An intelligent load balancing method based on zero-sample hybrid expert model

Abstract

The present invention discloses an intelligent load balancing method based on a zero-sample hybrid expert model, comprising the steps of: creating a task spectrum mapping relationship through zero-sample learning, generating virtual labels, refining the requirements of new tasks into resource requirement feature spectra through multi-dimensional feature cross-projection, and predicting the multi-dimensional resource requirements of new load balancing tasks; establishing a multi-perspective hybrid expert dynamic integration mechanism, configuring multi-perspective hybrid collaborative experts for tasks, applying a combination of several expert models to different demand dimensions, creating virtual image samples of new tasks, simulating resource requirements and their load effects in a virtual environment, and generating an initial resource allocation plan; in the face of dynamic load fluctuations, distributing task cells to nodes with sufficient resources; the method improves the robustness and efficiency of the system, simplifies the implementation process, and has broad application prospects and practical value.

Description

Technical Field

[0001] The invention relates to the field of network load balancing, and in particular to an intelligent load balancing method based on a zero-sample hybrid expert model. Background Art

[0002] Under the wave of digital transformation, the development of cloud computing and big data technology has provided unprecedented computing power and storage resources for data centers. As enterprises' demand for real-time data processing and high availability increases, the workloads of data centers have become more complex and diverse. These workloads include not only traditional batch processing and transaction processing tasks, but also a variety of emerging applications such as real-time stream processing, machine learning model training, and image processing. An effective load balancing strategy can ensure that computing resources are fully utilized, thereby improving system performance, reducing latency, and improving user experience. In the current technical context, traditional load balancing algorithms, such as polling, weighted polling, and least connections, often cannot adapt to dynamically changing workloads due to their static characteristics. These methods mainly rely on historical traffic data to make decisions, but in a rapidly changing environment, historical data may not be representative enough. Therefore, it is particularly important to develop an intelligent load balancing technology that can adapt to emerging tasks and reduce dependence on historical data.

[0003] The difficulties in the current research of load balancing tasks mainly include the following aspects. First, the workload of modern data centers may change significantly in a short period of time, such as peak periods of user access and traffic surges caused by specific events. This requires the load balancing algorithm to be able to monitor in real time and quickly adjust resource allocation to maintain the stability and performance of the system. Second, data scarcity and unknown tasks: In many cases, especially when new services are launched or new task types appear, there is a lack of historical data to support effective load balancing decisions. For these unknown tasks, traditional algorithms are usually unable to accurately judge their resource requirements, resulting in performance degradation. Processing of diverse tasks: Current data centers are faced with the simultaneous processing of different types of tasks, including compute-intensive, memory-intensive, and I / O-intensive tasks. How to dynamically select the optimal load balancing strategy among multiple tasks and ensure efficient processing of tasks is a complex challenge. Finally, existing load balancing models often rely on static training processes and lack sufficient adaptability when faced with new and unseen tasks. This makes these models unstable in real environments.

[0004] Current research focuses on load balancing algorithms based on machine learning, such as reinforcement learning and deep learning models. These algorithms learn load patterns from historical data and are able to optimize resource allocation under known conditions. However, these methods usually require a large amount of labeled data for training and have insufficient adaptability to unknown tasks. Hybrid expert models have been proposed as an effective solution that can dynamically select appropriate models for processing according to the task type, but there are still certain limitations when facing unseen tasks. These limitations mainly include: limitations of relying on historical data, insufficient model adaptability, complexity and computational cost, inability to handle very few samples, and lack of comprehensive intelligent solutions. Many load balancing algorithms based on machine learning rely on a large amount of historical data for training in order to predict future workloads. However, these methods cannot adapt effectively when faced with new and unseen task types due to the lack of relevant training samples. This leads to a significant decrease in the system's response speed and performance when new services are launched or traffic surges. Insufficient model adaptability: Existing hybrid expert models perform well when processing known tasks, but the model's adaptability is often insufficient for unseen tasks. Although these models can select appropriate experts for processing, they may lead to wrong resource allocation decisions and affect the overall performance of the system when there is a lack of understanding of the characteristics of new tasks. Complexity and computational cost: Although adaptive load balancing algorithms can dynamically adjust resource allocation, their implementation process is usually complex and involves real-time monitoring and data analysis. This not only increases the computational cost of the system, but may also cause delays, making the load balancing process inefficient and unable to meet application scenarios with high real-time requirements. Unable to handle very few samples: For some emerging application scenarios, especially when new task types appear, they often face very few samples or even no samples. Although the existing zero-shot learning technology can infer the characteristics of unseen tasks to a certain extent, it lacks effective combination with load balancing strategies, making it difficult to play its due role in practical applications. Lack of comprehensive intelligent solutions: Current solutions often focus on optimization of one aspect. For example: task type, resource allocation, etc., lack of intelligent mechanisms that comprehensively consider all factors. This leads to the inability to fully utilize system resources when multiple tasks are carried out simultaneously, resulting in resource waste and delays in task processing. Therefore, current load balancing methods have problems such as insufficient adaptability when facing unknown task types and dynamically changing workloads, over-reliance on historical data, and high computational costs that need to be urgently addressed. Summary of the invention

