Load balancing method and system in cloud computing environment

Abstract

The invention discloses a load balancing method and system in a cloud computing environment, and relates to the technical field of cloud computing, and the method comprises the steps: receiving a service request of a target user, obtaining the user portrait information of the target user, and determining a target model corresponding to the service request according to the user portrait information, obtaining model deployment information and a current load value of each node in the current service pool, calculating a matching degree score of the service request and each node according to the user portrait information, the model deployment information and the corresponding current load value, determining a target node according to the matching degree score, and if the matching degree score of the target node is higher than a preset score threshold, sending the target node to the server; if yes, scheduling the service request to a target node, otherwise, deploying a target model in the current service pool, and scheduling the service request to a compatible node in the current service pool; the method has the beneficial effects that the service request scheduling accuracy can be improved, the resource utilization rate is optimized, and the service continuity in a model unavailable scene is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cloud computing, and in particular to a load balancing method and system in a cloud computing environment. BACKGROUND

[0002] With the continuous development of cloud computing and artificial intelligence technologies, more and more enterprises migrate their core businesses to the cloud, forming a SaaS service mode characterized by "multi-tenant shared cloud infrastructure". In this mode, it has become a common practice to deploy dedicated models or specific versions of models for different customers to meet the differences in business logic, algorithm strategy or compliance requirements. In this context, multiple model versions are often deployed in parallel in the cloud platform for the same service, and the model paths required by different user requests also differ significantly, thus promoting the evolution of model scheduling and inference services towards refinement and customization.

[0003] The traditional load balancing strategy in the cloud computing environment mainly relies on the number of requests, CPU / GPU utilization, response time and other underlying resource indicators for task scheduling, lacking understanding of user personalized needs and model semantic levels. In the current complex scenario of high concurrency access, model heterogeneous deployment and tenant customization needs, simply relying on general resource indicators for scheduling has been difficult to guarantee the scheduling accuracy and response quality of the service, especially under the premise of highly heterogeneous user portraits. The model types, versions and inference paths relied on by user requests differ significantly. If the matching relationship between user features and model versions is not fully considered, it is easy to cause the request to be scheduled to a node with incompatible or semantically mismatched model versions, leading to inference errors, increased delays or resource waste, and it is difficult to meet the actual needs of large-scale multi-tenant cloud platforms for intelligent and accurate load balancing.

[0004] Therefore, a load balancing method and system in a cloud environment are proposed. SUMMARY

[0005] In view of the above prior art situation, the present application is proposed. The embodiments of the present application provide a load balancing method in a cloud computing environment, which can improve the scheduling accuracy of service requests, optimize resource utilization, and enhance service continuity in the case of model unavailability.

[0006] According to an aspect of the present application, a load balancing system in a cloud computing environment is provided, comprising: receiving a service request of a target user; obtaining user portrait information of the target user; determining a target model corresponding to the service request according to the user portrait information; obtaining model deployment information and current load values of each node in a current service pool; calculating a matching degree score of the service request and each node according to the user portrait information, model deployment information and corresponding current load values, and taking a node with the highest matching degree score as a target node; determining whether the matching degree score of the target node is higher than a preset score threshold, if yes, scheduling the service request to the target node, otherwise, deploying the target model on a standby node of the current service pool, and scheduling the service request to a compatible node, the compatible node being a node with the highest semantic similarity to the target model and a current load value lower than a preset load threshold among the nodes.

[0007] According to another aspect of the present application, a load balancing method and system in a cloud computing environment are provided, comprising: a request receiving module, receiving a service request of a target user; a portrait obtaining module, obtaining user portrait information of the target user; a model determining module, determining a target model corresponding to the service request according to the user portrait information; an information obtaining module, obtaining model deployment information and current load values of each node in a current service pool; a target node determining module, calculating a matching degree score of the service request and each node according to the user portrait information, model deployment information and corresponding current load values, and taking a node with the highest matching degree score as a target node; a decision module, determining whether the matching degree score of the target node is higher than a preset score threshold, if yes, scheduling the service request to the target node, otherwise, deploying the target model on a standby node of the current service pool, and scheduling the service request to a compatible node, the compatible node being a node with the highest semantic similarity to the target model and a current load value lower than a preset load threshold among the nodes.

