Hybrid model-based cloud desktop resource dynamic allocation method, device and equipment

Abstract

The application provides a hybrid model-based cloud desktop resource dynamic allocation method, device and equipment, and relates to the technical field of cloud desktop management. The method comprises the following steps: obtaining behavior data of a target user for a cloud desktop; inputting the behavior data into a double-channel intelligent prediction model, associating and evaluating the portrait of the target user and resource demand according to the behavior data, obtaining a first resource demand prediction result, and performing time sequence evaluation on the resource demand of the target user according to the first resource demand prediction result to obtain a resource demand prediction result of the target user, wherein the model parameters of the double-channel intelligent prediction model are determined based on the association mode between the pre-learned historical behavior data and resource demand; and adjusting the resource allocation of the cloud desktop according to the resource demand prediction result. Through the application, the resource utilization rate of the cloud desktop system is improved.

Description

Technical Field

[0001] This application relates to the field of cloud desktop management technology, and in particular to a method, apparatus and device for dynamic allocation of cloud desktop resources based on a hybrid model. Background Technology

[0002] Cloud desktops, also known as Virtual Desktop Infrastructure (VDI), are a key application in cloud computing and fall under the category of desktop virtualization and resource scheduling technologies. This technology centrally deploys desktop environments in data centers and delivers virtual desktops to users via remote protocols, achieving unified management and intensive use of computing resources. It has become a core supporting technology for enterprise digital transformation and remote work. Against this backdrop, one of the core challenges facing VDI technology is how to efficiently and flexibly allocate cloud desktop resources to meet diverse user computing needs and optimize resource utilization.

[0003] In current cloud desktop systems, resource scheduling optimization is key to improving overall efficiency. For example, existing technologies, such as patent CN103605574A, obtain historical data on host CPU utilization and then migrate virtual machines on hosts with high CPU utilization to achieve dynamic balancing of cloud desktop resources. However, this scheduling strategy, based on a single CPU utilization rate and triggered only when the load is high, often results in low overall resource utilization of the cloud desktop system. Summary of the Invention

[0004] This application provides a method, apparatus, and device for dynamic allocation of cloud desktop resources based on a hybrid model, in order to improve the resource utilization of cloud desktop systems.

[0005] Firstly, this application provides a method for dynamic allocation of cloud desktop resources based on a hybrid model, including:

[0006] Acquire target users' behavioral data regarding cloud desktops;

[0007] Behavioral data is input into the dual-channel intelligent prediction model. Based on the behavioral data, the profile of the target user and the resource demand are correlated and evaluated to obtain the first resource demand prediction result. Based on the first resource demand prediction result, the resource demand of the target user is evaluated in time series to obtain the resource demand prediction result of the target user. The model parameters of the dual-channel intelligent prediction model are determined based on the correlation pattern between historical behavioral data and resource demand learned in advance.

[0008] Adjust the resource allocation for cloud desktops based on resource demand forecasts.

[0009] In one possible implementation, the dual-channel intelligent prediction model includes a time-series prediction model and a profile inference model. Behavioral data is input into the dual-channel intelligent prediction model. Based on the behavioral data, the profile of the target user is correlated with their resource needs to obtain a first resource demand prediction result. Then, based on the first resource demand prediction result, a time-series evaluation of the target user's resource needs is performed to obtain the target user's resource demand prediction result, including:

[0010] Behavioral data is input into the profile reasoning model, global features of the behavioral data are extracted, and based on the global features, the profile of the target user and resource needs are correlated and evaluated to obtain the first resource needs prediction result.

[0011] The behavioral data and the first resource demand prediction results are input into the time series prediction model. Based on the behavioral data and the first resource demand prediction results, the resource demand of the target user is evaluated in a time series, and the resource demand prediction results are obtained.

[0012] In one possible implementation, behavioral data is input into a profile reasoning model, global features of the behavioral data are extracted, and based on these global features, the profile of the target user is correlated with resource requirements to obtain a first resource requirement prediction result, including:

[0013] Extracting multimodal features from behavioral data;

[0014] Multimodal features are fused to obtain the fused feature representation;

[0015] The feature representation is input into the profile reasoning model to extract global features from the feature representation. Based on the global features, the profile of the target user and the resource requirements are correlated and evaluated to obtain the first resource requirement prediction result.

[0016] In one possible implementation, multimodal features are fused to obtain a fused feature representation, including:

[0017] Based on the gating function, feature fusion is performed on multimodal features to obtain the fused feature representation;

[0018] Alternatively, based on an adaptive cross-attention mechanism, multimodal features can be interactively fused by constructing an intermodal semantic association matrix and a dynamic attention weight allocation strategy to generate a fused feature representation.

[0019] In one possible implementation, based on an adaptive cross-attention mechanism, multimodal features are interactively fused using a dynamic weight allocation strategy and cross-modal semantic alignment to generate a fused feature representation, including:

[0020] Intramodal feature encoding is performed on multimodal features to generate an initial modal feature representation containing local semantic information;

[0021] Based on the contextual information of the initial modal feature representation, calculate the semantic similarity between modalities;

[0022] Based on the semantic similarity between modalities, an intermodal semantic association matrix is ​​constructed. The intermodal semantic association matrix is ​​used to represent the semantic correspondence between features of each modality.

[0023] Based on the intermodal semantic association matrix and contextual information, an adaptive dynamic attention weight allocation strategy is determined.

[0024] Based on a dynamic attention weight allocation strategy, the initial modal feature representations are weighted and summed to generate fused feature representations.

[0025] In one possible implementation, behavioral data and the first resource demand prediction result are input into a time-series prediction model. Based on the behavioral data and the first resource demand prediction result, a time-series evaluation of the target user's resource demand is performed to obtain a resource demand prediction result, including:

[0026] The feature representation and the first resource demand prediction result are input into the time series prediction model. Based on the feature representation and the first resource demand prediction result, the resource demand of the target user is evaluated in time series to obtain the second resource demand prediction result.

[0027] The first resource demand forecast result and the second resource demand forecast result are weighted and summed, and the weighted sum is determined as the resource demand forecast result. The weights of the weighted sum are adaptively adjusted according to the prediction errors of the time series forecast model and the profile inference model.

[0028] In one possible implementation, the temporal prediction model is a long short-term memory network model that incorporates a gated attention mechanism; the profiling reasoning model is a decision model based on gradient boosting.

[0029] In one possible implementation, adjusting the resource allocation of cloud desktops based on resource demand forecasts includes:

[0030] Based on the resource demand forecast results, determine the resource priority of the cloud desktop;

[0031] Based on resource priority, the resource allocation of cloud desktops is adjusted using the virtualization resource hot-add mechanism.

