AI-driven multi-tenant cloud computing resource optimal allocation method and system

Abstract

The invention discloses an AI-driven multi-tenant cloud computing resource optimal allocation method and system, and belongs to the field of artificial intelligence, and the method comprises the steps: constructing a tenant resource dynamic portrait, and compressing the tenant resource dynamic portrait into a tensor containing a resource type, a time window and disturbance density through a three-dimensional residual attention network based on a resource request, a response time delay and a load jitter sequence; extracting resource competition conflicts among tenants by using a gating map neural encoder, and outputting a conflict risk map; a hierarchical scheduling space is constructed based on the atlas, a throughput maximization and return optimal strategy is trained, and a periodic switching controller is adopted to realize exploration and convergence balance; carrying out robustness screening on scheduling actions, and eliminating sensitive actions by utilizing a self-adaptive annealing mechanism to generate a stable strategy set; and executing resource allocation and continuously collecting microcosmic utilization data to guide subsequent strategy fine adjustment. The method has the beneficial effects that efficient and stable allocation of resource scheduling is realized, and the resource utilization rate and the system robustness in a multi-tenant cloud environment are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and more specifically, to an AI-driven multi-tenant cloud computing resource optimization allocation method and system. Background Art

[0002] With the rapid development of cloud computing technology, multi-tenant architectures have become the mainstream approach to resource sharing. Traditional resource allocation strategies often rely on static rules or heuristic algorithms. While these strategies offer some adaptability, they exhibit limitations when dealing with high concurrency, complex tenant behaviors, and dynamic load fluctuations. In recent years, machine learning-based scheduling methods have been gradually applied to cloud platforms, improving the intelligence of scheduling. However, most methods fail to effectively integrate system-level load perception, resource risk prediction, and real-time reward assessment, making it difficult to achieve comprehensive optimization in complex and changing resource environments.

[0003] The existing technologies have the following main shortcomings: First, path search strategies are mostly initialized with static indicators and lack dynamic load perception capabilities; second, resource scheduling reward modeling is crude, ignoring the stability of benefits and long-term trends; third, scheduling strategies are not robust enough in the face of sudden disturbances, which can easily lead to resource conflicts and performance degradation; fourth, disturbance modeling samples are scarce, which affects the model training effect. Summary of the Invention

[0004] The purpose of the present invention is to provide an AI-driven multi-tenant cloud computing resource optimization allocation method and system to solve the problems raised in the above-mentioned background technology: First, path search strategies are mostly initialized with static indicators and lack dynamic load perception capabilities; second, resource scheduling reward modeling is rough, ignoring the stability of benefits and long-term trends; third, the scheduling strategy is not robust enough in the face of sudden disturbances, which can easily lead to resource conflicts and performance degradation; fourth, disturbance modeling samples are scarce, which affects the model training effect.

[0005] Technical solution: The AI-driven multi-tenant cloud computing resource optimization allocation method includes the following steps:

[0006] S1. Build a dynamic profile of tenant resource usage. This is done by representing each tenant's resource request sequence, request-response latency sequence, and historical load jitter trajectory as a tensor. A three-dimensional residual attention network is used for representation compression, ultimately mapping each tenant into a three-dimensional dense tensor structure. This structure includes dimensions for resource type, time window, and perturbation density.

[0007] S2. Extract mutual exclusion features from the compressed tenant tensor representation and use a gated graph neural encoder to topologically model resource competition conflicts between tenants. The graph neural network introduces multi-dimensional edge types. Node state updates are jointly influenced by historical conflict weights and resource isomorphism coefficients, and a conflict risk factor map is output.

[0008] S3. Construct a hierarchical scheduling candidate space based on the conflict risk factor graph, use a distributed multi-objective deep policy network to learn a throughput maximization strategy and a value reward optimization strategy, respectively, and introduce a periodic switching controller between the throughput maximization strategy and the value reward optimization strategy. The periodic switching controller uses the cross-difference between the perturbation gradient and the reward stability as a period adjustment factor to achieve an exploration-convergence balance in resource scheduling.

[0009] S4. Robustness screening is performed on the generated scheduling action set. An adaptive annealing mechanism is used to dynamically update the temperature factor based on the historical fluctuation window response curve. Scheduling action sequences that are sensitive to future load changes are eliminated, and a convergent scheduling strategy set is output.