[0008] According to another aspect of the present application, an electronic device is provided, comprising a memory and a processor, the memory being used to store computer executable instructions, and the processor being used to execute the computer executable instructions, the computer executable instructions being executed by the processor to implement steps of the method as described above.

[0009] According to another aspect of the present application, a computer storage medium is provided, having computer executable instructions stored thereon, the computer executable instructions being executed by a processor to implement steps of the method as described above.

[0010] Compared with the prior art, the cloud computing environment load balancing method and system according to the embodiment of the application can dynamically calculate the node matching degree score through user portrait information, model deployment information and load value, realize accurate scheduling of service requests, automatically select compatible nodes with similar semantics and low load when the target model is unavailable, solve the problems of model version mismatch, resource waste and service interruption caused by traditional load balancing strategies, and has the advantages of improving the accuracy of service request scheduling, optimizing resource utilization, and enhancing service continuity in the scenario of unavailable model. BRIEF DESCRIPTION OF DRAWINGS

[0011] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of the application when taken in conjunction with the accompanying drawings. The drawings provided in the present application are used to provide further understanding of the embodiments of the present application, and constitute a part of the specification, and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0012] Figure 1 A flowchart of the cloud computing environment load balancing method of the present application.

[0013] Figure 2 A block diagram of the cloud computing environment load balancing method and system of the present application.

[0014] Figure 3 A block diagram of the electronic device of the present application. DETAILED DESCRIPTION

[0015] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein.

[0016] SUMMARY

[0017] In the traditional cloud computing load balancing system, the resource scheduling strategy mainly relies on the hardware resource indicators of the server nodes and the number of requests to allocate tasks, and fails to take the user portrait features and model semantic matching degree into account in the decision dimension. When the user request involves a specific version of the business model or a customized algorithm, the existing system cannot effectively identify the relevance between the user demand and the node model deployment, resulting in that the request may be allocated to a computing node with incompatible model version or semantic mismatch. This scheduling deviation will directly cause the model inference error rate to rise, the service response time to increase, and at the same time cause the load distribution imbalance among nodes, and some nodes will produce additional consumption of computing resources due to processing of non-adaptive requests.

[0018] If the above problems are not solved, the system will face the dual loss of model inference path redundancy and resource mismatch for a long time. The frequent cross-node model calls in the request processing process not only increase the network transmission overhead, but also produce context reconstruction cost due to model version switching. The nonlinear relationship between node load indicators and actual service capacity will lead to the failure of automatic scaling mechanism, and some nodes will be in a high-load and low-efficiency running state, while the nodes with model adaptation capability cannot trigger resource expansion due to the load indicators not reaching the threshold.

[0019] When facing the above problems, the present application first analyzes the root cause of the model matching failure caused by the traditional load balancing strategy relying only on hardware indicators, finds that there is a lack of dynamic association mechanism between user portrait and node model deployment characteristics, and to solve this problem, explores the possibility of multi-dimensional matching of user portrait vectors and node model semantic vectors, while considering the influence of node load state on scheduling decisions, further studies how to build a comprehensive scoring model containing user behavior preferences, model adaptation degree and load state, and designs a scoring threshold triggering mechanism to automatically enable the standby node to deploy the target model when the optimal node does not meet the conditions, and realizes the precise alignment of model version and user demand.

[0020] Exemplary method

[0021] Figure 1 A load balancing method in a cloud computing environment according to an embodiment of the present application is illustrated, including: receiving a service request of a target user; obtaining user portrait information of the target user; determining a target model corresponding to the service request according to the user portrait information; obtaining model deployment information and current load values of each node in the current service pool; calculating the matching degree score of the service request and each node according to the user portrait information, the model deployment information and the corresponding current load values, and taking the node with the highest matching degree score as the target node; judging whether the matching degree score of the target node is higher than a preset score threshold, if higher, scheduling the service request to the target node, otherwise, deploying the target model on the standby node of the current service pool, and scheduling the service request to the compatible node, the compatible node being: the node with the highest semantic similarity with the target model and the corresponding current load value being lower than a preset load threshold among all nodes.

