Method for intelligent scheduling and load balancing of computing power resources

Abstract

The invention relates to the technical field of digital computing and resource scheduling, in particular to a computing power resource intelligent scheduling and load balancing method. The method comprises the following steps: collecting performance indexes of a computing power node on different resource dimensions, and constructing a resource capability vector of the computing power node; extracting demand features of the calculation task on different resource dimensions, and constructing a task feature vector; matching the resource capability vector of the computing power node with the task feature vector to obtain an adaptation degree score; operation indexes of the computing power nodes are collected, comprehensive load scoring is carried out on the computing power nodes, and the current load states of the computing power nodes are obtained; and distributing calculation tasks based on the current load state of the computing power node and the adaptation degree score, and updating the current load state of the computing power node in real time. The problems that the resource utilization rate is low, the task response delay is high and the node load is unbalanced due to the fact that the computing power resources cannot be dynamically matched according to the task characteristics and the computing power node states in a traditional computing power resource scheduling method are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of digital computing and resource scheduling, and in particular to a method for intelligent scheduling and load balancing of computing resources. Background Art

[0002] With the rapid development of cutting-edge technologies such as artificial intelligence, big data, blockchain, and the Internet of Things, compute-intensive tasks are increasing, driving a continuous rise in demand for computing resources. Typical applications such as image recognition, natural language processing, deep learning model training, and scientific simulations are increasingly dependent on high-performance computing power. At the same time, modern computing scenarios are gradually shifting from traditional centralized data centers to distributed and heterogeneous computing architectures, such as edge computing and cloud-edge collaborative computing. These emerging models pose greater challenges to the scheduling of computing resources.

[0003] Traditional computing resource scheduling methods are mostly based on static preset policies or historical load data. These methods lack the sensitivity and adaptability to real-time resource status and are unable to promptly respond to fluctuations in task requests and changes in resource node status. In a multi-tenant environment, different tasks have significantly different computing resource requirements, such as those that are compute-intensive, memory-intensive, and I / O-intensive. Existing scheduling mechanisms struggle to accurately identify task characteristics and match them with appropriate compute nodes, leading to irrational resource allocation.

[0004] In addition, traditional computing resource scheduling methods generally have the following problems: First, computing resource scheduling is coarse-grained and lacks a detailed modeling mechanism, which cannot fully utilize the heterogeneous advantages of nodes; second, the load balancing mechanism is simple, relying only on single indicators such as queue length or CPU utilization, ignoring the coordination of multi-dimensional resources such as GPU, memory capacity, and bandwidth; third, when faced with sudden high-concurrency tasks or resource constraints, the scheduling strategy lacks the ability to make global optimal decisions, which easily leads to an imbalance where nodes are overloaded while other nodes are idle, reducing overall throughput and task processing efficiency. Therefore, there is an urgent need for an intelligent scheduling mechanism that can dynamically perceive resource status, accurately match task characteristics, and achieve multi-node load balancing to improve overall system efficiency. Summary of the Invention

[0005] The present invention provides a method for intelligent scheduling and load balancing of computing power resources to solve the technical problems in traditional computing power resource scheduling methods that computing power resources cannot be dynamically matched according to task characteristics and computing power node status, resulting in low resource utilization, high task response delay and node load imbalance.

[0006] The present invention provides a method for intelligent scheduling and load balancing of computing resources, which specifically includes the following technical solutions: A method for intelligent scheduling and load balancing of computing resources, comprising the following steps: S1. Collect performance indicators of computing nodes in different resource dimensions and normalize them to obtain normalized performance indicators. Based on the normalized performance indicators, construct the resource capability vector of the computing node. S2. By parsing the computing task execution script, configuration files, and historical resource usage records, we extract the demand characteristics of the computing task in different resource dimensions and construct a task feature vector. S3. Match the resource capability vector of the computing power node with the task feature vector to obtain the fitness score between the computing task and the computing power node, and construct the fitness score matrix; S4. Collect the operating indicators of the computing power nodes, perform a comprehensive load score on the computing power nodes, and obtain the current load status of the computing power nodes; based on the current load status and fitness score of the computing power nodes, calculate the scheduling priority score and construct a scheduling priority score matrix; S5. Based on the scheduling priority scoring matrix, computing tasks are allocated, and the current load status of the computing power nodes is updated in real time. The scheduling strategy is dynamically adjusted according to the execution status of the computing tasks.