[0010] S5. Execute resource allocation under the final selected scheduling action and continuously collect micro-resource utilization traces during the allocation cycle, including CPU, memory, I / O, and GPU periodic occupancy rates as feedback signals to guide policy fine-tuning and updates during subsequent training.

[0011] Preferably, the step of constructing a dynamic profile of tenant resource usage in S1 further includes the following steps:

[0012] S1-1. Obtain the tenant's resource request frequency sequence, CPU utilization gradient sequence, and bandwidth time series density sequence as ternary time series input. Use a fixed 30-second window, extract overlapping data using a 5-second sliding window, and map them to a three-dimensional dense tensor structure. Each dimension of the three-dimensional dense tensor represents a resource item, including CPU, memory, bandwidth, GPU, time window, and load disturbance index.

[0013] S1-2. The three-dimensional tensor structure is input into the three-dimensional residual attention network for embedding compression. The three-dimensional residual attention network adopts a three-layer convolution unit and attention fusion. The channel dimensions of each layer are 64, 128, and 64. The attention module introduces an anti-normalized scalar reinforcement mechanism with the maximum load perturbation point as the anchor, and embeds a cross-scale skip connection structure in the residual path to enhance the low-frequency perturbation retention rate.

[0014] Preferably, the step S1-1 further comprises the following steps:

[0015] S1-1-1. Perform nonlinear registration on the unified time axis for resource request frequency sequences from different tenants. Use a fifth-order B-spline kernel for time domain alignment. Interpolate to fill in missing segments of low-activity tenants to ensure that the input tensor is evenly distributed across all tenants.

[0016] S1-1-2. Introduce the disturbance index based on weighted information entropy as the third-dimensional data source of the tensor. The disturbance index based on weighted information entropy is calculated by weighted sliding entropy of the tenant's resource jitter variance in the past hour. The disturbance entropy uses a 0.3-second period window and performs bidirectional exponential smoothing, and finally is normalized to the range of [0,1].

[0017] Preferably, the step S2 of extracting mutual exclusion features from the compressed tenant tensor further comprises the following steps:

[0018] S2-1. Use the principal components of tensor embedding to construct a resource competition graph with the structure G = (V, E, W)

[0019] V is the tenant set, E is the competitive edge set, and W is the edge weight. The edge weight is constructed by the cross-utilization function between resource types; the form is:

[0020]

[0021] wherein u i is the resource occupancy main vector of tenant i, is the square of the F norm of the tensor globally;

[0022] S2-2. A heterogeneous gated graph neural network is used to encode the constructed resource competition graph. In the resource competition graph, an independent edge type is introduced for each type of resource competition. The node update function integrates the time dimension gradient and the resource coupling factor. The edge update introduces a type penalty weight. During training, a residual aggregation method with attention is used to realize the dynamic feature fusion of multiple resource conflicts.

[0023] Preferably, the step S2-2 further comprises the following steps:

[0024] S2-2-1. Introduce a resource coupling factor during the update process of the heterogeneous gated graph neural network. The resource coupling factor is expressed as the correlation coefficient between the current resource main vector and the moving average main vector of the past 24 hours, and perform softmax smoothing activation during forward propagation;

[0025] S2-2-2. Add the type penalty weight in the edge update, the value of which is the ratio of the average tenant occupancy rate of the current resource type to the resource surplus rate, and use it as an explicit regularization term in the loss function to suppress the formation of misleading coupling paths on scarce resources.

[0026] Preferably, constructing the hierarchical scheduling candidate space in S3 further includes the following steps:

[0027] S3-1. Based on the conflict risk factor map, tenants are divided into different resource levels. Resource levels are unsupervisedly sorted by the cosine similarity of the principal components of the tenant and resource tensors. Asymmetric hash clustering is used, and tenants are mapped to the scheduling subspace using the locality-sensitive hashing + cluster center enforcement constraint method. Cross-tenants are eliminated by boundary distance projection;

[0028] S3-2. A multi-objective policy network training process is constructed in parallel for each scheduling layer. The policy network consists of a dual-branch structure, which trains the optimal throughput path and the value return path respectively. The action space uses an enhanced pruning mechanism to remove low-contribution actions, and finally combines them to obtain a set of candidate scheduling actions.