[0022] The user portrait information refers to a set of characteristic attributes of a target user described by multidimensional data, which can be obtained by user behavior log analysis, business configuration data analysis, and compliance policy database query, and is used to represent the individual characteristics of the user in terms of service preference, business priority, and compliance requirements, and to realize accurate matching of user needs and model capabilities in load balancing decision-making. The model deployment information refers to the types and version identifiers of the algorithm models deployed on the service node, as well as the records of the algorithm models in processing service requests, which can be queried in real time through a node registration center or a model metadata database. The current load value refers to a quantitative indicator of the real-time resource occupation state of the service node, which can be collected by a monitoring system using GPU memory occupancy rate, concurrent request processing number, and thread pool utilization rate indicators to provide a basis for judging the availability of the node for load balancing. The preset scoring threshold refers to a critical decision value for triggering direct scheduling or model deployment decision, which can be dynamically adjusted by historical scheduling data analysis combined with service quality indicators to balance the conflicting demands of scheduling efficiency and resource overhead. The standby node refers to a scalable resource unit reserved in the service pool, which can be implemented by a virtual machine hot backup instance or a containerized elastic scaling cluster.

[0023] The core innovation of the present application lies in constructing a load balancing mechanism combining user portrait driving and model semantic matching, which dynamically generates a matching score by calculating the multidimensional similarity between the user feature vector and the model adaptation vector, and combining the real-time load state of the node, to realize intelligent scheduling decision considering user individual needs, model version compatibility, and resource utilization efficiency, and effectively solve the problems of scheduling errors and resource waste caused by ignoring user-model matching in traditional methods.

[0024] For example, the scheme of the present application is implemented as follows:

[0025] First, a service request from user A is received, and the portrait information of user A is extracted from the user database, including the historical transaction behavior, credit rating, risk preference, and other characteristics of user A. According to the portrait information of user A, it is determined that the anti-fraud model V2.0 version dedicated for high-risk users needs to be used as the target model.

[0026] Then, scan all nodes in the current service pool to obtain the model version information and the current load value deployed on each node, calculate the matching degree score of the service request and each node according to the portrait information of user A, the model deployment information and the load value of each node, and in the calculation process, each factor is assigned different weights. Assuming that the preset score threshold is 0.8, the matching degree score of node 1 is the highest, which is 0.85, then 0.85 is compared with the preset score threshold 0.8, and it is determined that it is higher than the threshold, therefore, the service request of user A is directly scheduled to node 1 for processing. If the score of node 1 is 0.75, which is lower than the threshold 0.8, the system will deploy the anti-fraud model V2.0 version on the standby node X, at the same time, scan all nodes, find that node 2 does not deploy V2.0 version, but deploys V1.9 version with high semantic similarity, and the current load value is lower than the preset threshold, then the request is temporarily scheduled to node 2 for processing, so as to ensure the timeliness of service response.

[0027] Through the above scheme, the present application realizes accurate load balancing based on user portrait and model semantics. By introducing user portrait analysis and model semantic matching mechanism, the system can accurately schedule user requests to the most suitable processing node, avoiding the problem of requests being allocated to nodes with incompatible model versions or mismatched semantics. This method significantly improves the accuracy and efficiency of request processing, reduces processing errors and delays caused by model mismatch.

[0028] In some of the above schemes of the present application, determining the target model corresponding to the service request according to the user portrait information specifically includes: performing multi-label vector encoding processing on the user portrait information to obtain a user portrait vector containing user behavior preference, business priority, and compliance requirement dimension; generating a multi-dimensional label classification result of the target user according to the user portrait vector through a pre-trained user classification model; and determining the target model from a preset target model mapping table according to the multi-dimensional label classification result and the business type of the service request.