[0032] In one possible implementation, the resource allocation of the cloud desktop is adjusted according to resource priority using a virtualization resource hot-adding mechanism, including:

[0033] Obtain resource usage metrics for cloud desktops;

[0034] When the resource usage index exceeds the usage threshold continuously within a set period, a resource demand event is generated.

[0035] In response to resource demand events, a virtual resource pool is built from the physical server cluster;

[0036] Based on resource priority, formulate a virtualization resource pool allocation strategy;

[0037] Based on the allocation strategy, the hot-add interface of the virtualization platform is invoked to add the required virtualization resource pool to the cloud desktop.

[0038] In one possible implementation, adjusting the resource allocation of cloud desktops based on resource demand forecasts includes:

[0039] The resource allocation problem of cloud desktops is modeled as a Markov decision model. The state space of the Markov decision model includes the current resource usage of cloud desktops, the predicted resource demand, and the historical resource allocation records. The action space of the Markov decision model includes the resource adjustment allocation amount of cloud desktops. The reward function of the Markov decision model includes a weighted combination based on resource utilization, user response time, and resource allocation cost.

[0040] Construct a reinforcement learning agent based on the Markov decision model;

[0041] By using a reinforcement learning agent, iterative training is conducted with the cloud desktop to learn the cloud desktop's resource allocation strategy. The training process aims to maximize the reward function.

[0042] Adjust the resource allocation of the cloud desktop according to the resource allocation strategy.

[0043] Secondly, this application provides a cloud desktop resource dynamic allocation device based on a hybrid model, comprising:

[0044] The acquisition module is used to acquire behavioral data of the target user regarding the cloud desktop;

[0045] The determination module is used to input behavioral data into the dual-channel intelligent prediction model, evaluate the correlation between the target user's profile and resource demand based on the behavioral data, obtain the first resource demand prediction result, and evaluate the target user's resource demand in a time series based on the first resource demand prediction result to obtain the target user's resource demand prediction result. The model parameters of the dual-channel intelligent prediction model are determined based on the correlation pattern between historical behavioral data and resource demand learned in advance.

[0046] The adjustment module is used to adjust the resource allocation of cloud desktops based on the resource demand prediction results.

[0047] In one possible implementation, the dual-channel intelligent prediction model includes a time-series prediction model and a profiling inference model, and the determination module is specifically used for:

[0048] Behavioral data is input into the profile reasoning model, global features of the behavioral data are extracted, and based on the global features, the profile of the target user and resource needs are correlated and evaluated to obtain the first resource needs prediction result.

[0049] The behavioral data and the first resource demand prediction results are input into the time series prediction model. Based on the behavioral data and the first resource demand prediction results, the resource demand of the target user is evaluated in a time series, and the resource demand prediction results are obtained.

[0050] In one possible implementation, the determining module is specifically used for:

[0051] Extracting multimodal features from behavioral data;

[0052] Multimodal features are fused to obtain the fused feature representation;

[0053] The feature representation is input into the profile reasoning model to extract global features from the feature representation. Based on the global features, the profile of the target user and the resource requirements are correlated and evaluated to obtain the first resource requirement prediction result.

[0054] In one possible implementation, the cloud desktop resource dynamic allocation device based on the hybrid model further includes a fusion module, which is specifically used for:

[0055] Based on the gating function, feature fusion is performed on multimodal features to obtain the fused feature representation;

[0056] Alternatively, based on an adaptive cross-attention mechanism, multimodal features can be interactively fused by constructing an intermodal semantic association matrix and a dynamic attention weight allocation strategy to generate a fused feature representation.

[0057] In one possible implementation, the fusion module is specifically used for:

[0058] Intramodal feature encoding is performed on multimodal features to generate an initial modal feature representation containing local semantic information;

[0059] Based on the contextual information of the initial modal feature representation, calculate the semantic similarity between modalities;

[0060] Based on the semantic similarity between modalities, an intermodal semantic association matrix is ​​constructed. The intermodal semantic association matrix is ​​used to represent the semantic correspondence between features of each modality.

[0061] Based on the intermodal semantic association matrix and contextual information, an adaptive dynamic attention weight allocation strategy is determined.

[0062] Based on a dynamic attention weight allocation strategy, the initial modal feature representations are weighted and summed to generate fused feature representations.

[0063] In one possible implementation, the determining module is specifically used for:

[0064] The feature representation and the first resource demand prediction result are input into the time series prediction model. Based on the feature representation and the first resource demand prediction result, the resource demand of the target user is evaluated in time series to obtain the second resource demand prediction result.

[0065] The first resource demand forecast result and the second resource demand forecast result are weighted and summed, and the weighted sum is determined as the resource demand forecast result. The weights of the weighted sum are adaptively adjusted according to the prediction errors of the time series forecast model and the profile inference model.

[0066] In one possible implementation, the temporal prediction model is a long short-term memory network model that incorporates a gated attention mechanism; the profiling reasoning model is a decision model based on gradient boosting.

[0067] In one possible implementation, the adjustment module is specifically used for:

[0068] Based on the resource demand forecast results, determine the resource priority of the cloud desktop;

[0069] Based on resource priority, the resource allocation of cloud desktops is adjusted using the virtualization resource hot-add mechanism.

[0070] In one possible implementation, the adjustment module is specifically used for:

[0071] Obtain resource usage metrics for cloud desktops;

[0072] When the resource usage index exceeds the usage threshold continuously within a set period, a resource demand event is generated.

[0073] In response to resource demand events, a virtual resource pool is built from the physical server cluster;

[0074] Based on resource priority, formulate a virtualization resource pool allocation strategy;

[0075] Based on the allocation strategy, the hot-add interface of the virtualization platform is invoked to add the required virtualization resource pool to the cloud desktop.

[0076] In one possible implementation, the adjustment module is specifically used for:

[0077] The resource allocation problem of cloud desktops is modeled as a Markov decision model. The state space of the Markov decision model includes the current resource usage of cloud desktops, the predicted resource demand, and the historical resource allocation records. The action space of the Markov decision model includes the resource adjustment allocation amount of cloud desktops. The reward function of the Markov decision model includes a weighted combination based on resource utilization, user response time, and resource allocation cost.

[0078] Construct a reinforcement learning agent based on the Markov decision model;

[0079] By using a reinforcement learning agent, iterative training is conducted with the cloud desktop to learn the cloud desktop's resource allocation strategy. The training process aims to maximize the reward function.

[0080] Adjust the resource allocation of the cloud desktop according to the resource allocation strategy.