[0005] To this end, it is necessary to provide a method that integrates hybrid expert models and zero-sample learning technology to achieve feature inference of unseen tasks and dynamically adjust server resource allocation, thereby improving the system's responsiveness and load balancing effect when new tasks appear. This intelligent load balancing solution not only improves the robustness and efficiency of the system, but also simplifies the implementation process. It has broad application prospects and practical value.

[0006] To achieve the above object, the inventor provides an intelligent load balancing method based on a zero-sample hybrid expert model, comprising the steps of:

[0007] S1, creates task spectrum mapping relationships through zero-shot learning, maps new task features to similar known task distributions, generates virtual labels, refines new task requirements into resource requirement feature spectra through multi-dimensional feature cross-projection, and predicts multi-dimensional resource requirements for new load balancing tasks;

[0008] S2, through the dynamic combination of expert models from multi-dimensional task perspectives, a dynamic integration mechanism of multi-perspective hybrid experts is established. Based on the resource demand feature spectrum, multi-perspective hybrid collaborative experts are configured for tasks, and several expert model combinations are applied to different demand dimensions;

[0009] S3, creates virtual image samples of new tasks through resource demand feature spectra, simulates resource demand and its load effect in a virtual environment, and generates an initial resource allocation plan;

[0010] S4, in the face of dynamic load fluctuations, if the load of a node exceeds the threshold, the expert model performs cellular decomposition, decomposing the high-load task into several task cells, and distributes the task cells to nodes with sufficient resources according to the load status.

[0011] As a preferred embodiment of the present invention, step S1 comprises:

[0012] S101, for new load balancing tasks , using the feature extractor Convert it into a high-dimensional embedding representation , capture the complex characteristics of the task, and then calculate the new load balancing task through the kernel function K With known task characteristics Similarity, find the matching task embedding , used for subsequent analysis and label generation, the expression is:

[0013] ;

[0014] S102, embedding according to the matched task , use the mapping function G to generate virtual labels, and project the virtual labels into the multidimensional feature spectrum through the projection matrix P to generate the resource requirement feature spectrum of the new task , including computing requirements Storage requirements and network requirements , the expression is:

[0015] ;

[0016] S103, by using the resource demand characteristic spectrum Input into the deep reasoning network H to predict the multi-dimensional resource requirements of new load balancing tasks , the expression is:

[0017] .

[0018] As a preferred embodiment of the present invention, step S2 includes:

[0019] S201, multi-dimensional resource requirements of new load balancing tasks As input, a pool of candidate expert models is generated through the task feature analysis function M , the expression is:

[0020] ;

[0021] S202, the resource demand characteristic spectrum Decomposition into a computational perspective , I / O perspective and bandwidth perspective , and assign weights based on these perspectives to determine the expert weights corresponding to each perspective , the expression is:

[0022] ;

[0023] in, It is to normalize the resource requirements of each perspective;

[0024] S203, based on the expert weights of the task at different perspectives , combined candidate expert model Collaborative applications to meet task requirements and form comprehensive processing results of tasks , the expression is:

[0025] ;

[0026] in, Represents different expert model weights, represents different expert models, and N represents the total number of expert models;

[0027] Each expert model Perform task processing on its focused perspective, and output results according to the assigned expert model weights Perform weighted summation to form a comprehensive processing result of the task;

[0028] S204, the system introduces a dynamic feedback mechanism to monitor the execution effect of tasks in real time and and expected resource allocation The difference between Adjusting the weights of expert models , the expression is:

[0029] ;

[0030] Among them, the parameters Used to control the amplitude and speed of adjustment. Represents the adjusted expert weight.