[0029] Among them, the multi-label vector encoding processing adopts a neural network architecture based on an attention mechanism to jointly embed user historical behavior data, business configuration parameters and compliance policy documents to generate a 128-dimensional dense vector. The user classification model adopts a hierarchical classification mechanism, the first layer identifies the industry field to which the user belongs, the second layer divides the business urgency level, and the third layer detects the data compliance constraint type. The target model mapping table is stored in the form of key-value pair, the key is the hash code of the multi-dimensional label combination, and the value is the corresponding model identifier and version number, and the hash code is generated by label priority weighting.

[0030] Specifically, after the user portrait information is processed by multi-label vector encoding, a structured vector representation containing user behavior preference, business priority, and compliance requirement dimensions is formed. The vector is input into a pre-trained three-layer classification model. The first layer extracts local features through convolution operation and outputs industry classification probability distribution. The second layer uses a gated recurrent unit to capture temporal dependency and generates business priority score. The third layer identifies compliance constraint conditions through a rule matching engine. The multi-dimensional label classification result and the business type of the service request jointly constitute a composite query condition, triggering the fast retrieval mechanism of the target model mapping table. For example, when the user portrait vector is classified and outputs the "financial industry-high priority-EU GDPR compliance" label combination, the system automatically matches a model instance that supports financial anti-fraud algorithms and meets the GDPR encryption standard.

[0031] Through the above technical solutions, the present application realizes accurate analysis and classification of user portrait information, and can more accurately select a target model matching user demand. This method improves the accuracy of model selection, reduces the probability of selecting an unsuitable model, and thus optimizes overall service quality and resource utilization efficiency. At the same time, through the preset mapping table, the present application can quickly respond to different types of service requests, improving the processing speed and flexibility of the system.

[0032] In some of the above schemes of the present application, the calculation of the matching degree score comprises: obtaining a portrait adaptation vector of the deployed model according to the model deployment information ; calculating the cosine similarity between the portrait adaptation vector and the user portrait vector of the target user to obtain a first score value :

[0033] ;

[0034] normalizing the current load value of the node according to a preset maximum load value to obtain a second score value :

[0035] ;

[0036] extracting a request semantic vector of each service request and a model semantic vector of the node , calculating the semantic similarity between the request semantic vector and the model semantic vector to obtain a third score value : ​​​​The first, second, and third scores are weighted and summed to obtain the service node and node... Match score : ,in, , and The preset weighting coefficients satisfy... .

[0037] Specifically, when a service request arrives, the user profile vector... Adaptation vectors for each node's image Vector space similarity calculations are performed. For example, vectors with higher weights in the user preference dimension will achieve a higher matching degree in cosine similarity calculations, along with node load values. through After normalization, the load score is converted to a 0-1 range. Each 10% increase in load value will result in... Decreased by 0.1. Request semantic vector. Generated through a natural language processing model, and related to the model's semantic vector. Similarity calculation can employ a deep matching model based on an attention mechanism. During the weighted summation process, when the three scores are 0.85, 0.7, and 0.9 respectively, and the weight coefficients are 0.5, 0.25, and 0.25, the final matching score is 0.85×0.5 + 0.7×0.25 + 0.9×0.25 = 0.825. This comprehensive scoring mechanism ensures optimal node selection across the three dimensions of user feature matching, load balancing, and semantic association, effectively improving service scheduling accuracy and resource utilization.

[0038] Through the above technical solutions, this application achieves a comprehensive consideration of user characteristics, node load, and request semantics, improving the matching accuracy between service requests and processing nodes. This allows for more precise allocation of user requests to the most suitable processing nodes, effectively enhancing load balancing in cloud computing environments. Simultaneously, by introducing semantic similarity scoring, a high degree of matching between request content and model functionality is ensured, reducing unnecessary model calls and resource waste. Furthermore, the scoring mechanism considering node load helps prevent overload of individual nodes, achieving more balanced resource utilization.

[0039] In some of the solutions described above in this application, the image adaptation vector The acquisition includes: extracting nodes from model deployment information. The set of users historically processed by the deployed model Based on user set For each user's user profile vector, calculate the mean vector of the model's profile features as the profile adaptation vector. :

[0040] ;

[0041] wherein, represents a user portrait vector of a user in the user set .