[0081] Thirdly, this application provides an electronic device, including: a memory and a processor;

[0082] The memory stores the instructions that the computer executes;

[0083] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0084] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0085] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0086] This application provides a method, apparatus, and device for dynamic allocation of cloud desktop resources based on a hybrid model, relating to the field of cloud desktop management technology. The method includes: acquiring behavioral data of a target user regarding cloud desktops; inputting the behavioral data into a dual-channel intelligent prediction model; performing a correlation assessment between the target user's profile and resource needs based on the behavioral data to obtain a first resource demand prediction result; and performing a time-series assessment of the target user's resource needs based on the first resource demand prediction result to obtain a final resource demand prediction result for the target user. The model parameters of the dual-channel intelligent prediction model are determined based on pre-learned correlation patterns between historical behavioral data and resource needs; and adjusting the resource allocation of the cloud desktops based on the resource demand prediction result. This application acquires behavioral data of target users regarding cloud desktops and inputs this data into a dual-channel intelligent prediction model. The model predicts the resource needs of the target users based on the behavioral data, thus obtaining the predicted resource needs. The model parameters of the dual-channel intelligent prediction model are determined based on pre-learned correlation patterns between historical behavioral data and resource needs. This dual-channel intelligent prediction model ensures that the predicted resource needs more closely match the actual results. Based on the predicted resource needs, the resource allocation of the cloud desktop is adjusted, thereby improving the resource utilization rate of the cloud desktop system. Attached Figure Description

[0087] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0088] Figure 1 A flowchart illustrating the cloud desktop resource dynamic allocation method based on a hybrid model provided in this application embodiment. Figure 1 ;

[0089] Figure 2 A schematic diagram illustrating real-time behavioral data and user profiles provided in the embodiments of this application;

[0090] Figure 3 This is a schematic diagram of the LSTM structure provided in the embodiments of this application;

[0091] Figure 4 A flowchart illustrating the cloud desktop resource dynamic allocation method based on a hybrid model provided in this application embodiment. Figure 2 ;

[0092] Figure 5 A schematic diagram of the structure of a cloud desktop resource dynamic allocation device based on a hybrid model provided in an embodiment of this application;

[0093] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0094] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0095] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0096] Currently, VDI typically delivers desktop resources using static resource allocation and predefined templates. This involves creating different types and configurations of desktop pools based on template virtual machines, then authorizing each desktop pool to a matching user based on their application scenario, and elastically scheduling the number of cloud desktops in the pool based on the number of user terminal connections. However, this approach has the following problems:

[0097] As the core organizational logic of VDI, the desktop pool achieves efficient delivery through centralized management, standardized configuration and resource pooling. The elastic scheduling of the desktop pool can improve the efficiency of creating and recycling cloud desktops within the desktop pool. However, the desktops automatically created by the elastic scheduling of the desktop pool will not automatically adjust their configurations according to the actual needs of users.

[0098] All desktops in the same desktop pool have identical resource configurations, making further granular management impossible. Desktop configurations within the same desktop pool are also uniform, failing to meet users' actual resource needs.

[0099] In a cloud desktop environment, application software is usually packaged as a software suite. In the same scenario, users have similar needs for general application software, operating system version, etc. However, different user habits, different levels of software usage, and different workloads can easily lead to different user needs for resources. This results in different user experiences in the same desktop pool, and even different user experiences for the same user at different times.

[0100] To address the issues of low resource utilization, uneven distribution of user desktop resources, large fluctuations in user experience, and inability to accurately match real-time user needs in the aforementioned cloud desktop pools, this application provides a dynamic resource allocation method for cloud desktops based on a hybrid model. This method involves acquiring behavioral data of target users regarding cloud desktops; inputting this behavioral data into a dual-channel intelligent prediction model to predict the resource needs of target users; and adjusting the resource allocation of cloud desktops based on the predicted resource needs.

[0101] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0102] Figure 1 A flowchart illustrating the cloud desktop resource dynamic allocation method based on a hybrid model provided in this application embodiment. Figure 1 ,like Figure 1 As shown, the method includes:

[0103] S101. Obtain the target user's behavior data regarding the cloud desktop.

[0104] In this step, it can be understood that when allocating resources to the cloud desktop, it is necessary to obtain the target user's behavioral data regarding the cloud desktop. The specific implementation method for obtaining the target user's behavioral data regarding the cloud desktop can be selected according to the actual situation.

[0105] For example, the cloud desktop system collects real-time behavioral data of target users by deploying a resource monitoring component. After collecting the real-time behavioral data, it analyzes it to determine the user profile of the target user. It should be noted that the behavioral data described in S101 includes real-time behavioral data and the corresponding user profile.

[0106] Figure 2 This is a schematic diagram illustrating real-time behavioral data and user profiles provided in embodiments of this application. Figure 2 As shown, real-time behavioral data includes application activity, real-time load of computing resources such as CPU, memory, and GPU, storage resource performance such as disk IOPS, network latency, and mouse and keyboard event frequency, where mouse and keyboard event frequency can also be referred to as mouse and keyboard input frequency. User profiles include historical resource peaks, frequently used application combinations, task sensitivity tags, and GPU memory consumption patterns.

[0107] Furthermore, the user profile dimensions include static profiles, dynamic profiles, and potential demand profiles, which are represented as follows:

[0108] Static profile: User roles (developers, designers, clerks, students), departments / projects, list of commonly used software (CAD, VS Code, Office, Chrome), and baseline basic resource requirements.

[0109] Dynamic profile: resource usage patterns (peak hours, typical load), application preference combinations, task sensitivity (such as tolerance for latency), and historical performance experience scores.

[0110] Potential demand profile: Based on historical data mining, identify resources that users may need but have not explicitly expressed, such as a user who frequently needs more memory after opening large files.

[0111] S102. Input the behavioral data into the dual-channel intelligent prediction model, evaluate the correlation between the target user's profile and resource demand based on the behavioral data, obtain the first resource demand prediction result, and evaluate the target user's resource demand in a time series based on the first resource demand prediction result to obtain the target user's resource demand prediction result. The model parameters of the dual-channel intelligent prediction model are determined based on the correlation pattern between historical behavioral data and resource demand learned in advance.

[0112] After obtaining the behavioral data via S102, the data needs to be input into the dual-channel intelligent prediction model. This allows the model to predict the resource needs of the target user based on the behavioral data, thus obtaining the predicted resource needs. The model parameters of the dual-channel intelligent prediction model are determined based on the pre-learned correlation patterns between historical behavioral data and resource needs.

[0113] This means that the dual-channel intelligent prediction model is a pre-trained model, the training process of which aims to endow it with the ability to identify and understand complex data relationships. Specifically, the training set of the dual-channel intelligent prediction model includes historical behavioral data and a model of the inherent correlation between this historical behavioral data and resource requirements.

[0114] By inputting these data as samples into the dual-channel prediction model for learning, the dual-channel intelligent prediction model can automatically extract, identify, and memorize the evolution patterns between historical behavioral data and resource demands.

[0115] S103. Adjust the resource allocation of the cloud desktop based on the resource demand forecast results.