[0031] As a preferred embodiment of the present invention, step S3 includes:

[0032] S301, using the resource demand characteristic spectrum , through the mapping function Create a sample virtual image of a new task , the expression is:

[0033] ;

[0034] S302, using the generated virtual image sample in the virtual environment Test multiple expert combinations, each expert model Representing different types of load handling capabilities, by comparing virtual image samples Conduct simulation and generate simulation plans , the expression is:

[0035] ;

[0036] in, represents the weight of matching the perspective of task requirements;

[0037] S303, using loss function Compare all simulation scenarios and target resource requirements The difference between The smallest solution , determine the initial resource allocation plan, the expression is;

[0038] .

[0039] As a preferred embodiment of the present invention, step S4 includes:

[0040] S401, combined with static threshold Detect node load and historical load trends, predict models for intelligent triggering, and use time series model T to predict future node loads , used to deal with load fluctuations in advance, the expression is:

[0041] ;

[0042] in, Indicates historical load trends, Indicates the trigger condition, which is used to determine whether a response operation is required. The value is Boolean True or False. If the condition is met, , then the trigger condition is True, and the time series model T is used to predict the future node load , if the condition is not met, no operation will be performed;

[0043] S402, high load tasks Divide into several task cells , intelligently adjust according to resource types and dependencies, and separate high-load tasks during the segmentation process. Based on its cell characteristic spectrum and the execution context Intelligent segmentation and analysis through expert models and context-aware analysis functions The resource requirements of each task cell are marked and optimized, and the expression is:

[0044] ;

[0045] in, represents the resource dependency feature vector of the i-th task cell;

[0046] S403, combining the node load level, network delay and real-time feedback, optimizes the task allocation decision in a multi-dimensional way, and uses a cell-level adaptive distribution algorithm to allocate task cells to target nodes with sufficient resources and low delay. , real-time feedback loop It is used to monitor the distribution effect and dynamically adjust the distribution strategy. The expression is:

[0047] ;

[0048] in, represents the current load level of the jth node, Indicates that network delay is used as a weight factor for node selection, through the parameter Controlling the impact of network delay in task allocation decisions, feedback function Used to evaluate the actual resource usage of node j and expected resource allocation The difference between Represents the weight of resource feedback error.

[0049] Different from the prior art, the above technical solution has the following beneficial effects:

[0050] (1) This method intelligently adapts to unknown task types: In the case of a lack of historical data for new services or traffic patterns at the initial launch, this method uses zero-sample learning technology to infer the resource demand characteristics of new tasks without training samples, so that the load balancing algorithm can intelligently identify and adapt to unseen task types. Through this method, the system can quickly respond to different types of load requests and reduce performance fluctuations caused by the launch of new tasks.

[0051] (2) This method dynamically allocates server resources in real time: In a complex load environment, this method uses a hybrid expert model to process known tasks and combines zero-sample learning technology to predict the load requirements of unknown tasks, thereby achieving dynamic resource allocation. By adjusting resources in real time, the load balancing of the server cluster at different times is always efficient, ensuring that the load of each server tends to be balanced, and avoiding resource bottlenecks caused by static allocation;

[0052] (3) This method improves resource utilization and system efficiency: This method comprehensively considers the resource demand characteristics of various tasks such as computing-intensive, storage-intensive, and I / O-intensive, and avoids resource idleness and overuse through an intelligent allocation mechanism, thereby significantly improving resource utilization; by effectively scheduling server resources, the system can achieve precise matching in resource allocation, maximizing the computing and storage capabilities of a data center or server cluster;

[0053] (4) This method simplifies the deployment and maintenance of load balancing: This method designs a simple intelligent resource management process so that the implementation of load balancing no longer relies too much on historical data, reducing the need for complex data preprocessing and specific configuration. This method can automatically make adaptive adjustments based on the current load type and task characteristics, making the load balancing algorithm easier to deploy and maintain, especially suitable for large-scale data centers and multi-task concurrent environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a framework diagram of the method described in the specific implementation method. DETAILED DESCRIPTION

[0055] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.

[0056] like Figure 1 As shown, this embodiment provides an intelligent load balancing method based on a zero-sample hybrid expert model, comprising the steps of:

[0057] S1, creates a task spectrum mapping relationship through zero-sample learning, maps new task features to similar known task distributions, generates virtual labels, and refines the requirements of new tasks into resource requirement feature spectra such as computing, storage, and network requirements through multi-dimensional feature cross-projection, and predicts the multi-dimensional resource requirements of new load balancing tasks; this process gives this embodiment an intelligent initial reasoning capability, accurately predicts the resource requirements of new tasks when there is no historical data, realizes rapid identification of the load type of new tasks, and forms a basic basis for resource allocation.

