergence scenarios, traditional scheduling ignores dynamic changes in network topology and bandwidth fluctuations, leading to inaccurate transmission overhead estimates and potential task delays or resource waste. Furthermore, their resource allocation conflict handling mechanisms are simplistic and cannot effectively address priority adjustments and reallocation during multi-task concurrency, thus impacting overall system performance and SLA (Service Level Agreement) guarantees. Therefore, achieving optimized scheduling of computing resources in cloud-network convergence environments has become a pressing issue for the industry. Summary of the Invention

[0004] This application provides a method and system for scheduling computing resources based on cloud-network convergence, which can realize the optimized scheduling of computing resources in a cloud-network convergence environment.

[0005] Firstly, this application provides a method for scheduling computing resources based on cloud-network convergence, comprising the following steps: Obtain the queue of computing power tasks to be scheduled from the cloud network computing power cluster in the cloud network converged environment, and select one computing power task to be scheduled from the queue of computing power tasks to be scheduled as the current scheduling task; Monitor and acquire the running task queues and resource status of each cloud-network converged resource node in the cloud-network computing power cluster, and generate the node resource adaptation matrix of the cloud-network computing power cluster based on the real-time load data of each node, the cloud-network topology relationship between nodes and dynamic bandwidth data. Based on the node resource adaptation matrix, the minimum scheduling cost of each cloud-network converged resource node is calculated, and the dynamic adjustment factor corresponding to each node when a resource allocation conflict occurs in the current scheduling task is determined accordingly. Based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated cost of transmitting the data required by the current scheduling task in the cloud network, the final scheduling cost generated by scheduling the current scheduling task to each node is calculated. Based on the final scheduling overhead, the current scheduling task is assigned to the target node, and the running task queue of the target node is updated. If a resource conflict occurs during the allocation process, the dynamic adjustment factor of the corresponding node is invoked to perform resource reallocation for the current scheduling task, and the running task queue of the corresponding node is updated after the reallocation is successful. The above process is repeated until all tasks in the queue of computing power tasks to be scheduled are scheduled.

[0006] In some embodiments, obtaining the queue of computing power tasks to be scheduled from the cloud-network converged computing power cluster in the cloud-network converged environment, and selecting a computing power task to be scheduled from the queue of computing power tasks to be scheduled as the current scheduling task specifically includes: Monitor changes in computing resource demand in the cloud-network converged environment and dynamically update the queue of computing tasks to be scheduled; The queue of computing power tasks to be scheduled is sorted according to task priority, deadline, and computing power requirement. Select the highest priority task from the sorted queue as the current scheduling task, and mark its computing power requirement parameters, data location and data transmission requirements.

[0007] In some embodiments, monitoring and acquiring the running task queues and resource status of each cloud-network converged resource node in the cloud-network computing power cluster, and generating the node resource adaptation matrix of the cloud-network computing power cluster based on the real-time load data of each node, the cloud-network topology relationship between nodes, and dynamic bandwidth data specifically includes: Real-time load data of each cloud-network converged resource node is calculated based on resource status. The cloud network topology relationship between nodes is analyzed, which includes link latency and available paths, and the interconnection efficiency between nodes is calculated through dynamic bandwidth data. Based on real-time load data, cloud network topology, and resource and bandwidth requirements of current scheduled tasks, a multi-dimensional node resource adaptation matrix is ​​constructed.

[0008] In some embodiments, the minimum scheduling cost of each cloud-network converged resource node is calculated based on the node resource adaptation matrix, and the dynamic adjustment factor corresponding to each node when a resource allocation conflict occurs in the current scheduling task is determined accordingly. Specifically, this includes: Extract the load balancing metrics and remaining resource capacity of each node from the node resource adaptation matrix; The minimum scheduling cost for each node for the current scheduled task is calculated using the load balancing metrics and the remaining resource capacity. Based on the distribution of the lowest scheduling cost, the priority of the current scheduled task, and the service SLA level, a dynamic adjustment factor is determined. The dynamic adjustment factor is proportional to the probability of node resource conflict, and the dynamic adjustment factor is used to quantify the impact of conflict based on load balancing indicators and remaining resource capacity, and is used for the redistribution of priority tasks.

[0009] In some embodiments, based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated cost of transmitting the data required by the current scheduling task in the cloud network, the final scheduling cost generated by scheduling the current scheduling task to each node is specifically calculated as follows: Quantify the computing power requirements of the current scheduled task, including floating-point operation requirements and parallel processing requirements; The estimated data transmission overhead is obtained by weighting and fusing the minimum scheduling cost, the computing power requirement, and the estimated data transmission overhead to obtain the final scheduling overhead value for each node.