[0116] In this step, it can be understood that the resource allocation of the cloud desktop is adjusted based on the resource demand prediction results output by the dual-channel intelligent prediction model. Different resource demand prediction results correspond to different resource allocation methods.

[0117] This application embodiment acquires target user behavior data regarding cloud desktops and inputs this data into a dual-channel intelligent prediction model. The dual-channel intelligent prediction model predicts the target user's resource needs based on the behavior data, thus obtaining the target user's resource demand prediction result. The model parameters of the dual-channel intelligent prediction model are determined based on pre-learned correlation patterns between historical behavior data and resource needs. The dual-channel intelligent prediction model ensures that the obtained resource demand prediction result is more consistent with the actual result. Based on the resource demand prediction result, the resource allocation of the cloud desktop is adjusted, thereby improving the resource utilization rate of the cloud desktop system.

[0118] Based on the above embodiments, in some examples, the dual-channel intelligent prediction model described in S102 includes a time-series prediction model and a profile inference model. Behavioral data is input into the dual-channel intelligent prediction model, and the profile of the target user and resource needs are correlated and evaluated based on the behavioral data to obtain a first resource demand prediction result. Then, based on the first resource demand prediction result, the resource needs of the target user are evaluated in a time-series manner to obtain the target user's resource demand prediction result. This includes: inputting behavioral data into the profile inference model, extracting global features of the behavioral data, and evaluating the profile of the target user and resource needs based on the global features to obtain a first resource demand prediction result; inputting the behavioral data and the first resource demand prediction result into the time-series prediction model, and evaluating the resource needs of the target user in a time-series manner based on the behavioral data and the first resource demand prediction result to obtain the resource demand prediction result.

[0119] In the example above, it can be understood that "dual-channel" refers to the time-series prediction channel and the profile inference channel, and the corresponding dual-channel intelligent prediction model includes a time-series prediction model and a profile inference model. Accordingly, to determine the resource demand prediction result for the target user, behavioral data must first be input into the profile inference model, where global features of the behavioral data are extracted. Then, based on these global features, the profile of the target user and their resource demand are correlated and evaluated to obtain the first resource demand prediction result.

[0120] After determining the first resource demand forecast result, the behavioral data and the first resource demand forecast result are used as input features of the time series forecast model. The time series forecast model performs a time series evaluation of the target user's resource demand based on the behavioral data and the first resource demand forecast result, and obtains the resource demand forecast result.

[0121] It should be noted that the time-series prediction model is used to capture the second-level fluctuation pattern of resource demand; the profile reasoning model is used to explore the non-linear relationship between user profiles and resources, convert dynamic profiles (such as commonly used CAD software) into resource demand weights, and calculate the current behavior-profile baseline to generate abnormal features (such as sudden mutation operation detection).

[0122] The embodiments of this application, through a dual-channel prediction model, can improve the accuracy of resource demand prediction results for target users.

[0123] Optionally, behavioral data is input into the profile inference model to extract global features from the behavioral data. Based on these global features, the profile of the target user and resource needs are correlated and evaluated to obtain a first resource demand prediction result. This includes: extracting multimodal features from the behavioral data; fusing the multimodal features to obtain a fused feature representation; inputting the feature representation into the profile inference model to extract global features from the feature representation; and correlating and evaluating the profile of the target user and resource needs based on the global features to obtain a first resource demand prediction result. It is understood that the behavioral data needs to be processed before being input into the profile inference model.

[0124] Specifically, multimodal features are extracted from behavioral data, such as clickstream, time series, and device information. The extracted multimodal features are then fused to obtain a fused feature representation. For example, fusing multimodal features to obtain a fused feature representation includes: performing feature fusion based on a gating function to obtain the fused feature representation; or, based on an adaptive cross-attention mechanism, interactively fusing multimodal features by constructing an intermodal semantic association matrix and a dynamic attention weight allocation strategy to generate the fused feature representation.

[0125] By fusing multimodal features, richer semantic information can be provided for the subsequent dual-channel intelligent prediction model, improving its prediction accuracy. The fused feature representation is then input into the profile inference model to extract global features. Based on these global features, the profile of the target user and their resource needs are correlated and evaluated to obtain the first resource demand prediction result.

[0126] There are two main methods for fusing multimodal features: the first is to introduce a gating function to fuse multimodal features; the second is to introduce an adaptive cross-attention mechanism to fuse multimodal features.

[0127] In the first implementation, the gating function is the core mechanism in fusing multimodal features. This gating function serves as a learnable, data-driven information filtering and weight allocation mechanism. Specifically, after obtaining the initial representations of each modality's features, the gating function dynamically generates a set of gating weights based on the inherent semantics of these initial representations and the correlations between modalities. These gating weights act as a filter, quantifying the contribution of each modality's features to the final fused representation, thereby selectively enhancing key information and effectively suppressing redundant or noisy information. By weighted summing or weighted concatenation of each modality's features with its corresponding gating weights, followed by a nonlinear transformation, the fused feature representation is obtained.

[0128] In the second implementation, the adaptive cross-attention mechanism is the core mechanism for fusing multimodal features. Specifically, based on the adaptive cross-attention mechanism, multimodal features are interactively fused by constructing an intermodal semantic association matrix and a dynamic attention weight allocation strategy to generate a fused feature representation. This includes: performing intramodal feature encoding on the multimodal features to generate an initial modal feature representation containing local semantic information; calculating intermodal semantic similarity based on the context information of the initial modal feature representation; constructing an intermodal semantic association matrix based on the intermodal semantic similarity, which is used to characterize the semantic correspondence between each modal feature; adaptively determining the dynamic attention weight allocation strategy based on the intermodal semantic association matrix and context information; and performing a weighted summation of the initial modal feature representation based on the dynamic attention weight allocation strategy to generate the fused feature representation.

[0129] In the second approach, the core of the adaptive cross-attention mechanism lies in allowing one modality feature to focus on another or more modality features, thereby capturing the deep semantic relationships and dependencies between different modality features. Specifically, after generating initial modality feature representations containing local semantic information, this approach further includes the following steps: constructing semantic associations between different modalities, determining dynamic adaptive attention weights, and fusing different modality features.

[0130] The construction of semantic associations between different modalities is achieved by generating an intermodal semantic association matrix. The principle behind this matrix construction is to use the feature representation of one modality as a query and interact with the feature representations of other modalities as keys for computation. This interaction computation can include dot product operations, concatenation followed by linear transformation, or multilayer perceptrons, aiming to quantify and generate an attention score or similarity matrix that reflects the semantic relevance between features of different modalities—that is, the intermodal semantic association matrix. The intermodal semantic association matrix quantifies the semantic relevance between features of different modalities, providing a data-driven foundation for subsequent attention allocation.