[0058] S2, through the dynamic combination of expert models from multi-dimensional task perspectives, establishes a dynamic integration mechanism of multi-perspective hybrid experts. This integration mechanism breaks the limitation of the single selection of traditional static expert models. Based on the spectrum of resource demand characteristics, it configures multi-perspective hybrid collaborative experts for tasks and applies several expert model combinations to different demand dimensions, namely, computationally intensive, I / O intensive, and bandwidth sensitive. Its core innovation lies in its dynamic adaptation capability, which can adaptively integrate multiple experts, so that each task can be processed by the optimal model combination in different dimensions, ensuring that load balancing decisions are highly flexible and targeted.

[0059] S3, through the resource demand characteristic spectrum, creates virtual image samples of new tasks, simulates resource demand and its load effect in a virtual environment, and generates an initial resource allocation plan; that is, a virtual pre-loaded image method is proposed to simulate the optimal resource allocation of tasks before the real load, and before the task is officially run, the virtual pre-loaded image sample will test different expert combination resource configurations to generate a set of optimal initial resource allocation plans, which greatly reduces the subsequent real-time adjustments; this unique image method not only ensures the accuracy of resource allocation, but also effectively reduces resource conflicts and task delays by optimizing resource demand in advance;

[0060] S4, in the face of dynamic load fluctuations, if the load of a node exceeds the threshold, the expert model will perform cellular decomposition, decompose the high-load task into several task cells, and distribute the task cells to nodes with sufficient resources according to the load status; that is, in the face of dynamic load fluctuations, when it is detected that the load of a certain node exceeds the threshold, the hybrid expert model will perform "cellular" decomposition, decompose the high-load task into multiple "task cells", and distribute them to nodes with sufficient resources according to the load status. Different from traditional task migration, this embodiment uses an expert model to analyze the resource dependence and transmission requirements of the task, and "fragment" the task to the most suitable small-load node, achieving the dual effects of self-healing adjustment and refined distribution; this cellular-level task distribution not only optimizes the system load balancing, but also enhances the system's resilience to sudden loads.

[0061] In the specific implementation process of step S1 in the above embodiment, the following steps are included:

[0062] S101, for new load balancing tasks , the system uses a feature extractor Convert it into a high-dimensional embedding representation , capturing the complex features of the task. This transformation can be achieved using deep neural networks, such as variational autoencoders (VAE) or deep convolutional networks, to obtain rich expressions of the task in terms of resource requirements and traffic characteristics. Then, the kernel function K is used to calculate With known task characteristics similarity, so as to find the most matching task embedding , used for subsequent analysis and label generation, the expression is:

[0063] ;

[0064] S102, based on the matched embedding , use the mapping function G to generate virtual labels. Through the projection matrix P, the virtual labels are projected into the multidimensional feature spectrum to generate the resource demand feature spectrum , including computing requirements Storage requirements and network requirements This process enables the system to refine the resource requirements of new tasks and provide a basis for load balancing decisions, expressed as:

[0065] ;

[0066] S103, by using the resource demand characteristic spectrum Input into the deep reasoning network H to predict the multi-dimensional resource requirements of new load balancing tasks The deep reasoning network H is able to learn complex nonlinear relationships between input features and provide accurate resource demand estimates. , identify the load type of the task, such as: computing intensive, storage intensive or network sensitive, to optimize resource allocation and load balancing strategy, the expression is:

[0067] ;

[0068] In the specific implementation process of step S2 in the above embodiment, the following steps are included:

[0069] S201, after completing the new load balancing task After feature extraction, embedding mapping, virtual label generation and resource demand prediction, the multi-dimensional resource requirements of the new load balancing task are As input, a pool of candidate expert models is generated through the task feature analysis function M , the expression is:

[0070] ;

[0071] Based on multi-dimensional resource requirements , analyze the resource requirement characteristics of the task, such as computing, storage, and network bandwidth, to select the appropriate expert model to ensure that the task can obtain the best processing effect from the expert combination.