[0007] Preferably, the S1 specifically includes: Assign a weight coefficient to each resource dimension, combine the exponential function and the logarithmic function, perform a weighted combination of the normalized performance indicators and the weight coefficients, and construct the resource capacity vector of the computing power node.

[0008] Preferably, the S2 specifically includes: Analyze the computing task running scripts, configuration files and historical resource usage records to obtain the specific dependency characteristics of the computing task on computing, storage and communication resources; based on the specific dependency characteristics of the computing task on the said resources, extract the demand of the computing task for resources of each dimension as the demand characteristics of the computing task in the resource dimension.

[0009] Preferably, the S2 specifically includes: The resource demand weight is set for each resource dimension. The exponential function and logarithmic function are combined to perform a weighted combination of the demand characteristics of the computing task in the resource dimension and the resource demand weight to construct the task feature vector.

[0010] Preferably, the S4 specifically includes: The operating indicators of the computing power nodes are collected and normalized to obtain normalized operating indicators; the operating indicators include the CPU utilization rate of the computing power nodes, the GPU utilization rate of the computing power nodes, the task queue length of the computing power nodes, and the average task waiting time of the computing power nodes.

[0011] Preferably, the S4 specifically includes: A weighted coefficient is set for each normalized operating indicator. By taking the weighted sum of the normalized operating indicators, a comprehensive load score is given to the computing power node to obtain the current load status of the computing power node.

[0012] Preferably, the S4 specifically includes: A scheduling trade-off factor is introduced to perform weighted fusion of the fitness score and the current load status of the computing power node to generate a scheduling priority score for the computing task on the computing power node.

[0013] Preferably, the S5 specifically includes: Based on the scheduling priority score matrix and the task allocation status array, the computing tasks are matched with the computing power nodes with the highest scheduling priority scores to achieve task allocation and update the current load status of each computing power node in real time.

[0014] The beneficial effects of the technical solution of the present invention are: 1. By collecting performance indicators of computing nodes across different resource dimensions (CPU, GPU, memory capacity, and network bandwidth) and combining them with nonlinear function transformations and weight coefficients, we construct a resource capability vector for computing nodes. Compared to traditional linear models, this improves the ability to depict complex node differences, providing more discriminative basic data for resource matching.

[0015] 2. By constructing a task feature vector that integrates the demand characteristics of computing tasks in various resource dimensions and task attributes (such as processing latency requirements and task urgency), and matching it with the computing power node capability vector, efficient adaptation between computing tasks and computing power nodes is achieved, effectively improving the matching accuracy of task scheduling and the overall system throughput.

[0016] 3. Based on the current load status of the computing power node and the adaptability score between the computing task and the computing power node, the scheduling priority score of the computing task on the computing power node is obtained, which realizes the flexible trade-off between multiple goals of "resource matching priority" and "load balancing priority", and improves the scheduling stability and service quality in high-load environments.

[0017] 4. By introducing the task allocation status array, the task allocation status is judged and calculated before task allocation to avoid repeated scheduling or allocation conflicts, ensuring the rigor of the scheduling logic and the controllability of the scheduling strategy.