[0029] Preferably, the S3-2 further comprises the following steps:

[0030] S3-2-1. The optimal throughput path search prioritizes the most concurrent tenants, initializes the policy queue by sorting tenant node-level load expansion rates, and employs a gradient-guided Monte Carlo tree search. Action combinations in high-density conflict areas in the resource risk graph are eliminated.

[0031] S3-2-2. A value guidance factor is introduced in the search for the optimal return path. The value guidance factor is evaluated based on the partial derivative of the resource allocation action and the average return surface of the historical stable interval, and a priority scheduling table based on the time-weighted lag factor is constructed.

[0032] Preferably, the step of performing robustness screening on the scheduling action set in S4 further includes:

[0033] S4-1. Construct a disturbance response training set, including all burst load events, resource scheduling failures, and sampling points with sudden response latency exceeding 50ms in the past three months. Use a conditional generative adversarial network to generate disturbance-augmented samples, using the concatenation of the tenant's historical state matrix and the scheduling graph context.

[0034] S4-2. Input the candidate scheduling action sequence into the disturbance response discrimination module. The disturbance response discrimination module internally constructs a multi-channel variation gradient detector to detect the disturbance response gradient trend of the scheduling behavior under each type of resource dimension. Samples with significantly positive slope growth are eliminated, and scheduling actions with a disturbance response curve slope converging near 0 are retained.

[0035] Preferably, the GAN perturbation generator of the S4-1 perturbation response training set comprises the following steps:

[0036] S4-1-1. The conditional input vector is constructed by embedding the resource request frequency, resource history status window, and scheduling graph into three sets of concatenated representations. The concatenated vector dimension is [128, 256, 64] and is uniformly mapped to the generated input vector through a two-layer feedforward network.

[0037] S4-1-2. The discriminator uses a forward time residual network with 4 residual modules and a basic unit of Conv1D-ReLU-Dropout combination. The final output discrimination value is based on the weighted integral scoring function of the standard deviation of the time window fluctuation of the scheduling failure rate and the mean error of resource usage.

[0038] Preferably, the AI-driven multi-tenant cloud computing resource optimization allocation system relates to any AI-driven multi-tenant cloud computing resource optimization allocation method in claims 1-8, and the AI-driven multi-tenant cloud computing resource optimization allocation system is designed based on the AI-driven multi-tenant cloud computing resource optimization allocation method.

[0039] Compared with the prior art, the advantages of the present invention are:

[0040] (1) A scheduling strategy combining load expansion rate sorting and Monte Carlo tree search is introduced to achieve intelligent initialization and conflict avoidance of throughput path selection, thereby improving concurrent scheduling efficiency.

[0041] (2) By adopting the value guidance factor based on the partial derivative of the return surface and combining it with the time-weighted lag mechanism, a priority scheduling table is dynamically constructed to significantly enhance the return optimization capability of resource scheduling.

[0042] (3) Establish a disturbance response training set and introduce conditional generative adversarial network (cGAN) for sample enhancement to build a highly robust scheduling behavior screening system for emergencies.

[0043] (4) A multi-channel variation gradient detection mechanism is introduced to detect the response trend of scheduling actions to resource dimension disturbances, thereby filtering out risky scheduling behaviors and ensuring scheduling stability.

[0044] (5) Design a residual network-based discriminator and an embedded high-dimensional input vector structure to effectively improve the authenticity and relevance of GAN-generated samples and enhance the perturbation modeling capability.

[0045] (6) Implement a joint evaluation discriminant scoring function for scheduling failure rate and resource usage error, perform fine screening of GAN generated samples, and optimize the quality of robust training data. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a schematic diagram of the overall process of the AI-driven multi-tenant cloud computing resource optimization allocation method of the present invention. DETAILED DESCRIPTION

[0047] For examples, see Figure 1 , an AI-driven multi-tenant cloud computing resource optimization allocation method includes the following steps:

[0048] S1. Build a dynamic profile of tenant resource usage. This is done by representing each tenant's resource request sequence, request-response latency sequence, and historical load jitter trajectory as a tensor. A three-dimensional residual attention network is used for representation compression, ultimately mapping each tenant into a three-dimensional dense tensor structure. This structure includes dimensions for resource type, time window, and perturbation density.