[0131] By constructing an intermodal semantic association matrix and contextual information, a dynamic attention weight allocation strategy is adaptively determined. This means that the uncertainty between different modal feature representations is considered when determining the dynamic attention weight allocation strategy. By modeling the uncertainty between different modal feature representations, it is possible to ensure dynamic enhancement of key semantic information in the subsequent fusion process between different modal features, thereby reducing the impact of low-quality multimodal data on the fusion result.

[0132] Based on the above embodiments, in some examples, behavioral data and the first resource demand prediction result are input into a time-series prediction model. Based on the behavioral data and the first resource demand prediction result, a time-series evaluation of the target user's resource demand is performed to obtain a resource demand prediction result. This includes: inputting feature representation and the first resource demand prediction result into the time-series prediction model; performing a time-series evaluation of the target user's resource demand based on the feature representation and the first resource demand prediction result to obtain a second resource demand prediction result; and performing a weighted summation of the first and second resource demand prediction results, and determining the weighted summation result as the resource demand prediction result. The weights of the weighted summation are adaptively adjusted based on the prediction errors of the time-series prediction model and the profile reasoning model.

[0133] In the above example, it can be understood that the resource demand results of the target user are determined based on the first resource demand prediction result and the second resource demand prediction result, wherein the first resource demand prediction result is determined by the profile inference model; and the second resource demand prediction result is determined by the time series prediction model.

[0134] Specifically, to obtain the second resource demand prediction result, the first resource demand prediction result and the aforementioned feature representation need to be used as input features and fed into the time series prediction model. In the time series prediction model, based on the feature representation and the first resource demand prediction result, a time series evaluation of the target user's resource demand is performed to obtain the second resource demand prediction result.

[0135] Optionally, the time-series prediction model is a Long Short-Term Memory (LSTM) network model incorporating a gated attention mechanism; the profile reasoning model is a gradient-boosting decision model (Light Gradient Boosting Machine, or LightGBM). The gated attention mechanism operates at the top-level output sequence of the LSTM encoder by assigning dynamic weights to the hidden states at all time steps. This is where the gated attention mechanism excels in handling long sequences with LSTM. By assigning dynamic weights to different time steps, the LSTM's attention can focus more on resource usage with similar prediction times, thereby improving the prediction accuracy of the LSTM.

[0136] It should be noted that LSTM is a special type of recurrent neural network structure. Its core idea is to use memory units to store, update, and transfer information, while controlling the flow of information through input gates, forget gates, and output gates. Figure 3 This is a schematic diagram of the LSTM structure provided in an embodiment of this application. Figure 3 As shown, the calculation formula is as follows:

[0137] Forget Gate: ;

[0138] Input Gate: ;

[0139] Cell State Update: ;

[0140] Output Gate: .

[0141] in, The input represents time t; The output represents time t; This indicates the hidden state at the previous moment; Indicates the state of the memory unit at the previous moment; Represents the sigmoid function; represents element-wise product; W and b represent the weight matrix and bias vector, respectively.

[0142] Furthermore, the first resource demand forecast result and the second resource demand forecast result are weighted and summed, and the result of the weighted sum is determined as the resource demand forecast result for the target user. The resource demand forecast result can be expressed by the following formula:

[0143] .

[0144] in, This represents the prediction result of the LSTM model, i.e., the second resource demand prediction result; This represents the prediction result of the LightGBM model, i.e., the first resource demand prediction result; This represents the weighting coefficients adjusted based on model performance.

[0145] Furthermore, the weights of the weighted summation are adaptively adjusted based on the prediction errors of the time-series prediction model and the profiling inference model. This is typically achieved by optimizing the prediction errors of the time-series prediction model and the profiling inference model on the cross-validation process or validation set.

[0146] Optionally, the dynamic adjustment mechanism for the weights is as follows: when (current GPU utilization > 50% of the image baseline) and (mouse movement acceleration) are detected, the weights are temporarily increased. Weight 0.2; Initial login period (first 2 minutes): =0.8 This is to enhance time series prediction.

[0147] It should be noted that the time-series prediction model captures the temporal characteristics of user resource needs and strengthens the impact of key time points through an attention mechanism; the profile inference model explores the non-linear relationship between user profiles and resource needs, such as the high GPU requirements associated with commonly used CAD software.

[0148] Based on the above embodiments, S103 describes adjusting the resource allocation of the cloud desktop according to the resource demand prediction results, including: determining the resource priority of the cloud desktop according to the resource demand prediction results; and adjusting the resource allocation of the cloud desktop according to the resource priority by using the virtualization resource hot-adding mechanism.

[0149] In the above implementation, it can be understood that when adjusting the resource allocation of cloud desktops based on resource demand forecasts, the resource priority of the cloud desktops must first be determined based on the resource demand forecasts. Then, based on the determined resource priorities, the resource allocation of the cloud desktops is adjusted using the virtualization resource hot-adding mechanism. The virtualization resource hot-adding mechanism refers to the ability to dynamically add computing resources such as CPU and memory to running virtual machines without interrupting virtual machine services.

[0150] The aforementioned hot-add mechanism for virtualized resources allows users to dynamically adjust resource allocation based on actual needs. For example, when users are performing high-load tasks such as 3D modeling or rendering, the system can automatically allocate high-performance GPUs and sufficient memory; while during document writing or simple code editing, resource allocation can be reduced to save costs. This on-demand allocation method ensures efficient resource utilization and avoids over-allocation or under-allocation of resources.

[0151] Furthermore, after determining the resource demand forecast results, some resources can be pre-allocated based on these forecasts to reduce delays in actual resource allocation. Even further, before allocating resources to cloud desktops, resource caches can be reserved in advance for high-frequency users or critical business tasks to ensure rapid response to sudden demand surges.

[0152] In one example, based on resource priority, the resource allocation of the cloud desktop is adjusted using the virtualization resource hot-add mechanism, including: obtaining the resource usage indicators of the cloud desktop; generating a resource demand event when the resource usage indicators continuously exceed the usage threshold within a set time period; responding to the resource demand event and building a virtualization resource pool from the physical server cluster; formulating an allocation strategy for the virtualization resource pool based on resource priority; and, based on the allocation strategy, calling the hot-add interface of the virtualization platform to add the required virtualization resource pool to the cloud desktop.

[0153] In this example, when adjusting resource allocation for cloud desktops, the process first determines whether the resource usage metrics of the acquired cloud desktops have continuously exceeded the usage threshold within a set time period. For example, it first determines whether the acquired resource usage metrics of the acquired cloud desktops have continuously exceeded the usage threshold within a time period T1. When it is detected that the resource usage metrics of the cloud desktops have continuously exceeded the usage threshold within a set time period, this means that there is a shortage of cloud desktop resources. At this time, a resource demand event is generated and responded to, and a virtualization resource pool is built from the physical server to increase the available resources of the cloud desktops. When increasing the available resources of cloud desktops, resource priorities must be considered, and a refined virtualization resource pool allocation strategy must be formulated based on the priorities, current resource demands, and other factors. Based on the formulated allocation strategy, the hot-add interface of the virtualization platform is called to add the required virtualization resource pool to the cloud desktops.