[0018] 5. After the computing task is allocated, the current load status of the computing power node is automatically updated in real time. The scheduling strategy can be reconstructed based on events such as computing power node load exceeding the limit and computing task failure, thereby realizing the dynamic adaptability and sustainable optimization capability of the resource allocation strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1This is a flow chart of a method for intelligent scheduling and load balancing of computing resources described in the present invention. DETAILED DESCRIPTION

[0020] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0021] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0022] The following describes in detail a method for intelligent scheduling and load balancing of computing resources provided by the present invention with reference to the accompanying drawings.

[0023] Refer to the attached Figure 1 , which shows a flow chart of a method for intelligent scheduling and load balancing of computing resources provided by an embodiment of the present invention, the method comprising the following steps: S1. Collect performance indicators of computing nodes in different resource dimensions and normalize them to obtain normalized performance indicators. Based on the normalized performance indicators, construct the resource capability vector of the computing node. Collect and schedule the performance indicators of each computing node in multiple resource dimensions, including: CPU ( ) performance indicators (such as main frequency, unit is ), Graphics Processing Unit ( ) performance indicators (such as floating-point computing power, units of ), memory capacity performance indicators (such as available memory capacity, unit is ) and network bandwidth performance indicators (such as available transmission rate, in The performance indicators reflect the processing power, concurrency and data throughput of the computing power node; in order to unify the standards and eliminate the impact of inconsistent dimensions, the minimum-maximum normalization method is used to standardize the performance indicators to interval, and obtain the normalized performance index; the minimum-maximum normalization method is a technical means well known to those skilled in the art and will not be described in detail here; Then, according to the resource focus of different application scenarios (such as AI reasoning, scientific computing, and edge analysis), a corresponding weight coefficient is assigned to each resource dimension. , the normalized performance indicators are weighted and combined with the corresponding weight coefficients to construct the resource capacity vector of the computing power node. The specific calculation formula is: , in, Indicates the The resource capability vector of each computing node; Respectively , GPU, memory and network bandwidth weight coefficients, whose values ​​are determined according to expert experience, and the range is ; Indicates the The normalized CPU performance index value of each computing power node, including the CPU main frequency and number of CPU cores of the current computing power node, can be selected as the CPU performance index based on the actual requirements of the computing task through expert experience. The value range is [0, 1]; Indicates the The normalized GPU performance index value of each computing power node, including the theoretical peak computing power (TFLOPS, trillion floating-point operations per second) of the GPU of the current computing power node, the number of GPU cores, the memory bandwidth, etc., can be selected as the GPU performance index based on the actual requirements of the computing task through expert experience. The value range is [0, 1]; Indicates the The normalized memory capacity index value of each computing power node, including the available memory capacity and total memory capacity of the current computing power node, can be selected as the memory capacity performance index based on the actual requirements of the computing task through expert experience. The value range is [0, 1]; Indicates the The normalized network bandwidth index value of each computing power node, including a comprehensive consideration of the current computing power node's uplink bandwidth and downlink bandwidth (the average of the two), and the maximum available transmission rate. One of these performance indicators can be selected as the network bandwidth performance indicator based on the actual needs of the computing task through expert experience. The value range is [0, 1]; Indicates the Normalized CPU performance index value of each computing power node The nonlinear adjustment factor is used to control the The influence of the CPU performance indicators of a computing power node, when When , it means that the influence of CPU performance indicators increases. When , it means that the influence of CPU performance indicators is reduced. According to the expert experience method, the value range is ; Indicates the Normalized GPU performance index value of each computing power node The nonlinear adjustment factor is used to control the influence of GPU performance indicators. When , it means that the influence of GPU performance indicators increases. When , it means that the influence of GPU performance indicators is reduced. According to the expert experience method, the value range is ; In the above formula, the introduction of exponential function can more flexibly adjust the weight coefficients of CPU and GPU. For example, for deep learning tasks, GPU is more important. In this case, we can increase To enhance the performance of GPU; for data processing tasks, CPU may be more important, using a larger This allows the CPU to be given a higher priority during scheduling; a logarithmic function is introduced to ensure that the memory capacity indicator value gradually becomes less sensitive as resources increase.