[0049] S2. Extract mutual exclusion features from the compressed tenant tensor representation and use a gated graph neural encoder to topologically model resource competition conflicts between tenants. The graph neural network introduces multi-dimensional edge types. Node state updates are jointly influenced by historical conflict weights and resource isomorphism coefficients, and a conflict risk factor map is output.

[0050] S3. Construct a hierarchical scheduling candidate space based on the conflict risk factor graph, use a distributed multi-objective deep policy network to learn a throughput maximization strategy and a value reward optimization strategy, respectively, and introduce a periodic switching controller between the throughput maximization strategy and the value reward optimization strategy. The periodic switching controller uses the cross-difference between the perturbation gradient and the reward stability as a period adjustment factor to achieve an exploration-convergence balance in resource scheduling.

[0051] S4. Robustness screening is performed on the generated scheduling action set. An adaptive annealing mechanism is used to dynamically update the temperature factor based on the historical fluctuation window response curve. Scheduling action sequences that are sensitive to future load changes are eliminated, and a convergent scheduling strategy set is output.

[0052] S5. Execute resource allocation under the final selected scheduling action and continuously collect micro-resource utilization traces during the allocation cycle, including CPU, memory, I / O, and GPU periodic occupancy rates as feedback signals to guide policy fine-tuning and updates during subsequent training.

[0053] Specifically, the "building of a dynamic profile of tenant resource usage" is achieved in the following ways:

[0054] First, the resource requests (including CPU, GPU, memory, IO) of each tenant in the continuous period T = 30s are encoded into tensors is the resource dimension, 6 is the number of sliding subintervals in the time window, and 1 is the initial channel for the disturbance density. The resource request data is first normalized from 0 to 1 and then extracted using a sliding window with a step size of 5 seconds.

[0055] The tensor representation is then compressed using a three-dimensional residual attention network. This network consists of three layers of 3D convolution (channel count: 64 → 128 → 64), each followed by a channel-wise attention module. The attention module incorporates a perturbation anchor weighting mechanism: the moment of maximum resource jitter is strengthened as the primary focus of attention, and cross-scale skip connections are embedded to enhance anomaly retention.

[0056] The entire system schedules data every minute. The resource scheduler is deployed on the control node. Agents collect Prometheus metrics on the tenant side and report them to the central policy server using Kafka. The system supports switching between batch and real-time scheduling, with a specific scheduling window of two minutes. Decisions are then pushed to the container orchestration system (such as Kubernetes) for execution.

[0057] The step of constructing a dynamic profile of tenant resource usage in S1 further includes the following steps:

[0058] S1-1. Obtain the tenant's resource request frequency sequence, CPU utilization gradient sequence, and bandwidth time series density sequence as ternary time series input. Use a fixed 30-second window, extract overlapping data using a 5-second sliding window, and map them to a three-dimensional dense tensor structure. Each dimension of the three-dimensional dense tensor represents a resource item, including CPU, memory, bandwidth, GPU, time window, and load disturbance index.

[0059] S1-2. The three-dimensional tensor structure is input into the three-dimensional residual attention network for embedding compression. The three-dimensional residual attention network adopts a three-layer convolution unit and attention fusion. The channel dimensions of each layer are 64, 128, and 64. The attention module introduces an anti-normalized scalar reinforcement mechanism with the maximum load perturbation point as the anchor, and embeds a cross-scale skip connection structure in the residual path to enhance the low-frequency perturbation retention rate.

[0060] Specifically, in the specific implementation, the time window is fixed to 30 seconds, and the sliding extraction is performed once every 5 seconds, that is, 6 frames of input data (6×5s) are generated for each tenant to form a set of three-dimensional tensors The first dimension of the tensor is the resource item (CPU, memory, bandwidth, GPU), the second dimension is the sliding time slice, and the third dimension is the disturbance density indicator.

[0061] The perturbation density is constructed using tenant resource request jitter information. First, the temporal gradients of CPU utilization, GPU occupancy, and bandwidth utilization are calculated. Their variances are then calculated using a sliding window, processed using exponential smoothing (period 0.3s), and normalized to obtain the perturbation value. The maximum perturbation value is used to determine the attention anchor point. Subsequent convolutional layers use a channel-weighted mask with the anchor point location.