[0042] Specifically, when the node processes a service request, a mapping relationship between a user identifier and a model version is automatically recorded to form a historical user set. In the calculation of the portrait adaptation vector, portrait vectors of all users in the set are obtained in parallel through a distributed database, and a MapReduce framework is used for dimension calculation of the vector. Finally, a dimension-aligned portrait feature mean vector is generated by aggregation, which can reflect the common characteristics of the user groups served by the model. When performing similarity matching with a new user portrait vector, the degree of fit between the model service capability and the user demand can be effectively identified. For example, a compliance audit model deployed on a node has historically served financial industry users, and the portrait adaptation vector of the node presents high weight features in the compliance demand dimension. When a new user request with strong compliance demand is received, the first score value of the node will be significantly improved, thereby improving the discrimination of the matching degree score.

[0043] For example, assume that a node has historically processed 3 users, and their user portrait vectors are , , respectively. The portrait adaptation vector Mi(u) of the node is calculated as follows: .

[0044] Through the above technical solutions, the portrait adaptation vector of the model can be dynamically generated based on historical user data, reflecting the adaptation degree of the model to different types of users. This improves the matching accuracy between user requests and models, reduces the probability of mismatched scheduling. Further, the scheme smoothes the fluctuation of individual user data by calculating the mean vector, improving the stability and representativeness of the portrait adaptation vector. At the same time, the scheme uses existing user portrait data, without additional data collection and storage overhead, and has high implementation efficiency.

[0045] In some of the above schemes of the present application, when the model deployed on the node has not processed any user request, the user set is empty, resulting in the inability to calculate the portrait adaptation vector, causing data missing in the matching degree score calculation process, and affecting the accuracy of the load balancing decision.

[0046] The present application further proposes that before calculating the portrait feature mean vector of the model, it further includes: judging whether the user set ​whether empty, if yes, obtaining the service type of the model, and extracting a general portrait vector corresponding to the model from a preset service type-general portrait vector mapping table as the portrait adaptation vector of the model according to the service type .

[0047] In the service type-general portrait vector mapping table, standardized user feature vectors corresponding to different service types are stored in advance, for example, a general vector corresponding to high compliance requirements for a financial risk control type service, and a general vector corresponding to high computing resource requirements for an image processing type service. The general portrait vector is generated by analyzing historical user data of the same type of service, and can reflect the typical user features of the service type.

[0048] Specifically, when a newly deployed model first joins the service pool, since there is a lack of historical processing records, a preset general portrait vector is matched according to the service type. For example, when a compliance audit model is deployed, if it has not processed user requests, a general vector containing high compliance weight and medium response speed requirement is extracted from the mapping table as a temporary adaptation basis. The vector and the portrait vector of the target user are calculated for similarity, until the model accumulates enough user data and switches to the mean vector based on actual processing. This mechanism guarantees the effectiveness of scoring in the cold start stage, and avoids the problem that requests are incorrectly scheduled to incompatible nodes due to data loss.

[0049] Through the above technical solutions, the application solves the problem that a newly deployed model lacks historical user data and cannot generate a portrait adaptation vector. By introducing a preset service type-general portrait vector mapping table, an initial portrait adaptation vector is provided for the new model, so that the system can immediately start load balancing scheduling based on user portraits without waiting for enough user data to be accumulated. This method improves the adaptability of the system to newly deployed models, ensures the continuity and stability of the load balancing mechanism, and provides support for the rapid access and use of new models.

[0050] In some of the above schemes of the application, the matching degree score is calculated based on the user portrait, model deployment information and load value, and scheduling is performed, but the dynamic performance change of the node in processing requests is not considered, which may cause the score to not reflect the actual service capability of the node in time, resulting in deviation in subsequent request scheduling.

[0051] The application further proposes: obtaining the average service delay and the average error rate of the target node in processing a plurality of service requests in a preset monitoring period; calculating a score adjustment factor of the target stage according to the average service delay and the average error rate; and adjusting the matching degree score calculated by the target node in the next round of service requests according to the score adjustment factor.