[0024] S2. By parsing the computing task execution script, configuration files, and historical resource usage records, we extract the demand characteristics of the computing task in different resource dimensions and construct a task feature vector. Through static analysis technology, a script parser built based on an open source parsing library, and a computing task historical data mining method, the computing task running scripts, configuration files, and historical resource usage records submitted by users are analyzed to obtain the specific dependency characteristics of computing tasks on computing, storage, and communication resources. For example, deep learning reasoning tasks are usually more dependent on GPUs, while log aggregation or data cleaning tasks tend to rely on network bandwidth. The static analysis technology and historical data mining method are well known to those skilled in the art and will not be described in detail here. Based on the specific dependency characteristics of computing tasks on resources, the computing tasks in each resource dimension ( , GPU, memory capacity and network bandwidth), that is, the demand characteristics of the computing task for resources in each dimension, and combined with the task attribute information such as the task urgency and processing delay requirements set by the user when submitting the task, a corresponding resource demand weight is assigned to each resource dimension. Then, the demand characteristics of each resource dimension and the corresponding resource demand weight are weighted and calculated to construct the task feature vector. The specific formula is: , in, Indicates the The task feature vector of the computing tasks to be scheduled; Represents the computing task pair , GPU, memory capacity and network bandwidth resources , that is, the resource demand weight, is set according to the expert experience method. In actual application, it can be adjusted according to the task type or scenario. The value range is ; Indicates the The CPU computing intensity required by a computing task can be selected based on the actual needs of the computing task, such as the computational complexity of the computing task, the number of required threads, and other indicators as the CPU computing intensity, and normalized to ensure that the value range is [0, 1]; Indicates the The demand for GPU computing power of a computing task can be selected based on the actual demand of the computing task, such as GPU call frequency, required video memory size, etc., as the demand for GPU computing power, and normalized to ensure that the value range is [0, 1]; Indicates the The memory requirement of each computing task can be selected based on the actual needs of the computing task, such as the maximum resident memory, memory allocation peak, etc., as the memory requirement, and normalized to ensure that the value range is [0, 1]; Indicates the The network bandwidth demand of each computing task includes indicators such as data transmission rate, network latency, and number of I / O requests per unit time. One of these indicators can be selected as the I / O bandwidth demand based on the actual needs of the computing task and normalized so that the value range is [0, 1]. Indicates the nonlinear enhancement factor for the CPU computing intensity requirement, which is used to control the The influence of CPU resource requirements of a computing task, when When the CPU resource demand influence increases, When , it means that the influence of CPU resource demand is weakened. According to the expert experience method, the value range is ; Indicates adjusting the impact of computing tasks on CPU resources based on the computing intensity required by the computing tasks; By introducing a logarithmic function to regulate the impact of computing tasks on GPU resources, we can avoid excessive impact caused by excessive demand. This function uses the sigmoid function to adjust the impact of computing tasks on network bandwidth requirements, thereby smoothing the impact of excessive demand on task scheduling.

[0025] S3. Match the resource capability vector of the computing power node with the task feature vector to obtain the fitness score between the computing task and the computing power node, and construct the fitness score matrix; Obtaining the resource capability vector of the computing power node and the task feature vector Then, the resource capability vector of the computing power node is calculated using the existing and well-known cosine similarity algorithm. and the task feature vector The cosine value of the angle between the computing task and the computing power node is used as the fitness score between the computing task and the computing power node. , to measure the matching degree of computing tasks to computing node resource capabilities; finally, the adaptability score Arrange by rows and construct a fitness score matrix between computing tasks and computing power nodes Unlike traditional distance metrics, the cosine similarity algorithm focuses more on directional consistency, that is, judging whether the resource demand structure of the computing task matches the resource capacity structure of the computing power node, and can more effectively reflect "which computing power node the computing task is suitable for running on."