[0010] In some embodiments, based on the final scheduling overhead, the current scheduling task is assigned to the target node, and the running task queue of the target node is updated; if a resource conflict occurs during the allocation process, the dynamic adjustment factor of the corresponding node is invoked to perform resource reallocation for the current scheduling task, and the running task queue of the corresponding node is updated after successful reallocation; the above process is repeated until all tasks in the queue of computing power tasks to be scheduled are scheduled, specifically including: Select the node with the lowest final scheduling overhead as the target node, and insert the current scheduling task into its running task queue; The system detects whether resource allocation leads to resource overload. If a conflict occurs, it adjusts the task's resource requests based on a dynamic adjustment factor, migrates the task to other nodes, or migrates some subtasks. If the number of consecutive reassignment failures exceeds a preset threshold, task degradation or resource expansion will be triggered. After reallocation, verify resource availability and update the running task queue of the relevant nodes; repeat the above scheduling loop until the running task queue is empty.

[0011] In some embodiments, the scheduling process further includes: Monitor network fluctuations in the cloud-network converged environment. If a bandwidth change exceeds a preset threshold, dynamically update the node resource adaptation matrix and recalculate the final scheduling overhead. Record scheduling history data to optimize the priority ranking and adjustment factor calculation of subsequent tasks, and avoid scheduling deadlock.

[0012] Secondly, this application provides a cloud-network convergence-based computing resource scheduling system, comprising: The acquisition module is used to acquire the queue of computing power tasks to be scheduled from the cloud network computing power cluster in the cloud network converged environment, and select a computing power task to be scheduled from the queue of computing power tasks to be scheduled as the current scheduling task. The processing module is used to monitor and obtain the running task queues and resource status of each cloud-network converged resource node in the cloud-network computing power cluster, and generate the node resource adaptation matrix of the cloud-network computing power cluster based on the real-time load data of each node, the cloud-network topology relationship between nodes and dynamic bandwidth data. The processing module is also used to calculate the minimum scheduling cost of each cloud-network converged resource node based on the node resource adaptation matrix, and thereby determine the dynamic adjustment factor corresponding to each node when a resource allocation conflict occurs in the current scheduling task. The processing module is also used to calculate the final scheduling cost of scheduling the current scheduling task to each node based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated cost of transmitting the data required by the current scheduling task in the cloud network. The allocation module is used to allocate the current scheduling task to the target node based on the final scheduling overhead and update the running task queue of the target node. If a resource conflict occurs during the allocation process, the dynamic adjustment factor of the corresponding node is called to perform resource reallocation for the current scheduling task, and the running task queue of the corresponding node is updated after the reallocation is successful. The above process is repeated until all tasks in the queue of computing power tasks to be scheduled are scheduled.

[0013] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described cloud-network converged computing resource scheduling method.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned cloud-network converged computing resource scheduling method.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: In this embodiment, the system obtains the queue of scheduled computing tasks of the cloud-network converged computing power cluster in the cloud-network converged environment, and selects a scheduled computing task from the queue as the current scheduled task; it monitors and obtains the running task queues and resource status of each cloud-network converged resource node in the cloud-network converged computing power cluster, and generates a node resource adaptation matrix of the cloud-network converged computing power cluster based on the real-time load data of each node, the cloud-network topology relationship between nodes, and dynamic bandwidth data; based on the node resource adaptation matrix, it calculates the minimum scheduling cost of each cloud-network converged resource node, and thereby determines the corresponding resource allocation conflict of each node when the current scheduled task occurs. The system dynamically adjusts the scheduling factor. Based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated cost of transmitting the data required by the current scheduling task in the cloud network, the system calculates the final scheduling cost of scheduling the current scheduling task to each node. According to the final scheduling cost, the system allocates the current scheduling task to the target node and updates the running task queue of the target node. If a resource conflict occurs during the allocation process, the system calls the dynamic adjustment factor of the corresponding node to perform resource reallocation for the current scheduling task, and updates the running task queue of the corresponding node after successful reallocation. The system repeats the above process until all tasks in the queue of computing power tasks to be scheduled are scheduled.

[0016] Therefore, this application achieves task allocation and resource redistribution in the event of conflicts by calculating the minimum scheduling cost and dynamic adjustment factor based on the node resource adaptation matrix and combining it with the final scheduling overhead. First, by monitoring the running task queues and resource status of each node in the cloud-network computing cluster and integrating real-time load data, cloud-network topology relationships, and dynamic bandwidth data to generate a node resource adaptation matrix, it can comprehensively capture the dynamic characteristics of the cloud-network converged environment and achieve refined modeling of resource nodes, thereby improving the accuracy and adaptability of scheduling decisions from the source. Second, by calculating the minimum scheduling cost of each node based on the node resource adaptation matrix and determining the dynamic adjustment factor, it can quantify the potential scheduling cost and predict conflict scenarios, providing a low-overhead guidance mechanism for subsequent allocation and effectively reducing resource waste and performance issues caused by blind allocation. The system first identifies the bottleneck; then, based on the minimum scheduling cost, the scale of task computing power requirements, and the estimated data transmission overhead, it calculates the final scheduling cost. This allows for a comprehensive evaluation of the feasibility of the scheduling scheme from multiple dimensions, ensuring that task allocation balances computational efficiency and network transmission optimization, avoiding the limitations of traditional methods that ignore cloud-network coupling effects. Finally, tasks are allocated to target nodes based on the final scheduling cost, and a dynamic adjustment factor is invoked to perform reallocation when resource conflicts occur. This enables adaptive scheduling paths, dynamically resolving conflicts and updating the task queue, significantly improving the overall system resource utilization and task completion time. This mechanism, through iterative execution until all tasks are scheduled, constructs an efficient and robust resource scheduling framework in a cloud-network converged environment, effectively overcoming the inefficiency caused by static evaluation or neglect of conflict handling in existing scheduling methods. In summary, the proposed solution can achieve optimized scheduling of computing resources in a cloud-network converged environment. Attached Figure Description