[0072] S202, the resource demand characteristic spectrum Decomposition into a computational perspective , I / O perspective and bandwidth perspective , and assign weights based on these perspectives to determine the expert weights corresponding to each perspective ,This step decomposes the task from a multi-dimensional perspective, identifies the priorities of different resource requirements, and dynamically allocates the weights of each expert model so that it can play the best role in its respective areas of expertise. The expression is:

[0073] ;

[0074] in, The resource requirements of each perspective are normalized to ensure that the contribution of each perspective is within a standardized range in the weight allocation, making it easier to compare and weight.

[0075] S203, based on the weight of the task at different perspectives , system combination candidate expert model Collaborative applications to meet task requirements, task output The expression is:

[0076] ;

[0077] in, Represents different expert model weights, represents different expert models, and N represents the total number of expert models. Perform task processing on its focused perspective and output results according to the assigned weights The weighted sum is performed to form a comprehensive processing result of the task. This step realizes the coordination of multiple expert models to meet the complex task requirements of different resource dimensions.

[0078] S204, the system introduces a dynamic feedback mechanism to monitor the execution effect of tasks in real time and and expected resource allocation The difference between Adjust expert weights :

[0079] ;

[0080] This feedback mechanism dynamically adjusts the weight of the expert model by comparing the actual and expected resource usage, ensuring that the system can adaptively adjust the strategy in the face of real-time load changes and maintain efficient and stable load balancing effects. Used to control the amplitude and speed of adjustment. For the new expert weights.

[0081] In the specific implementation process of step S3 in the above embodiment, the following steps are included:

[0082] S301, using the resource demand feature spectrum generated from the task feature extraction and mapping process , through the mapping function Create a sample virtual image of a new task . Mapping function It is based on deep generative models, such as Generative Adversarial Networks (GANs) or Variational Autoencoders (VAEs), to capture the complex characteristics of tasks in different resource dimensions, including computing requirements, storage requirements, and network requirements. The global characteristics of the new task are simulated to provide data support for subsequent simulation tests. The expression is:

[0083] ;

[0084] S302, the system uses the generated virtual image sample in the virtual environment Test multiple expert combinations. Each expert model Represents different types of load processing capabilities, such as computationally intensive, I / O intensive, or bandwidth sensitive. By simulating image samples, the output of each solution is is the weighted result of the expert model combination, where is the weight that matches the task requirement perspective. This step simulates the response and resource allocation effects of different expert model combinations under virtual tasks to help the system identify the optimal configuration strategy. The expression is:

[0085] ;

[0086] S303, using loss function Compare all simulation scenarios and target resource requirements The loss function can comprehensively consider factors such as resource conflicts, task delays, and system load balancing. The smallest solution , the optimal initial resource allocation scheme is determined. Applying this scheme before the actual task runs can greatly reduce the need for real-time adjustment and improve the response efficiency and stability of the overall system. The expression is:

[0087] ;

[0088] In the specific implementation process of step S4 in the above embodiment, the following steps are included:

[0089] S401, the system is not only based on static thresholds Detect node loads and combine historical load trends and forecasting models for intelligent triggering to predict potential overloads; use time series model T to predict future node loads , to ensure that load fluctuations are dealt with in advance, the expression is:

[0090] ;

[0091] in, Indicates historical load trends, Indicates the trigger condition, which is used to determine whether a certain response operation is required. The value is Boolean True or False. If the condition is met, , then the trigger condition is True, and the time series model T is used to predict the future node load If the condition is not met, the trigger condition is False and no action is taken. By combining real-time load detection and predictive analysis, the system triggers task cellular decomposition when it detects that the current load or predicted load is about to exceed the threshold. This innovative mechanism allows the system to take active adjustment measures before the load exceeds the tolerance range.

[0092] S402, task decomposition not only divides high-load tasks Divide into multiple cells , and intelligently adjusts according to resource types and dependencies, using expert models and context-aware analysis functions The resource requirements of each task cell are marked and optimized, and the expression is:

[0093] ;

[0094] The decomposition process divides the tasks into their characteristic spectra and runtime context Intelligent subdivision enables each task cell to better adapt to the specific resource conditions of different nodes when allocated; this approach ensures the accuracy and adaptability of task decomposition; It represents the resource dependency feature vector of the i-th task cell, which is a description of the specific resource requirements and dependencies required for each subdivided task cell during execution.