[0026] S4. Collect the operating indicators of the computing power nodes, perform a comprehensive load score on the computing power nodes, and obtain the current load status of the computing power nodes; based on the current load status and fitness score of the computing power nodes, calculate the scheduling priority score and construct a scheduling priority score matrix; In the fitness score matrix After the construction is completed, the operating indicators of the computing power nodes are collected and normalized to obtain the normalized operating indicators. Based on the normalized operating indicators, the computing power nodes are comprehensively loaded and scored to obtain the current load status of each computing power node. The operating indicators include the CPU utilization rate of the computing power node, the GPU utilization rate of the computing power node, the task queue length of the computing power node, and the average task waiting time of the computing power node. The calculation formula can be defined as follows: , in, Indicates the The current load status of the computing power node, the value range is ; Indicates the Normalized CPU utilization of each computing node; Indicates the Normalized GPU utilization of each computing node; Indicates the The normalized task queue length of each computing power node (0 indicates an empty queue, 1 indicates a saturated queue); Indicates the The normalized average task waiting time of a computing power node is obtained by calculating the difference between the submission time and the start time of all computing tasks on the computing power node and taking the average; 、 、 、 It is used to measure the The current load status of each computing node The multi-factor weighted coefficients represent the relative importance of the CPU utilization, GPU utilization, task queue length, and average task waiting time of the computing power node in the comprehensive load score. They are determined by expert experience and the value range is ; Current load status of the computing power node The larger the value, the busier the computing node; The scheduling trade-off factor is introduced to perform weighted fusion of the fitness score and the current load status of the computing power node to generate the scheduling priority score of the computing task on the computing power node. The specific formula is: , in: Indicates the The computation task is Scheduling priority score on each computing node; Indicates the scheduling trade-off factor, which is used to control the weight distribution between "matching priority" and "load balancing". It is set according to expert experience and has a value range of ; Reasonably set scheduling trade-off factors The value of (such as 0.5-0.7) helps to achieve optimization between balancing the resource adaptability of computing tasks and the load level of computing nodes, thereby achieving a more robust and efficient scheduling strategy; based on all scheduling priority scores, a scheduling priority score matrix is ​​constructed to directly guide the allocation process of computing tasks.

[0027] S5. Based on the scheduling priority scoring matrix, computing tasks are allocated, and the current load status of the computing power nodes is updated in real time. The scheduling strategy is dynamically adjusted according to the execution status of the computing tasks.

[0028] Based on the scheduling priority score matrix, a one-to-one match is performed between computing tasks and computing nodes, and computing tasks are assigned to the computing nodes with the highest scheduling priority score. At the same time, the current load status of each computing node is updated in real time to maintain the overall balance of resource allocation. To avoid duplicate allocation, a task allocation status array is set. , real-time recording of the allocation of each computing task ( Indicates the total number of computing tasks to be scheduled), the initial values ​​of the task allocation status array are all set to 0, Indicates the Computational tasks have not yet been assigned. Indicates the The computing task has been successfully assigned to a computing node. The current allocation status of the computing tasks , if The computing task has been assigned to a node, then skip the current assignment operation; if the computing tasks are not assigned (i.e. ), then the tasks are assigned according to the following task assignment formula and a task assignment matrix is ​​constructed; the task assignment formula is as follows: , in, Represents an element in the task allocation matrix, indicating whether the first The computing task is assigned to Computing nodes, 1 means allocation, 0 means no allocation; Indicates the The scheduling priority score of a computing task on all computing nodes The maximum value in Indicates a traversal variable, representing one of all computing power nodes, and is used to exhaustively enumerate all computing power nodes, that is, to determine which node is most "suitable" for the current task; Indicates the node index, indicating the computing power nodes; Indicates the task index, indicating the Computing tasks to be scheduled; Indicates the The computation task is Scheduling priority score on computing power nodes The maximum value of the scheduling priority scores of the computing task on all computing nodes. The above task allocation method allocates the computing task to the computing node with the highest scheduling priority score, ensuring the consistency and uniqueness of the scheduling, and avoiding task conflicts or resource competition. After the task scheduling is completed, the current load status of each computing node is updated in real time , and trigger rescheduling conditions (such as task failure) to maintain a dynamic balance in global resource distribution.