[0017] Figure 1 This is an exemplary flowchart of a cloud-network converged computing resource scheduling method according to some embodiments of this application; Figure 2 This is an exemplary flowchart illustrating the determination of dynamic allocation factors according to some embodiments of this application; Figure 3 This is a schematic diagram of a node resource adaptation matrix according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a cloud-network converged computing resource scheduling system according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device that implements a cloud-network converged computing resource scheduling method according to some embodiments of this application. Detailed Implementation

[0018] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific implementation methods. (Reference) Figure 1The figure is an exemplary flowchart of a cloud-network converged computing resource scheduling method according to some embodiments of this application. The method mainly includes the following steps: In step 101, the queue of computing power tasks to be scheduled for the cloud network computing power cluster in the cloud network converged environment is obtained, and a computing power task to be scheduled for the current scheduling task is selected from the queue of computing power tasks to be scheduled for the current scheduling task.

[0019] Among them, the cloud-network converged environment refers to the deep integration architecture of cloud computing, network resources and computing power resources. For example, edge cloud nodes are connected to the core cloud data center through 5G / fiber optic networks, and the cloud-network computing power cluster includes multiple distributed nodes, such as edge servers and cloud servers. The queue of computing power tasks to be scheduled can include computing tasks such as AI inference and data analysis, which will not be elaborated here.

[0020] In some embodiments, obtaining the queue of computing power tasks to be scheduled from the cloud-network converged computing power cluster in the cloud-network converged environment, and selecting a computing power task to be scheduled from the queue of computing power tasks to be scheduled as the current scheduling task specifically includes: The system monitors changes in computing resource demand in the cloud-network converged environment and dynamically updates the queue of computing tasks to be scheduled. In practice, this can be achieved by collecting task requests submitted by users in real time and adjusting the queue priority based on fluctuations in resource utilization, such as adjusting the queue priority when the central processing unit (CPU) utilization exceeds 80%. This is just an example and is not intended to limit the scope of this application. The queue of tasks to be scheduled is sorted according to task priority, deadline, and computing power requirement. Specifically, different weights (totaling 1) are assigned to the three key dimensions of task priority, deadline, and computing power requirement. A weighted score is calculated for each task, and tasks with higher scores are placed at the front of the queue and scheduled first. For example: There are two tasks to be scheduled. Calculate their scores and sort them: Task A: Core AI inference task (priority = 5), total cycle 10 hours, remaining 2 hours (urgency = 10 / 2 = 5), requires 4 TFLOPS computing power (scale = 4), the weighted score of task A = 0.5×5 + 0.3×5 + 0.2×4 = 2.5 + 1.5 + 0.8 = 4.8; Task B: Ordinary log analysis task (priority = 2), total cycle 5 hours, remaining 1 hour (urgency = 5 / 1 = 5), requires 6 TFLOPS computing power (scale = 6), the weighted score of task B = 0.5×2 + 0.3×5 + 0.2×6 = 1.0 + 1.5 + 1.2 = 3.7; Sorting results: Task A has a higher weighted score and will be scheduled first; The task with the highest priority is selected from the sorted queue as the current scheduling task, and its computing power requirement parameters, data location and data transmission requirements are marked. In specific implementation, for example, the task is marked with a GPU requirement of 4 cores, a data location of the storage volume of edge node A, and a transmission requirement of latency <50ms.

[0021] In step 102, the running task queues and resource status of each cloud-network converged resource node in the cloud-network computing power cluster are monitored and obtained. Based on the real-time load data of each node, the cloud-network topology relationship between nodes and dynamic bandwidth data, a node resource adaptation matrix of the cloud-network computing power cluster is generated.

[0022] Among them, cloud-network converged resource nodes refer to nodes that integrate computing, networking and storage functions, such as edge computing nodes or cloud data center servers. Resource status includes CPU / GPU utilization, etc., which are not specifically limited here.