[0062] Said S1-1 further comprises the following steps:

[0063] S1-1-1. Perform nonlinear registration on the unified time axis for resource request frequency sequences from different tenants. Use a fifth-order B-spline kernel for time domain alignment. Interpolate to fill in missing segments of low-activity tenants to ensure that the input tensor is evenly distributed across all tenants.

[0064] S1-1-2. Introduce the disturbance index based on weighted information entropy as the third-dimensional data source of the tensor. The disturbance index based on weighted information entropy is calculated by weighted sliding entropy of the tenant's resource jitter variance in the past hour. The disturbance entropy uses a 0.3-second period window and performs bidirectional exponential smoothing, and finally is normalized to the range of [0,1].

[0065] Specifically, for low-activity tenants, their request sequences may have some missing time periods. In this case, the tenant time series will be aligned with the unified time axis using B-spline registration (fifth-order kernel), and the missing parts will be interpolated to ensure that all tensors are aligned in shape to 4×6×1. The third dimension of the perturbation density is obtained as follows:

[0066] Statistics on CPU, bandwidth, and GPU jitter variance of tenants in the past hour;

[0067] Construct the perturbation entropy index: Using these variances as weights, calculate the information entropy of the resource distribution within the sliding window with a period of 0.3s;

[0068] The final perturbation entropy is normalized to the interval [0,1] after double exponential smoothing and used as the third-dimensional data channel. The S2 further extracts the mutual exclusion feature of the compressed tenant tensor and includes the following steps:

[0069] S2-1. Use the principal components of tensor embedding to construct a resource competition graph with the structure G = (V, E, W)

[0070] V is the tenant set, E is the competitive edge set, and W is the edge weight. The edge weight is constructed by the cross-utilization function between resource types; the form is:

[0071]

[0072] wherein u i is the resource occupancy main vector of tenant i, is the square of the F norm of the tensor globally;

[0073] S2-2. A heterogeneous gated graph neural network is used to encode the constructed resource competition graph. In the resource competition graph, an independent edge type is introduced for each type of resource competition. The node update function integrates the time dimension gradient and the resource coupling factor. The edge update introduces a type penalty weight. During training, a residual aggregation method with attention is used to realize the dynamic feature fusion of multiple resource conflicts.

[0074] Said S2-2 further comprises the following steps:

[0075] S2-2-1. Introduce a resource coupling factor during the update process of the heterogeneous gated graph neural network. The resource coupling factor is expressed as the correlation coefficient between the current resource main vector and the moving average main vector of the past 24 hours, and perform softmax smoothing activation during forward propagation;

[0076] S2-2-2. Add the type penalty weight in the edge update, the value of which is the ratio of the average tenant occupancy rate of the current resource type to the resource surplus rate, and use it as an explicit regularization term in the loss function to suppress the formation of misleading coupling paths on scarce resources.

[0077] Specifically, the resource coupling factor is calculated as follows:

[0078]

[0079] in It is the moving average of the tenant's main vector in the past 24 hours. The correlation coefficient is calculated using Pearson and activated by softmax, which is used as the weighted input for node status update.

[0080] Type penalty weight in edge update:

[0081]

[0082] Add a regularization term to the training loss function:

[0083]

[0084] Used to suppress the propagation of spurious correlations on scarce resource paths.

[0085] The construction of the hierarchical scheduling candidate space in S3 further includes the following steps:

[0086] S3-1. Based on the conflict risk factor map, tenants are divided into different resource levels. Resource levels are unsupervisedly sorted by the cosine similarity of the principal components of the tenant and resource tensors. Asymmetric hash clustering is used, and tenants are mapped to the scheduling subspace using the locality-sensitive hashing + cluster center enforcement constraint method. Cross-tenants are eliminated by boundary distance projection;

[0087] S3-2. A multi-objective policy network training process is constructed in parallel for each scheduling layer. The policy network consists of a dual-branch structure, which trains the optimal throughput path and the value return path respectively. The action space uses an enhanced pruning mechanism to remove low-contribution actions, and finally combines them to obtain a set of candidate scheduling actions.