[0154] This application embodiment constructs a virtualized resource pool from a physical server cluster, determines the allocation strategy of the virtualized resource pool, and calls the hot-add interface of the virtualization platform to complete the addition of the required virtualized resource pool to the cloud desktop. This can quickly meet the resource needs of the cloud desktop and improve the user experience when there is a shortage of resources on the cloud desktop.

[0155] Based on the above embodiments, the cloud desktop resource dynamic allocation method based on a hybrid model provided in this application adjusts the resource allocation of the cloud desktop according to the resource demand prediction results, including: modeling the cloud desktop resource allocation problem as a Markov decision model, wherein the state space of the Markov decision model includes the current resource usage of the cloud desktop, the resource demand prediction results, and historical resource allocation records; the action space of the Markov decision model includes the resource adjustment allocation amount of the cloud desktop; and the reward function of the Markov decision model includes a weighted combination based on resource utilization, user response time, and resource allocation cost; constructing a reinforcement learning agent based on the Markov decision model; iteratively training the reinforcement learning agent with the cloud desktop to learn the resource allocation strategy of the cloud desktop, wherein the training process is to maximize the reward function; and adjusting the resource allocation of the cloud desktop according to the resource allocation strategy.

[0156] In this embodiment, it can be understood that this application also provides an implementation method for adjusting the resource allocation of cloud desktops based on resource demand prediction results. In this implementation method, the cloud desktop resource allocation problem is modeled as a Markov decision model. Markov decision models are applicable to problems in an uncertain environment where a series of decision actions are taken to achieve a certain goal. Since the resource demands and user behaviors of cloud desktops are inherently dynamic and uncertain, this application introduces a Markov decision model to optimize the resource allocation of cloud desktops.

[0157] Furthermore, a reinforcement learning agent is constructed using a Markov decision model. This reinforcement learning agent is endowed with the ability to perceive the state space of the Markov decision model, understand and select actions in the action space of the Markov decision model, and evaluate its behavior based on the reward function of the Markov decision model.

[0158] The fundamental purpose of constructing this reinforcement learning agent is to enable it to learn and ultimately form an optimal or suboptimal resource allocation strategy through continuous interaction with the cloud desktop. This strategy guides the agent to choose the resource adjustment action that maximizes long-term cumulative rewards in any given state. This method of constructing a reinforcement learning agent based on a Markov decision model gives the resource allocation strategy the ability to adapt and continuously learn, allowing a deterministic resource allocation strategy to effectively cope with the uncertainty of the cloud desktop.

[0159] Next, taking the temporal prediction model as a long short-term memory network model that integrates gating attention mechanism, also known as LSTM-Attention, and the portrait reasoning model as LightGBM as an example, we will explain how to use the cloud desktop resource dynamic allocation method based on the hybrid model provided in the embodiments of this application. Figure 4A flowchart illustrating the cloud desktop resource dynamic allocation method based on a hybrid model provided in this application embodiment. Figure 2 .like Figure 4 As shown, the method includes:

[0160] 1. The cloud desktop system collects real-time behavior data of cloud desktop users by deploying resource monitoring components, and analyzes the real-time behavior data to obtain user profiles.

[0161] 2. Generate feature engineering based on real-time behavioral data and user profiles.

[0162] 3. Perform dual-channel prediction on feature engineering to obtain the user's resource demand prediction results.

[0163] Dual-channel prediction for feature engineering refers to performing time-series prediction and profile inference on the feature engineering separately, resulting in a second resource demand prediction result and a first resource demand prediction result. Then, the second and first resource demand prediction results are dynamically weighted and fused to complete multi-dimensional resource demand prediction, such as predicting resource demands for vCPU, memory, GPU, and storage IOPS. It should be noted that the specific principles of resource demand prediction have been explained in detail in the previous embodiments; therefore, they will not be repeated here.

[0164] 4. Based on the results of the multi-dimensional resource demand forecast, resources are dynamically allocated in conjunction with the virtualization resource hot-addition mechanism.

[0165] In summary, this application proposes a dynamic resource allocation method for cloud desktops based on user profiles and behavior awareness. It intelligently predicts resource usage through machine learning, combining real-time data collection and user profile characteristics with a hybrid prediction model that integrates multimodal features: LSTM-Attention and LightGBM. Furthermore, by collecting real-time cloud desktop user data, including behavioral data, resource load data, and profile data, dual-channel prediction using LSTM-Attention and LightGBM is performed based on feature engineering. Finally, the prediction results of the two models are dynamically weighted and adjusted using a dynamic adjustment mechanism, which can efficiently and relatively accurately predict the real-time resource needs of users in the cloud desktop system. Moreover, based on the predictive capability, a virtualization resource hot-addition mechanism is combined to ultimately achieve dynamic adjustment of resource usage, realizing efficient and fine-grained resource allocation. Compared to the single elastic scaling mechanism of the cloud desktop pool, the overall resource utilization rate of the cloud desktop will be greatly improved, while significantly enhancing the user experience and increasing user stickiness to the cloud desktop.

[0166] It is understood that the cloud desktop resource dynamic allocation method based on user profiles and behavior awareness proposed in this application is based on existing desktop pool elastic scheduling and application layering and packaging. Through deep learning, it dynamically obtains the user's resource consumption over a certain period of time and dynamically adjusts the user's cloud desktop resources. This method is driven by user needs, builds refined user profiles, perceives user behavior and load in real time, and uses intelligent prediction algorithms to achieve on-demand, real-time elastic scaling of cloud desktop computing, memory, storage, GPU and other resources to maximize the use of overall resources. Unlike the elastic scaling of the number of cloud desktops in the desktop pool, this method focuses elastic scaling on the user's needs for the cloud desktop resources they are using, thereby improving resource utilization, ensuring the performance of critical businesses, improving user experience and reducing overall operating costs.

[0167] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0168] Figure 5 A schematic diagram of the structure of the cloud desktop resource dynamic allocation device based on a hybrid model provided in this application embodiment is shown below. Figure 5 As shown, the cloud desktop resource dynamic allocation device 500 based on a hybrid model provided in this embodiment includes:

[0169] Module 501 is used to acquire behavioral data of the target user regarding the cloud desktop.