[0023] In some embodiments, monitoring and acquiring the running task queues and resource status of each cloud-network converged resource node in the cloud-network computing power cluster, and generating the node resource adaptation matrix of the cloud-network computing power cluster based on the real-time load data of each node, the cloud-network topology relationship between nodes, and dynamic bandwidth data specifically includes: Real-time load data for each cloud-network converged resource node is calculated based on resource status. In practice, the utilization of four key resources—CPU, GPU, memory, and storage—is integrated into a unified load data value. For example, the load data is calculated as (CPU utilization 0.4 + GPU utilization 0.3 + memory utilization 0.2 + storage utilization 0.1) to quantify the overall node load. Here, 0.4, 0.3, 0.2, and 0.1 represent the weight percentages of the four resource types, and their sum is 1, reflecting the priority of different resources' impact on the overall node load. In practice, this can be set according to the characteristics of computing tasks in cloud-network converged scenarios. This is only an example and is not intended to be a specific limitation on the weight of different resources on nodes. The analysis focuses on the cloud network topology relationships between nodes, including link latency and available paths. Interconnection efficiency is calculated using dynamic bandwidth data. The cloud network topology describes the network connection structure between resource nodes (e.g., edge servers, core cloud nodes) within the cloud computing cluster. Link latency refers to the time consumed by data transmission in the physical / logical link between two nodes, a key indicator of network transmission real-time performance. Available paths refer to all independent connection routes between two nodes that enable data transmission, reflecting the redundancy and reliability of network connections. In practice, real-time link latency can be obtained through monitoring. For example, the direct link latency from edge node A to core cloud node B is 20ms, while the link latency via relay node C is 35ms. It should be noted that lower latency means shorter data transmission time, making it more suitable for tasks with high real-time requirements, such as AI inference and real-time data processing. In addition, available paths can be identified through topology analysis. For example, available paths from node X to node Y include: 1. Direct fiber optic link; 2. 5G link via edge gateway; 3. Ethernet link via regional aggregation node. It should be noted that the more available paths there are, the more alternative paths can be switched when a link fails or becomes congested, reducing the risk of task transmission interruption.

[0024] In addition, interconnection efficiency is a quantitative indicator that integrates cloud-network topology relationships (link latency, available paths) and dynamic bandwidth data. It is used to evaluate the efficiency of data transmission between nodes and directly supports the construction of the node resource adaptation matrix. In specific implementation, it can be done in the following way: First, the optimal path is selected from the available paths, usually prioritizing the path with the lowest cost. The path cost calculation formula is: Path cost = λ × (latency / latency_max) + μ × bandwidth utilization + ν × (packet loss rate / packet loss rate_max) (where λ + μ + ν = 1, λ, μ, and ν are weights, which are set according to the task type, such as λ = 0.5, μ = 0.3, ν = 0.2 for real-time tasks; latency_max and packet loss rate_max are the cluster's historical maximum latency and maximum packet loss rate). Secondly, based on the link delay and dynamic bandwidth utilization of the optimal path, the interconnection efficiency is calculated. Specifically, it can be calculated using the following formula: Interconnection efficiency = (1 - bandwidth utilization) / link delay.

[0025] Based on real-time load data, cloud-network topology, and resource and bandwidth requirements of the current scheduled task, a multi-dimensional node resource adaptation matrix is ​​constructed. This matrix quantifies the compatibility between each cloud-network converged resource node and the current scheduled task. Essentially, it uses a table format of node × resource dimension. In this application, the node resource adaptation matrix is ​​a task-aware dynamic matrix. Specifically, the load data of each node is arranged according to the topology to form a multi-dimensional matrix. In the implementation, the matrix dimension is the number of nodes × the number of resource dimensions, where rows correspond to each cloud-network converged resource node in the cloud-network computing cluster; columns correspond to the core resource dimensions related to the current scheduled task, mainly including the following three categories: In terms of computing resources, such as CPU and GPU, the computing power requirements of the corresponding task; Storage resources, such as memory and hard disk, correspond to the data storage requirements of the task; Transmission resource dimensions, such as interconnection efficiency, correspond to the data transmission requirements of the task; The element values ​​in the matrix represent the node's suitability for the current task in a certain resource dimension, using a score of 0-1 (where 1 = perfectly suited, 0 = completely unsuitable). A higher score indicates a more suitable node for the current task. It's important to note that the multi-level node resource suitability matrix is ​​a pre-built resource matching database that maps node resources and task types to suitability scores. (See reference...) Figure 3 This diagram is a specific embodiment of the node resource adaptation matrix. This embodiment is a node resource adaptation matrix of a cloud-network converged cluster. The node IDs can be Node1, Node2, and Node3. The resource types include, for example, CPU, GPU, memory, and transmission efficiency. The CPU adaptation degree of Node1 is 0.8, the GPU adaptation degree is 0.7, etc. This embodiment is only an example for illustration. The actual matrix can also be dynamically optimized based on machine learning. No specific limitations are made here.

[0026] In step 103, the minimum scheduling cost of each cloud-network converged resource node is calculated based on the node resource adaptation matrix, and the dynamic adjustment factor corresponding to each node when a resource allocation conflict occurs in the current scheduling task is determined.