[0052] The preset monitoring period is set as a minute-level or hour-level time window, and is used to collect real-time performance data of the node processing request. The average service delay is obtained by statistical arithmetic mean value of all request processing time in the period, and the average error rate is calculated by the ratio of the number of error responses to the total request amount.

[0053] Through the above technical solutions, the application can dynamically adjust the matching degree score of the node, so that the load balancing system can more accurately reflect the actual service capability and quality of each node when scheduling service requests. Specifically, for the node with high service delay or high error rate, the score adjustment factor will be correspondingly reduced, thereby reducing the probability of being selected in subsequent scheduling, which helps to allocate requests to more stable and efficient nodes. At the same time, this dynamic adjustment mechanism can also timely feedback the changes of node performance, so that the load balancing system can quickly adapt to various changes in the cloud environment, and improve the overall service quality and resource utilization efficiency.

[0054] In some schemes of the application, the calculation method of the score adjustment factor is specifically:

[0055] ;

[0056] Among them, is a preset maximum service delay threshold, is a preset maximum error rate threshold, and is a weight coefficient, satisfying .

[0057] Among them, the maximum service delay threshold is used to limit the upper limit of the delay of the node processing request, and the node exceeding the threshold will be determined as insufficient performance; the maximum error rate threshold is used to constrain the fault tolerance range of the node processing request, and the node exceeding the value indicates that the service quality is decreased. The weight coefficient and is set according to the business demand, for example, in the scene with high real-time requirement, may be 0.7, 0.3 is set to focus on the delay index.

[0058] Exemplary system

[0059] Figure 2The illustration shows a load balancing system in a cloud computing environment according to an embodiment of this application, including: a request receiving module for receiving service requests from target users; a profile acquisition module for acquiring user profile information of the target users; a model determination module for determining a target model corresponding to the service request based on the user profile information; an information acquisition module for acquiring model deployment information and current load values ​​of each node in the current service pool; a target node determination module for calculating the matching score between the service request and each node based on the user profile information, model deployment information, and corresponding current load values, and selecting the node with the highest matching score as the target node; and a decision module for determining whether the matching score of the target node is higher than a preset score threshold. If it is higher, the service request is scheduled to the target node; otherwise, the target model is deployed on a backup node in the current service pool, and the service request is scheduled to a compatible node, wherein the compatible node is the node with the highest semantic similarity to the target model and whose corresponding current load value is lower than a preset load threshold.

[0060] In one example, the model determination module determines the target model corresponding to the service request based on the user profile information by: performing multi-label vector encoding on the user profile information to obtain a user profile vector containing dimensions such as user behavior preferences, business priorities, and compliance requirements; generating a multi-dimensional label classification result of the target user based on the user profile vector using a pre-trained user classification model; and determining the target model from a preset target model mapping table based on the multi-dimensional label classification result and the business type of the service request.

[0061] In one example, the target node determination module calculates the matching score by: obtaining nodes based on model deployment information. Image adaptation vector of the deployed model Adaptation vectors for image User profile vectors of target users Calculate the cosine similarity to obtain the first score. :

[0062] ;

[0063] Based on the preset maximum load value For nodes Current load value Normalization is performed to obtain the second score. :

[0064] ;

[0065] Extract the request semantic vector of each service request and nodes Model semantic vector Calculate the request semantic vector the semantic similarity of the model semantic vector and the semantic of the model semantic vector : ; and sum the first score value, the second score value and the third score value by weighting to obtain a matching degree score of the service node and the node : , , and are preset weight coefficients, and satisfy .

[0066] In an example, the obtaining of the portrait adaptation vector by the target node determination module includes: extracting a user set processed by the model of the node from the model deployment information; and calculating a portrait feature mean vector of the model as the portrait adaptation vector :

[0067] ;

[0068] wherein, represents a user portrait vector of a user in the user set .