[0095] S403 combines the node load level, network latency and real-time feedback to optimize task allocation decisions in a multi-dimensional manner. Unlike traditional methods, it not only relies on current static information, but also introduces a dynamic feedback mechanism to better adapt to changing load conditions. It uses a cell-level adaptive distribution algorithm to allocate task cells to target nodes with sufficient resources and low latency. ; Real-time feedback loop It is used to monitor the distribution effect and dynamically adjust the distribution strategy. The expression is:

[0096] ;

[0097] in, represents the current load level of the jth node. Nodes with higher loads will increase the value of the formula, making them less preferred in allocation; Indicates that network delay is used as a weight factor in node selection, through the parameter Control the impact of network delay in task allocation decisions, the larger A value of means that network latency is more important for allocation decisions; the feedback function Evaluate the actual resource usage of node j and expected resource allocation The difference between The weight of the resource feedback error is used to control the importance of resource allocation error to task allocation decision. The difference indicates the effectiveness of the current task allocation, such as resource contention, inefficient utilization, or load imbalance.

[0098] In order to prove the effectiveness of this embodiment, a public data set is used for verification. Specifically, the GoogleCluster Data data set is a real cloud computing cluster data set that contains the tasks and resource usage of Google data centers. This data set can be used to simulate different load scenarios in task allocation and resource management, and verify the performance of the new method in a real environment. The performance comparison experimental results of different network anomaly detection methods are shown in Table 1.

[0099] Table 1: Experimental results of performance comparison of network anomaly detection methods

[0100]

[0101] Experimental results show that this method has obvious advantages in all key indicators of load balancing and resource management, significantly surpassing traditional load balancing algorithms. First, the resource utilization rate reaches 97%, which is much higher than the polling algorithm (75%), the least connection algorithm (80%) and the weighted polling algorithm (82%). This shows that this method can achieve near-maximum resource utilization under high load conditions, so that system resources are fully and effectively used, thereby reducing resource waste and improving overall efficiency.

[0102] Secondly, the load balancing degree (CV) of this method is 0.13, which is much lower than the polling algorithm (0.4), the least connection algorithm (0.3) and the weighted polling algorithm (0.25), indicating that it distributes the load more evenly in the cluster, effectively avoiding the situation where a single node is overloaded or other node resources are idle. This balanced distribution ensures the stability and continuous efficient operation of the system, especially under dynamic and peak loads.

[0103] In terms of response time, the 50 milliseconds of this method is much lower than the polling algorithm (120 milliseconds), the least connection algorithm (100 milliseconds), and the weighted polling algorithm (95 milliseconds). This means that in the process from user request submission to task processing, this method can respond at a faster speed, greatly reducing latency, improving the real-time performance of the system and user experience, which is especially important for tasks that require fast response and application scenarios with high real-time requirements.

[0104] The most impressive thing is the throughput (tasks / second). This method reaches 500 tasks / second, which is a significant improvement over the polling algorithm (200 tasks / second), the least connection algorithm (220 tasks / second), and the weighted polling algorithm (230 tasks / second). The high throughput shows that this method has excellent performance in handling concurrent tasks and coping with high-load environments, and can complete more tasks in a short time, meeting the needs of large-scale data centers and high-demand environments.

[0105] In general, this method achieves comprehensive optimization in four key indicators: resource utilization, load balancing, response time, and throughput. Its design not only improves the overall efficiency of the system, but also significantly enhances the adaptability and stability of the system in high-concurrency and dynamic load environments through efficient load distribution and rapid response capabilities. This makes this method an extremely efficient, reliable, and suitable load balancing solution for a variety of application scenarios.

[0106] It should be noted that, although the above embodiments have been described in this article, the patent protection scope of the present invention is not limited thereby. Therefore, based on the innovative concept of the present invention, changes and modifications made to the embodiments described herein, or equivalent structures or equivalent process changes made using the contents of the present invention specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included in the patent protection scope of the present invention.