[0027] Here, the minimum scheduling cost refers to the minimum overhead of node scheduling for the current task, and the dynamic adjustment factor refers to the quantitative parameter for resource coordination during conflicts. In some embodiments, it refers to... Figure 3 As shown, the dynamic adjustment factor for each node when a resource allocation conflict occurs in the current scheduled task can be determined in the following way: In step 1031, the load balancing index and remaining resource capacity of each node are extracted from the node resource adaptation matrix. The load balancing index is a parameter that quantifies the uniformity of load distribution among resource nodes within the cloud computing cluster. Its core purpose is to avoid overloading of some nodes, such as excessively high CPU / GPU utilization, or idle nodes, ensuring balanced overall resource utilization efficiency. The remaining resource capacity refers to the amount of remaining resources that a single cloud-network converged resource node can currently use to support new tasks. It directly reflects whether the node can meet the resource requirements of the currently scheduled task and is the core basis for judging node adaptability. The load balancing index is calculated using the load standard deviation coefficient. The specific steps are: 1. Calculate the normalized load data (L1, L2, ..., L...) of all nodes. n ∈[0,1]); 2. Calculate the average load L_avg=(ΣLᵢ) / n; 3. Calculate the standard deviation of the load σ=√[(Σ(Lᵢ-L_avg)²) / n], where n is the total number of cloud-network converged resource nodes, and the value of i ranges from 1 to n; 4. The original load balancing index = L_balance=σ / L_avg (L_avg≠0); 5. The normalized load balancing index L_norm=min(L_balance / L_balance_max,1), where L_balance_max is the original load balancing index of the cluster's historical maximum; In step 1032, the minimum scheduling cost for each node for the current scheduled task is calculated using the load balancing index and the remaining resource capacity. The core logic of this application in calculating the minimum scheduling cost for each node using the load balancing index and remaining resource capacity is to comprehensively judge whether the node can handle the task and whether the cluster load balance is maintained after scheduling, ultimately quantifying the minimum scheduling cost for each node for the current task. Specifically, before calculation, each parameter needs to be normalized: transmission delay cost normalization (D_norm = actual delay / maximum allowable delay) and energy consumption cost normalization (E_norm = actual energy consumption / maximum rated energy consumption of the node). Then, the minimum scheduling cost is calculated using the formula: Minimum scheduling cost = α × (L_norm × inverse weight of remaining resource capacity) + β × D_norm + γ×E_norm (where α+β+γ=1, α, β, and γ are weights, set according to task type, such as α=0.3, β=0.4, γ=0.3 for real-time tasks); The load balancing metric calculates the uniformity of node resource distribution through variance. Remaining resource capacity is the remaining resources a node has to support new tasks, such as the number of remaining CPU cores or GPU computing power. Inverse weighting means that the more resources remaining, the smaller the weight; for example, the weight of 10 remaining CPU cores < the weight of 2 remaining CPU cores, representing the pressure on the node to handle tasks. The load balancing metric is multiplied by E_norm. The inverse weight of remaining resource capacity means that if the node load balance is high and there are many remaining resources (small inverse weight, such as 0.1), then this part of the cost is low (0.9×0.1=0.09), indicating that the node has good adaptability and is not prone to overload or imbalance after scheduling tasks. The transmission delay cost is the time cost required to transmit the data of the current task, such as the dataset to be processed, to the target node. The higher the transmission delay cost, the more it will directly increase the total scheduling cost. The energy consumption cost refers to the energy cost consumed by the node when carrying the current task. It is usually positively correlated with the node load. For example, a high-load node consumes more energy when running a task. In step 1033, a dynamic adjustment factor is determined based on the distribution of the lowest scheduling cost, the priority of the current scheduled task, and the service SLA level. The dynamic adjustment factor is proportional to the node resource conflict probability and quantifies the conflict impact based on load balancing indicators and remaining resource capacity. It is used for the redistribution of priority tasks. The larger the dynamic adjustment factor value, the higher the priority for resource redistribution should be when a conflict occurs on that node. Specifically, the dynamic adjustment factor needs to be calculated after normalizing the numerator and denominator: Numerator_norm = (Priority × SLA Level) / (Highest Priority × Highest SLA Level) (Highest Priority and Highest SLA Level are the cluster's preset maximum values), Denominator_norm = (Lowest Scheduling Cost + Conflict Probability) / 2. The final formula is: Dynamic Adjustment Factor = min (Numerator_norm / Denominator_norm, 1) (If denominator_norm = 0, the factor is set to 1). Priority: This refers to the business importance of the currently scheduled task. For example, core AI inference tasks have high priority (set to 5), while ordinary log analysis tasks have low priority (set to 2). The larger the value, the more priority the task needs to be guaranteed.