[0069] In an example, before the target node determination module calculates the portrait feature mean vector of the model, the target node determination module further includes: judging whether the user set is empty, and if yes, obtaining a service type of the model, and extracting a general portrait vector corresponding to the model from a preset service type-general portrait vector mapping table according to the service type as the portrait adaptation vector of the model.

[0070] In an example, the system further includes a feedback module, which is configured to obtain an average service time delay and an average error rate of the target node in processing a plurality of service requests in a preset monitoring period; calculate a score adjustment factor of the target stage according to the average service time delay and the average error rate; and adjust the matching degree score calculated by the target node in the next round of service requests according to the score adjustment factor.

[0071] In an example, the feedback module calculates the score adjustment factor in the following manner:

[0072] ;

[0073] wherein, is a preset maximum service time delay threshold, is a preset maximum error rate threshold,​​ and are weight coefficients, satisfying .

[0074] Exemplary electronic device

[0075] Figure 3 An electronic device according to embodiments of the present application is illustrated. The electronic device can be the mobile device itself, or a stand-alone device independent from the mobile device, which can communicate with the mobile device to receive the acquired input signals therefrom and send the selected target driving behavior thereto.

[0076] Figure 3 A block diagram of an electronic device according to embodiments of the present application is illustrated.

[0077] As shown in Figure 3 , the electronic device includes one or more processors and a memory.

[0078] The processor can be a central processing unit (CPU) or other form of processing unit having data processing and / or instruction execution capabilities, and can control other components in the electronic device to perform desired functions.

[0079] The memory can include one or more computer program products, which can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory, for example, can include random access memory (RAM), cache, and / or the like. The non-volatile memory, for example, can include read only memory (ROM), hard disk, flash memory, and / or the like. One or more computer program instructions can be stored on the computer-readable storage media, and the processor can execute the program instructions to implement the driving behavior decision method of the embodiments of the present application described above and / or other desired functions.

[0080] In one example, the electronic device can further include input and output devices, which are interconnected through a bus system and / or other form of connection mechanism (not shown).

[0081] Of course, for simplicity, Figure 3 only some of the components in the electronic device related to the present application are shown in the figure, and components such as buses, input / output interfaces, and the like are omitted. In addition, the electronic device can include any other appropriate components according to specific application cases.

[0082] Exemplary computer-readable medium

[0083] Embodiments of the present application can also be computer readable storage media storing computer program instructions thereon which, when executed by a processor, cause the processor to perform the steps of the driving behavior decision method according to various embodiments of the present application described in the above "Exemplary Methods" section of the specification.

[0084] The computer readable storage medium can be any combination of one or more computer readable medium(s). The computer readable medium can be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include an electrical connection having one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0085] The above describes the basic principles of the present application in conjunction with specific embodiments, but it should be noted that the advantages, benefits, effects and the like mentioned in the present application are only examples and are not limiting, and these advantages, benefits, effects and the like cannot be considered as necessary for each embodiment of the present application. In addition, the above specific details are only for the purpose of example and understanding, and are not limiting, and the above details do not limit the present application to the above specific details.

[0086] The block diagrams of the devices, apparatuses, equipment, systems involved in the present application are only illustrative examples and are not intended to require or imply that the connections, arrangements, configurations shown in the block diagrams must be connected, arranged, configured. As those skilled in the art will recognize, these devices, apparatuses, equipment, systems can be connected, arranged, configured in any manner. Words such as "include", "contain", "have" and the like are open-ended words, mean "including but not limited to", and can be used interchangeably. The words "or" and "and" used herein mean the word "and / or", and can be used interchangeably unless the context clearly indicates otherwise. The word "such as" used herein means the phrase "such as but not limited to", and can be used interchangeably.

[0087] It should also be noted that in the devices, apparatuses and methods of the present application, each component or step can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions of the present application.

[0088] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other aspects without departing from the scope of the application. Thus, the present application is not intended to be limited to the aspects shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0089] The above description has been presented to enable any person skilled in the art to make or use the application. Numerous modifications and alterations to this application will be readily apparent to those skilled in the art, and it is intended to include all such modifications and alterations insofar as they come within the scope of the general concepts defined herein.