[0170] The determination module 502 is used to input behavioral data into the dual-channel intelligent prediction model, evaluate the correlation between the target user's profile and resource demand based on the behavioral data, obtain the first resource demand prediction result, and evaluate the target user's resource demand in a time series based on the first resource demand prediction result to obtain the target user's resource demand prediction result. The model parameters of the dual-channel intelligent prediction model are determined based on the correlation pattern between historical behavioral data and resource demand learned in advance.

[0171] Adjustment module 503 is used to adjust the resource allocation of the cloud desktop based on the resource demand prediction results.

[0172] In one possible implementation, the dual-channel intelligent prediction model includes a time-series prediction model and a profile reasoning model, and the determining module 502 is specifically used for:

[0173] Behavioral data is input into the profile reasoning model, global features of the behavioral data are extracted, and based on the global features, the profile of the target user and resource needs are correlated and evaluated to obtain the first resource needs prediction result.

[0174] The behavioral data and the first resource demand prediction results are input into the time series prediction model. Based on the behavioral data and the first resource demand prediction results, the resource demand of the target user is evaluated in a time series, and the resource demand prediction results are obtained.

[0175] In one possible implementation, the determining module 502 is specifically used for:

[0176] Extracting multimodal features from behavioral data;

[0177] Multimodal features are fused to obtain the fused feature representation;

[0178] The feature representation is input into the profile reasoning model to extract global features from the feature representation. Based on the global features, the profile of the target user and the resource requirements are correlated and evaluated to obtain the first resource requirement prediction result.

[0179] In one possible implementation, the cloud desktop resource dynamic allocation device based on the hybrid model further includes a fusion module (not shown), which is specifically used for:

[0180] Based on the gating function, feature fusion is performed on multimodal features to obtain the fused feature representation;

[0181] Alternatively, based on an adaptive cross-attention mechanism, multimodal features can be interactively fused by constructing an intermodal semantic association matrix and a dynamic attention weight allocation strategy to generate a fused feature representation.

[0182] In one possible implementation, the fusion module is specifically used for:

[0183] Intramodal feature encoding is performed on multimodal features to generate an initial modal feature representation containing local semantic information;

[0184] Based on the contextual information of the initial modal feature representation, calculate the semantic similarity between modalities;

[0185] Based on the semantic similarity between modalities, an intermodal semantic association matrix is ​​constructed. The intermodal semantic association matrix is ​​used to represent the semantic correspondence between features of each modality.

[0186] Based on the intermodal semantic association matrix and contextual information, an adaptive dynamic attention weight allocation strategy is determined.

[0187] Based on a dynamic attention weight allocation strategy, the initial modal feature representations are weighted and summed to generate fused feature representations.

[0188] In one possible implementation, the determining module 502 is specifically used for:

[0189] The feature representation and the first resource demand prediction result are input into the time series prediction model. Based on the feature representation and the first resource demand prediction result, the resource demand of the target user is evaluated in time series to obtain the second resource demand prediction result.

[0190] The first resource demand forecast result and the second resource demand forecast result are weighted and summed, and the weighted sum is determined as the resource demand forecast result. The weights of the weighted sum are obtained by optimizing the prediction errors of the time series forecast model and the profile inference model in the cross-validation process or validation set.

[0191] In one possible implementation, the temporal prediction model is a long short-term memory network model that incorporates a gated attention mechanism; the profiling reasoning model is a decision model based on gradient boosting.

[0192] In one possible implementation, the adjustment module 503 is specifically used for:

[0193] Based on the resource demand forecast results, determine the resource priority of the cloud desktop;

[0194] Based on resource priority, the resource allocation of cloud desktops is adjusted using the virtualization resource hot-add mechanism.

[0195] In one possible implementation, the adjustment module 503 is specifically used for:

[0196] Obtain resource usage metrics for cloud desktops;

[0197] When the resource usage index exceeds the usage threshold continuously within a set period, a resource demand event is generated.

[0198] In response to resource demand events, a virtual resource pool is built from the physical server cluster;

[0199] Based on resource priority, formulate a virtualization resource pool allocation strategy;

[0200] Based on the allocation strategy, the hot-add interface of the virtualization platform is invoked to add the required virtualization resource pool to the cloud desktop.

[0201] In one possible implementation, the adjustment module 503 is specifically used for:

[0202] The resource allocation problem of cloud desktops is modeled as a Markov decision model. The state space of the Markov decision model includes the current resource usage of cloud desktops, the predicted resource demand, and the historical resource allocation records. The action space of the Markov decision model includes the resource adjustment allocation amount of cloud desktops. The reward function of the Markov decision model includes a weighted combination based on resource utilization, user response time, and resource allocation cost.

[0203] Construct a reinforcement learning agent based on the Markov decision model;

[0204] By using a reinforcement learning agent, iterative training is conducted with the cloud desktop to learn the cloud desktop's resource allocation strategy. The training process aims to maximize the reward function.

[0205] Adjust the resource allocation of the cloud desktop according to the resource allocation strategy.

[0206] The cloud desktop resource dynamic allocation device based on the hybrid model provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0207] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, a processing module can be a separate processing element, or it can be integrated into a chip within the device. Alternatively, it can be stored as program code in the device's memory, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0208] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a System-On-a-Chip (SOC).

[0209] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 600 provided in this application embodiment may include: a processor 601, and a memory 602 communicatively connected to the processor, wherein:

[0210] The memory stores the instructions that the computer executes;

[0211] The processor executes computer execution instructions stored in memory to implement the method described in the foregoing method embodiments.

[0212] It should be understood that processor 601 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the application can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor. Memory 602 may include high-speed random access memory (RAM), and may also include non-volatile memory (NVM), such as at least one disk storage device, or a USB flash drive, external hard drive, read-only memory, disk, or optical disc, etc.

[0213] Optionally, the electronic device 600 may also include a communication interface 603. In specific implementations, if the communication interface 603, memory 602, and processor 601 are implemented independently, they can be interconnected via a bus to complete communication. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not imply that there is only one bus or one type of bus.

[0214] Optionally, in a specific implementation, if the communication interface 603, memory 602, and processor 601 are integrated on a single chip, then the communication interface 603, memory 602, and processor 601 can communicate through an internal interface.

[0215] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the methods described in any of the foregoing embodiments.

[0216] It is understood that the computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0217] An exemplary computer-readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the computer-readable storage medium. Of course, the computer-readable storage medium can also be a component of the processor. The processor and the computer-readable storage medium can reside in an ASIC. Alternatively, the processor and the computer-readable storage medium can exist as discrete components in an electronic device.

[0218] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a computer-readable storage medium, include several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application.

[0219] This application also provides a computer program product, including a computer program that, when executed, implements the method described in any of the foregoing embodiments.