[0028] SLA Level: Service Level Agreement Level, used to quantify the service assurance requirements of a task. For example, a high SLA requires latency <50ms and a success rate of 100%, and the level is set to 4; a low SLA has more lenient requirements and the level is set to 1. The higher the value, the more stringent the resource assurance required for the task.

[0029] Minimum scheduling cost: This is the minimum cost for a node to carry out the current task. The smaller the value, the more suitable the node is for the task.

[0030] Conflict probability: refers to the estimated probability of a conflict occurring when a node is assigning a current task due to insufficient resources. For example, if the node has a high load, the probability of conflict is high, so the value is set to 0.8; if the load is low, it is set to 0.2.

[0031] Priority × SLA Level): This reflects the degree to which a task needs to be prioritized for reallocation. The more important the task and the stricter the SLA requirements, the larger the numerator, indicating that resources should be coordinated for it first in case of conflict.

[0032] Minimum scheduling cost + conflict probability: This reflects the difficulty for a node to resolve conflicts for a task. The higher the node adaptability (lower minimum scheduling cost) and the lower the conflict risk (lower conflict probability), the smaller the denominator, indicating that the node is more likely to carry out tasks through reallocation.

[0033] In step 104, based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated cost of transmitting the data required by the current scheduling task in the cloud network, the final scheduling cost generated by scheduling the current scheduling task to each node is calculated.

[0034] In some embodiments, based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated cost of transmitting the data required by the current scheduling task in the cloud network, the final scheduling cost generated by scheduling the current scheduling task to each node is specifically calculated as follows: The computational power requirement of the current scheduled task is quantified. The computational power requirement includes floating-point operation requirements and parallel processing requirements. In a specific implementation, for example, the computational power requirement is calculated as: computational power requirement = floating-point operation requirement (TFLOPS) + number of parallel threads 0.1. Here, the number of parallel threads × 0.1 is the parallel processing adaptation requirement of the task. The number of parallel threads refers to the number of computing units that can be divided into and executed simultaneously, representing the task's requirement for the parallel processing capability of the nodes. Multiplying by 0.1 is because its impact on the overall computational power requirement is much smaller than the core floating-point operation requirement. It can also be adjusted according to the actual situation. This is only an example and is not intended to limit the specific scope of the invention. Before estimating data transmission overhead, the computing power requirement and the estimated data transmission overhead need to be normalized: computing power requirement normalization (C_norm = task computing power requirement / node maximum computing power, if requirement > maximum computing power then C_norm = 1.1), and estimated data transmission overhead normalization (T_norm = task data volume / node maximum storage bandwidth). Then, the minimum scheduling cost, the normalized computing power requirement, and the estimated data transmission overhead are weighted and fused to obtain the final scheduling overhead value for each node. In a specific implementation, for example, the minimum scheduling cost has a weight of 0.4, the normalized computing power requirement has a weight of 0.3, and the normalized estimated data transmission overhead has a weight of 0.3. The final scheduling overhead = 0.4 × minimum scheduling cost + 0.3 × C_norm + 0.3 × T_norm. This is only an example and is not intended to limit the specific scope of the invention.

[0035] In step 105, based on the final scheduling overhead, the current scheduling task is assigned to the target node, and the running task queue of the target node is updated. If a resource conflict occurs during the assignment process, the dynamic adjustment factor of the corresponding node is invoked to perform resource reallocation for the current scheduling task, and the running task queue of the corresponding node is updated after the reallocation is successful. The above process is repeated until all tasks in the queue of computing power tasks to be scheduled are scheduled.

[0036] In some embodiments, based on the final scheduling overhead, the current scheduling task is assigned to the target node, and the running task queue of the target node is updated; if a resource conflict occurs during the allocation process, the dynamic adjustment factor of the corresponding node is invoked to perform resource reallocation for the current scheduling task, and the running task queue of the corresponding node is updated after successful reallocation; the above process is repeated until all tasks in the queue of computing power tasks to be scheduled are scheduled, specifically including: Select the node with the lowest final scheduling overhead as the target node, and insert the current scheduled task into its running task queue. In practice, for example, a priority queue can be used to insert the task and update the node resource status. The system detects whether resource allocation leads to resource overload. If a conflict occurs, it adjusts the resource requests of the task based on a dynamic adjustment factor, migrates the task to other nodes, or migrates some subtasks. In specific implementation, for example, if the dynamic adjustment factor is >0.5, the subtasks are migrated to the second-best node first. If the reassignment fails consecutively for more than a preset threshold, the task will be downgraded or the resources will be expanded. In practice, for example, if the threshold is 3 failures, the task will be downgraded to a low priority or the cloud API will be called to expand the node. After reallocation, verify resource availability and update the running task queue of the relevant nodes; repeat the above scheduling loop until the running task queue is empty.