Claims

1. A load balancing method in a cloud computing environment, characterized in that, include: Receive service requests from target users; Obtain the user profile information of the target user; Determine the target model corresponding to the service request based on the user profile information; Obtain the model deployment information and current load value of each node in the current service pool; The matching score between the service request and each node is calculated based on the user profile information, model deployment information and the corresponding current load value, and the node with the highest matching score is taken as the target node. Determine whether the matching score of the target node is higher than a preset score threshold. If it is higher, schedule the service request to the target node. Otherwise, deploy the target model on the standby node of the current service pool and schedule the service request to a compatible node. The compatible node is the node with the highest semantic similarity to the target model and whose current load value is lower than a preset load threshold.

2. The load balancing method in a cloud computing environment according to claim 1, characterized in that, Determining the target model corresponding to the service request based on the user profile information includes: The user profile information is processed by multi-label vector encoding to obtain a user profile vector containing dimensions of user behavior preferences, business priorities, and compliance requirements. A multi-dimensional label classification result for the target user is generated based on the user profile vector using a pre-trained user classification model. Based on the multidimensional label classification results and the business type of the service request, the target model is determined from the preset target model mapping table.

3. The load balancing method in a cloud computing environment according to claim 2, characterized in that, The matching score is calculated as follows: Obtain nodes based on the model deployment information. Image adaptation vector of the deployed model ; Adaptation vector for the image User profile vectors of target users Calculate the cosine similarity to obtain the first score. : ； Based on the preset maximum load value For nodes Current load value Normalization is performed to obtain the second score. : ； Extract the request semantic vector of each service request and nodes Model semantic vector Calculate the request semantic vector With the model semantic vector The semantic similarity is used to obtain the third score. : ; The first score, the second score, and the third score are weighted and summed to obtain the service node and node. Match score : ,in, , and The preset weighting coefficients satisfy... .

4. A load balancing method in a cloud computing environment according to claim 3, characterized in that, The image adaptation vector The acquisition includes: Extract nodes from the model deployment information The set of users historically processed by the deployed model ; According to the user set For each user's user profile vector, the mean vector of the model's profile features is calculated as the profile adaptation vector. : ； in, Represents a set of users Chinese users User profile vectors.

5. A load balancing method in a cloud computing environment according to claim 4, characterized in that, The image feature mean vector of the computational model is preceded by: Determine the user set If the value is empty, then obtain the service type of the model, and extract the general image vector corresponding to the model from the preset service type-general image vector mapping table as the image adaptation vector of the model. .

6. A load balancing method in a cloud computing environment according to claim 1, characterized in that, Also includes: Obtain the average service latency and average error rate of the target node when processing multiple service requests within a preset monitoring period; The scoring adjustment factor for the target phase is calculated based on the average service latency and average error rate. The matching score of the target node in the next round of service requests is adjusted according to the scoring adjustment factor.

7. A load balancing method in a cloud computing environment according to claim 6, characterized in that, The scoring adjustment factor is calculated as follows: ； in, The preset maximum service latency threshold. The preset maximum error rate threshold, and For the weighting coefficients, satisfying .

8. A load balancing system in a cloud computing environment, characterized in that, include: The request receiving module receives service requests from target users. The profile acquisition module acquires the user profile information of the target user; The model determination module determines the target model corresponding to the service request based on the user profile information. The information acquisition module obtains the model deployment information and current load value of each node in the current service pool; The target node determination module calculates the matching score between the service request and each node based on the user profile information, model deployment information and the corresponding current load value, and takes the node with the highest matching score as the target node. The decision module determines whether the matching score of the target node is higher than a preset score threshold. If it is higher, the service request is scheduled to the target node. Otherwise, the target model is deployed on the standby node of the current service pool, and the service request is scheduled to a compatible node. The compatible node is the node with the highest semantic similarity to the target model and whose current load value is lower than a preset load threshold.

9. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 7.

10. A computer storage medium storing computer-executable instructions thereon, characterized in that: When the computer-executable instructions are executed by a processor, they implement the steps of the method as described in any one of claims 1 to 7.

Images