[0029] In summary, a method for intelligent scheduling and load balancing of computing resources has been completed.

[0030] The order in which the embodiments of the invention are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0031] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0032] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A method for intelligent scheduling and load balancing of computing resources, characterized in that: The following steps are involved: S1. Collect performance indicators of computing nodes in different resource dimensions and normalize them to obtain normalized performance indicators. Based on the normalized performance indicators, construct the resource capability vector of the computing node. S2. By parsing the computing task execution script, configuration files, and historical resource usage records, we extract the demand characteristics of the computing task in different resource dimensions and construct a task feature vector. S3. Match the resource capability vector of the computing power node with the task feature vector to obtain the fitness score between the computing task and the computing power node, and construct the fitness score matrix; S4. Collect the operating indicators of the computing power nodes, perform a comprehensive load score on the computing power nodes, and obtain the current load status of the computing power nodes; based on the current load status and fitness score of the computing power nodes, calculate the scheduling priority score and construct a scheduling priority score matrix; S5. Based on the scheduling priority scoring matrix, computing tasks are allocated, and the current load status of the computing power nodes is updated in real time. The scheduling strategy is dynamically adjusted according to the execution status of the computing tasks.

2. The method for intelligent scheduling and load balancing of computing resources according to claim 1, characterized in that: Said S1 specifically includes: Assign a weight coefficient to each resource dimension, combine the exponential function and the logarithmic function, perform a weighted combination of the normalized performance indicators and the weight coefficients, and construct the resource capacity vector of the computing power node.

3. The method for intelligent scheduling and load balancing of computing resources according to claim 1, characterized in that: Said S2 specifically includes: Analyze the computing task running scripts, configuration files and historical resource usage records to obtain the specific dependency characteristics of the computing task on computing, storage and communication resources; based on the specific dependency characteristics of the computing task on the said resources, extract the demand of the computing task for resources of each dimension as the demand characteristics of the computing task in the resource dimension.

4. The method for intelligent scheduling and load balancing of computing resources according to claim 3, characterized in that: Said S2 specifically includes: The resource demand weight is set for each resource dimension. The exponential function and logarithmic function are combined to perform a weighted combination of the demand characteristics of the computing task in the resource dimension and the resource demand weight to construct the task feature vector.

5. The method for intelligent scheduling and load balancing of computing resources according to claim 1, characterized in that: Said S4 specifically includes: The operating indicators of the computing power nodes are collected and normalized to obtain normalized operating indicators; the operating indicators include the CPU utilization rate of the computing power nodes, the GPU utilization rate of the computing power nodes, the task queue length of the computing power nodes, and the average task waiting time of the computing power nodes.

6. The method for intelligent scheduling and load balancing of computing resources according to claim 5, characterized in that: Said S4 specifically includes: A weighted coefficient is set for each normalized operating indicator. By taking the weighted sum of the normalized operating indicators, a comprehensive load score is given to the computing power node to obtain the current load status of the computing power node.

7. The method for intelligent scheduling and load balancing of computing resources according to claim 6, characterized in that: Said S4 specifically includes: A scheduling trade-off factor is introduced to perform weighted fusion of the fitness score and the current load status of the computing power node to generate a scheduling priority score for the computing task on the computing power node.

8. The method for intelligent scheduling and load balancing of computing resources according to claim 1, characterized in that: Said S5 specifically includes: Based on the scheduling priority score matrix and the task allocation status array, the computing tasks are matched with the computing power nodes with the highest scheduling priority scores to achieve task allocation and update the current load status of each computing power node in real time.

Images