[0037] In addition, in some embodiments of this application, the scheduling process may further include: Monitor network fluctuations in the cloud-network converged environment. If a bandwidth change exceeds a preset threshold, dynamically update the node resource adaptation matrix and recalculate the final scheduling overhead. In specific implementation, for example, if the threshold is 20% bandwidth fluctuation, matrix reconstruction will be triggered. Record scheduling history data to optimize the priority ranking and adjustment factor calculation of subsequent tasks, and avoid scheduling deadlocks. In specific implementations, log analysis can be used to avoid deadlocks caused by circular dependencies.

[0038] Furthermore, in another aspect of this application, in some embodiments, this application provides a cloud-network converged computing resource scheduling system, with reference to... Figure 4 The figure is a schematic diagram of the structure of a cloud-network converged computing resource scheduling system 400 according to some embodiments of this application. The cloud-network converged computing resource scheduling system 400 includes: an acquisition module 401, a processing module 402, and an allocation module 403, which are described below: The acquisition module 401 in this application is mainly used to acquire the queue of computing power tasks to be scheduled in the cloud network computing power cluster in the cloud network converged environment, and select a computing power task to be scheduled from the queue of computing power tasks to be scheduled as the current scheduling task. Processing module 402, in this application, is mainly used to monitor and obtain the running task queues and resource status of each cloud network converged resource node in the cloud network computing power cluster, and generate the node resource adaptation matrix of the cloud network computing power cluster based on the real-time load data of each node, the cloud network topology relationship between nodes and dynamic bandwidth data. The processing module 402 is further configured to calculate the minimum scheduling cost of each cloud-network converged resource node based on the node resource adaptation matrix, and thereby determine the dynamic adjustment factor corresponding to each node when a resource allocation conflict occurs in the current scheduling task. The processing module 402 is also used to calculate the final scheduling cost generated by scheduling the current scheduling task to each node based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated cost of transmitting the data required by the current scheduling task in the cloud network. The allocation module 403 in this application is mainly used to allocate the current scheduling task to the target node according to the final scheduling overhead, and update the running task queue of the target node; if a resource conflict occurs during the allocation process, the dynamic adjustment factor of the corresponding node is called to perform resource reallocation for the current scheduling task, and the running task queue of the corresponding node is updated after the reallocation is successful; the above process is repeated until all tasks in the queue of computing power tasks to be scheduled are scheduled.

[0039] In addition, this application also provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described cloud-network converged computing resource scheduling method.

[0040] In some embodiments, reference Figure 5 This figure is a schematic diagram of the structure of a computer device implementing a cloud-network converged computing resource scheduling method according to some embodiments of this application. The method in the above embodiments can be achieved through... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0041] The processor 501 may be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the cloud-network convergence-based computing resource scheduling method in this application.

[0042] The communication bus 502 may include a path for transmitting information between the aforementioned components.

[0043] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CDROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0044] The memory 503 stores program code that executes the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. In the above embodiments, the calculation of scheduling overhead can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0045] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0046] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single CPU) processor or a multi-core (multi CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0047] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0048] In addition, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described cloud-network converged computing resource scheduling method.

[0049] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0050] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of the invention. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. A method for scheduling computing resources based on cloud-network convergence, characterized in that, Includes the following steps: Obtain the queue of computing power tasks to be scheduled from the cloud network computing power cluster in the cloud network converged environment, and select one computing power task to be scheduled from the queue of computing power tasks to be scheduled as the current scheduling task; Monitor and acquire the running task queues and resource status of each cloud-network converged resource node in the cloud-network computing power cluster, and generate the node resource adaptation matrix of the cloud-network computing power cluster based on the real-time load data of each node, the cloud-network topology relationship between nodes and dynamic bandwidth data. Based on the node resource adaptation matrix, the minimum scheduling cost of each cloud-network converged resource node is calculated, and the dynamic adjustment factor corresponding to each node when a resource allocation conflict occurs in the current scheduling task is determined accordingly. Based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated cost of transmitting the data required by the current scheduling task in the cloud network, the final scheduling cost generated by scheduling the current scheduling task to each node is calculated. Based on the final scheduling overhead, the current scheduling task is assigned to the target node, and the running task queue of the target node is updated; If a resource conflict occurs during the allocation process, the dynamic adjustment factor of the corresponding node is invoked to perform resource reallocation for the current scheduled task, and the running task queue of the corresponding node is updated after the reallocation is successful; the above process is repeated until all tasks in the queue of computing power tasks to be scheduled are scheduled.

2. The method according to claim 1, characterized in that, Obtaining the queue of computing power tasks to be scheduled from the cloud-network converged environment and selecting one computing power task from the queue as the current scheduling task specifically includes: Monitor changes in computing resource demand in the cloud-network converged environment and dynamically update the queue of computing tasks to be scheduled; The queue of computing power tasks to be scheduled is sorted according to task priority, deadline, and computing power requirement. Select the highest priority task from the sorted queue as the current scheduling task, and mark its computing power requirement parameters, data location and data transmission requirements.