Claims

1. An intelligent load balancing method based on zero-sample hybrid expert model, characterized in that: Includes steps: S1, creates task spectrum mapping relationships through zero-shot learning, maps new task features to similar known task distributions, generates virtual labels, refines new task requirements into resource requirement feature spectra through multi-dimensional feature cross-projection, and predicts multi-dimensional resource requirements for new load balancing tasks; S2, through the dynamic combination of expert models from multi-dimensional task perspectives, a dynamic integration mechanism of multi-perspective hybrid experts is established. Based on the resource demand feature spectrum, multi-perspective hybrid collaborative experts are configured for tasks, and several expert model combinations are applied to different demand dimensions; S3, creates virtual image samples of new tasks through resource demand feature spectra, simulates resource demand and its load effect in a virtual environment, and generates an initial resource allocation plan; S4, in the face of dynamic load fluctuations, if the load of a node exceeds the threshold, the expert model performs cellular decomposition, decomposing the high-load task into several task cells, and distributes the task cells to nodes with sufficient resources according to the load status; The step S1 comprises: S101, for new load balancing tasks , using the feature extractor Convert it into a high-dimensional embedding representation , capture the complex characteristics of the task, and then calculate the new load balancing task through the kernel function K With known task characteristics Similarity, find the matching task embedding , used for subsequent analysis and label generation, the expression is: ； S102, embedding according to the matched task , use the mapping function G to generate virtual labels, and project the virtual labels into the multidimensional feature spectrum through the projection matrix P to generate the resource requirement feature spectrum of the new task , including computing requirements Storage requirements and network requirements , the expression is: ； S103, by using the resource demand characteristic spectrum Input into the deep reasoning network H to predict the multi-dimensional resource requirements of new load balancing tasks , the expression is: ； The step S2 comprises: S201, multi-dimensional resource requirements of new load balancing tasks As input, a pool of candidate expert models is generated through the task feature analysis function M , the expression is: ； S202, the resource demand characteristic spectrum Decomposition into a computational perspective , I / O perspective and bandwidth perspective , and assign weights based on these perspectives to determine the expert weights corresponding to each perspective , the expression is: ； in, It is to normalize the resource requirements of each perspective; S203, based on the expert weights of the task at different perspectives , combined candidate expert model Collaborative applications to meet task requirements and form comprehensive processing results of tasks , the expression is: ； in, Represents different expert model weights, represents different expert models, and N represents the total number of expert models; Each expert model Perform task processing on its focused perspective, and output results according to the assigned expert model weights Perform weighted summation to form a comprehensive processing result of the task; S204, the system introduces a dynamic feedback mechanism to monitor the execution effect of tasks in real time and and expected resource allocation The difference between Adjusting the weights of expert models , the expression is: ； Among them, the parameters Used to control the amplitude and speed of adjustment. Represents the adjusted expert weight.

2. The intelligent load balancing method based on zero-sample hybrid expert model according to claim 1 is characterized in that: The step S3 comprises: S301, using the resource demand characteristic spectrum , through the mapping function Create a sample virtual image of a new task , the expression is: ； S302, using the generated virtual image sample in the virtual environment Test multiple expert combinations, each expert model Representing different types of load handling capabilities, by comparing virtual image samples Conduct simulation and generate simulation plans , the expression is: ； in, represents the weight that matches the perspective of task requirements; S303, using loss function Compare all simulation scenarios and target resource requirements The difference between The smallest solution , determine the initial resource allocation plan, the expression is; 。 3. The intelligent load balancing method based on zero-sample hybrid expert model according to claim 2 is characterized in that: Step S4 includes: S401, combined with static threshold Detect node load and historical load trends, predict models for intelligent triggering, and use time series model T to predict future node loads , used to deal with load fluctuations in advance, the expression is: ； in, Indicates historical load trends, Indicates the trigger condition, which is used to determine whether a response operation is required. The value is Boolean True or False. If the condition is met, , then the trigger condition is True, and the time series model T is used to predict the future node load , if the condition is not met, no operation will be performed; S402, high load tasks Divide into several task cells , intelligently adjust according to resource types and dependencies, and separate high-load tasks during the segmentation process. Based on its cell characteristic spectrum and the execution context Intelligent segmentation and analysis through expert models and context-aware analysis functions The resource requirements of each task cell are marked and optimized, and the expression is: ； in, represents the resource dependency feature vector of the i-th task cell; S403, combining the node load level, network delay and real-time feedback, optimizes the task allocation decision in a multi-dimensional way, and uses a cell-level adaptive distribution algorithm to allocate task cells to target nodes with sufficient resources and low delay. , real-time feedback loop It is used to monitor the distribution effect and dynamically adjust the distribution strategy. The expression is: ； in, represents the current load level of the jth node, Indicates that network delay is used as a weight factor for node selection, through the parameter Controlling the impact of network delay in task allocation decisions, feedback function Used to evaluate the actual resource usage of node j and expected resource allocation The difference between Represents the weight of resource feedback error.

Images