[0220] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0221] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0222] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0223] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0224] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A cloud desktop resource dynamic allocation method based on a mixed model, characterized in that, The method comprises: obtaining behavior data of a target user for a cloud desktop; inputting the behavior data into a double-channel intelligent prediction model, performing correlation evaluation on the portrait and resource demand of the target user according to the behavior data, obtaining a first resource demand prediction result, and performing time series evaluation on the resource demand of the target user according to the first resource demand prediction result to obtain a resource demand prediction result of the target user, wherein model parameters of the double-channel intelligent prediction model are determined based on a pre-learned correlation mode between historical behavior data and resource demand; adjusting resource allocation of the cloud desktop according to the resource demand prediction result; The double-channel intelligent prediction model comprises a time series prediction model and a portrait inference model, and the step of inputting the behavior data into the double-channel intelligent prediction model, performing correlation evaluation on the portrait and resource demand of the target user according to the behavior data, obtaining a first resource demand prediction result, and performing time series evaluation on the resource demand of the target user according to the first resource demand prediction result to obtain a resource demand prediction result of the target user comprises: inputting the behavior data into the portrait inference model, extracting global features of the behavior data, and performing correlation evaluation on the portrait and resource demand of the target user according to the global features to obtain a first resource demand prediction result; inputting the behavior data and the first resource demand prediction result into the time series prediction model, and performing time series evaluation on the resource demand of the target user according to the behavior data and the first resource demand prediction result to obtain a resource demand prediction result.

2. The method of claim 1, wherein, The step of inputting the behavior data into the portrait inference model, extracting global features of the behavior data, and performing correlation evaluation on the portrait and resource demand of the target user according to the global features to obtain a first resource demand prediction result comprises: extracting multi-modal features from the behavior data; fusing the multi-modal features to obtain a fused feature representation; inputting the feature representation into the portrait inference model to extract global features of the feature representation, and performing correlation evaluation on the portrait and resource demand of the target user according to the global features to obtain the first resource demand prediction result.

3. The method of claim 2, wherein, The step of fusing the multi-modal features to obtain a fused feature representation comprises: performing feature fusion on the multi-modal features based on a gating function to obtain a fused feature representation; or, based on an adaptive cross-attention mechanism, by constructing an inter-modal semantic correlation matrix and a dynamic attention weight distribution strategy, the multi-modal features are interactively fused to generate a fused feature representation.

4. The method of claim 3, wherein, The step of interactively fusing the multi-modal features based on the adaptive cross-attention mechanism, by a dynamic weight distribution strategy and cross-modal semantic alignment, to generate a fused feature representation comprises: performing intra-modal feature encoding on the multi-modal features to generate an initial modal feature representation containing local semantic information; calculating inter-modal semantic similarity based on context information of the initial modal feature representation; According to the inter-modal semantic similarity, an inter-modal semantic correlation matrix is constructed, which is used to represent the semantic corresponding relationship between features of each modality; Based on the inter-modal semantic correlation matrix and the context information, the dynamic attention weight distribution strategy is adaptively determined; Based on the dynamic attention weight distribution strategy, the initial modality feature representation is weighted and summed to generate a fused feature representation.

5. The method according to any one of claims 2 to 4, characterized in that, The behavior data and the first resource demand prediction result are input into the time series prediction model, and the resource demand of the target user is time series evaluated according to the behavior data and the first resource demand prediction result, to obtain a resource demand prediction result, including: The feature representation and the first resource demand prediction result are input into the time series prediction model, and the resource demand of the target user is time series evaluated according to the feature representation and the first resource demand prediction result, to obtain a second resource demand prediction result; The first resource demand prediction result and the second resource demand prediction result are weighted and summed, and the result of the weighted sum is determined as the resource demand prediction result, wherein the weight of the weighted sum is adaptively adjusted according to the prediction error of the time series prediction model and the portrait inference model.

6. The method of claim 1, wherein, The time series prediction model is a long short-term memory network model with a fusion gate attention mechanism; and the portrait inference model is a gradient boosting-based decision model.

7. The method of claim 1, wherein, According to the resource demand prediction result, the resource allocation of the cloud desktop is adjusted, including: According to the resource demand prediction result, the resource priority of the cloud desktop is determined; According to the resource priority, the virtualization resource hot addition mechanism is used to adjust the resource allocation of the cloud desktop.

8. The method of claim 7, wherein, According to the resource priority, the virtualization resource hot addition mechanism is used to adjust the resource allocation of the cloud desktop, including: Obtain the resource usage index of the cloud desktop; When the resource usage index continuously exceeds the usage threshold within a set time length, a resource demand event is generated; In response to the resource demand event, a virtualization resource pool is constructed from a physical server cluster; According to the resource priority, an allocation strategy of the virtualization resource pool is formulated; Based on the allocation strategy, a hot addition interface of a virtualization platform is called to add the required virtualization resource pool to the cloud desktop.

9. The method of claim 1, wherein, According to the resource demand prediction result, the resource allocation of the cloud desktop is adjusted, including: The resource allocation problem of the cloud desktop is modeled as a Markov decision model, wherein the state space of the Markov decision model includes the current resource usage of the cloud desktop, the resource demand prediction result, and historical resource allocation records, the action space of the Markov decision model includes the resource adjustment allocation amount of the cloud desktop, and the reward function of the Markov decision model includes a weighted combination based on resource utilization, user response time, and resource allocation cost; Based on the Markov decision model, a reinforcement learning agent is constructed; The reinforcement learning agent is iteratively trained with the cloud desktop to learn a resource allocation strategy of the cloud desktop, wherein a training process is to maximize the reward function; According to the resource allocation strategy, the resource allocation of the cloud desktop is adjusted.

10. A cloud desktop resource dynamic allocation device based on a hybrid model, characterized in that, Comprising: An acquisition module is configured to acquire behavior data of a target user for a cloud desktop; A determination module is configured to input the behavior data into a double-channel intelligent prediction model, to perform resource demand prediction on the target user according to the behavior data, to obtain a resource demand prediction result of the target user, and to determine model parameters of the double-channel intelligent prediction model based on a pre-learned association mode between historical behavior data and resource demand; An adjustment module is configured to adjust resource allocation of the cloud desktop according to the resource demand prediction result; The double-channel intelligent prediction model comprises a time series prediction model and a portrait inference model, and the determination module is specifically configured to: Input the behavior data into the portrait inference model, extract global features of the behavior data, and perform association evaluation on a portrait and resource demand of the target user according to the global features, to obtain a first resource demand prediction result; Input the behavior data and the first resource demand prediction result into the time series prediction model, and perform time series evaluation on the resource demand of the target user according to the behavior data and the first resource demand prediction result, to obtain a resource demand prediction result.

11. An electronic device, comprising: Comprising: A memory, a processor; The memory stores computer execution instructions; The processor executes the computer execution instructions stored in the memory, so that the processor executes the method of any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by the processor to implement the method of any one of claims 1-9.

Images