3. The method according to claim 1, characterized in that, Monitoring and acquiring the running task queues and resource status of each cloud-network converged resource node in the cloud-network computing power cluster, and generating the node resource adaptation matrix of the cloud-network computing power cluster based on the real-time load data of each node, the cloud-network topology relationship between nodes, and dynamic bandwidth data, specifically includes: Real-time load data of each cloud-network converged resource node is calculated based on resource status. The cloud network topology relationship between nodes is analyzed, which includes link latency and available paths, and the interconnection efficiency between nodes is calculated through dynamic bandwidth data. Based on real-time load data, cloud network topology, and resource and bandwidth requirements of current scheduled tasks, a multi-dimensional node resource adaptation matrix is ​​constructed.

4. The method according to claim 1, characterized in that, Based on the node resource adaptation matrix, the minimum scheduling cost of each cloud-network converged resource node is calculated, and the dynamic adjustment factor for each node when a resource allocation conflict occurs in the current scheduling task is determined accordingly, including: Extract the load balancing metrics and remaining resource capacity of each node from the node resource adaptation matrix; The minimum scheduling cost for each node for the current scheduled task is calculated using the load balancing metrics and the remaining resource capacity. Based on the distribution of the lowest scheduling cost, the priority of the current scheduled task, and the service SLA level, a dynamic adjustment factor is determined. The dynamic adjustment factor is proportional to the probability of node resource conflict, and the dynamic adjustment factor is used to quantify the impact of conflict based on load balancing indicators and remaining resource capacity, and is used for the redistribution of priority tasks.

5. The method according to claim 1, characterized in that, Based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated overhead of transmitting the data required by the current scheduling task in the cloud network, the final scheduling overhead generated by scheduling the current scheduling task to each node is calculated, specifically including: Quantify the computing power requirements of the current scheduled task, including floating-point operation requirements and parallel processing requirements; The estimated data transmission overhead is obtained by weighting and fusing the minimum scheduling cost, the computing power requirement, and the estimated data transmission overhead to obtain the final scheduling overhead value for each node.

6. The method according to claim 1, characterized in that, Based on the final scheduling overhead, the current scheduling task is assigned to the target node, and the running task queue of the target node is updated; If a resource conflict occurs during the allocation process, the dynamic adjustment factor of the corresponding node is invoked to perform resource reallocation for the current scheduled task, and the running task queue of the corresponding node is updated after the reallocation is successful. The above process is repeated until all tasks in the queue of computing power tasks to be scheduled are scheduled. Specifically, this includes: Select the node with the lowest final scheduling overhead as the target node, and insert the current scheduling task into its running task queue; The system detects whether resource allocation leads to resource overload. If a conflict occurs, it adjusts the task's resource requests based on a dynamic adjustment factor, migrates the task to other nodes, or migrates some subtasks. If the number of consecutive reassignment failures exceeds a preset threshold, task degradation or resource expansion will be triggered. After reallocation, verify resource availability and update the running task queue of the relevant nodes; repeat the above scheduling loop until the running task queue is empty.

7. The method according to claim 1, characterized in that, The scheduling process also includes: Monitor network fluctuations in the cloud-network converged environment. If a bandwidth change exceeds a preset threshold, dynamically update the node resource adaptation matrix and recalculate the final scheduling overhead. Record scheduling history data to optimize the priority ranking and adjustment factor calculation of subsequent tasks, and avoid scheduling deadlock.

8. A computing resource scheduling system based on cloud-network convergence, characterized in that, include: The acquisition module is used to acquire the queue of computing power tasks to be scheduled from the cloud network computing power cluster in the cloud network converged environment, and select a computing power task to be scheduled from the queue of computing power tasks to be scheduled as the current scheduling task. The processing module is used to monitor and obtain the running task queues and resource status of each cloud-network converged resource node in the cloud-network computing power cluster, and generate the node resource adaptation matrix of the cloud-network computing power cluster based on the real-time load data of each node, the cloud-network topology relationship between nodes and dynamic bandwidth data. The processing module is also used to calculate the minimum scheduling cost of each cloud-network converged resource node based on the node resource adaptation matrix, and thereby determine the dynamic adjustment factor corresponding to each node when a resource allocation conflict occurs in the current scheduling task. The processing module is also used to calculate the final scheduling cost of scheduling the current scheduling task to each node based on the minimum scheduling cost, the computing power requirement of the current scheduling task, and the estimated cost of transmitting the data required by the current scheduling task in the cloud network. The allocation module is used to allocate the current scheduling task to the target node based on the final scheduling overhead, and update the running task queue of the target node; If a resource conflict occurs during the allocation process, the dynamic adjustment factor of the corresponding node is invoked to perform resource reallocation for the current scheduled task, and the running task queue of the corresponding node is updated after the reallocation is successful; the above process is repeated until all tasks in the queue of computing power tasks to be scheduled are scheduled.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute the cloud-network converged computing resource scheduling method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the cloud-network convergence-based computing resource scheduling method as described in any one of claims 1 to 7.

Images