[0088] Specifically, in the construction of the strategy layer, the following are introduced:

[0089] Throughput maximization policy network: uses DeepQ-Network, with inputs including the tenant's current state tensor embedding and graph neural conflict score, and outputs the Q value of the resource allocation option in the action space;

[0090] Value-reward optimal strategy network: Adopts the Actor-Critic architecture. The state input is the same as above. The Actor outputs the probability of resource scheduling action, and the Critic evaluates the corresponding reward (such as load stability and SLA achievement rate).

[0091] The periodic switching controller checks the policy divergence every 10 scheduling cycles, and the indicator is:

[0092] Δ=KL(P actor , P DQN )

[0093] If Δ exceeds a threshold, the switching period is shortened to enhance exploration. The perturbation gradient change rate is used to dynamically adjust the controller weights to achieve an adaptive balance between exploration (throughput priority) and convergence (stability priority).

[0094] The S3-2 further comprises the following steps:

[0095] S3-2-1. The optimal throughput path search prioritizes the most concurrent tenants, initializes the policy queue by sorting tenant node-level load expansion rates, and employs a gradient-guided Monte Carlo tree search. Action combinations in high-density conflict areas in the resource risk graph are eliminated.

[0096] S3-2-2. A value guidance factor is introduced in the search for the optimal return path. The value guidance factor is evaluated based on the partial derivative of the resource allocation action and the average return surface of the historical stable interval, and a priority scheduling table based on the time-weighted lag factor is constructed.

[0097] Specifically, the optimal path search constructs a multi-tenant scheduling graph model, extracts the instantaneous CPU, memory, and I / O load and average load offset for each tenant node, calculates the expansion rate, and sorts them to form the initial policy queue. Monte Carlo tree search, combined with a gradient guidance factor, prioritizes the expansion of nodes with the largest gradient changes in the current state, avoiding local optima. During path evaluation, tenant scheduling history is integrated to construct a resource risk map, dynamically marking high-conflict areas and eliminating action combinations containing such areas from the search. Regarding the optimal return path, the system continuously monitors the profit curves after each scheduling action is executed, constructing a three-dimensional profit tensor. Partial derivatives of the current action are calculated based on the historical profit curve to form a real-time value guidance factor. This factor is weighted by a time lag window (window size T = 30 seconds, update period 5 seconds) to form a priority scheduling table. This allows for rapid identification of cost-effective action combinations when scheduling high-concurrency tasks, thereby improving overall system profitability.

[0098] The step of performing robustness screening on the scheduling action set in S4 further includes:

[0099] S4-1. Construct a disturbance response training set, including all burst load events, resource scheduling failures, and sampling points with sudden response latency exceeding 50ms in the past three months. Use a conditional generative adversarial network to generate disturbance-augmented samples, using the concatenation of the tenant's historical state matrix and the scheduling graph context.

[0100] S4-2. Input the candidate scheduling action sequence into the disturbance response discrimination module. The disturbance response discrimination module internally constructs a multi-channel variation gradient detector to detect the disturbance response gradient trend of the scheduling behavior under each type of resource dimension. Samples with significantly positive slope growth are eliminated, and scheduling actions with a disturbance response curve slope converging near 0 are retained.

[0101] Specifically, the system first summarizes the key disturbance events that occurred in the past three months and constructs a disturbance response training set, which includes failed scheduling records caused by insufficient resources or scheduling conflicts, and scheduling samples with response delays exceeding 50 milliseconds. These samples are enhanced using a conditional generative adversarial network (cGAN), with the input conditions being the tenant's historical resource usage trajectory matrix (a time series containing dimensions such as CPU, memory, and network) and the current scheduling graph context embedding vector. The trained generator is used to generate simulated scheduling scenarios under high-noise situations for robustness screening. Subsequently, the system inputs all scheduling action sequences to be screened into the disturbance response discrimination module, which constructs a multi-channel variation gradient detector inside the module. Each channel corresponds to a resource dimension (such as CPU, memory, bandwidth, etc.) to monitor the changing trend of the response curve of the scheduling action under small disturbance conditions. If the slope of the response curve of a certain scheduling action continues to grow rapidly (for example, within the disturbance range of 10%, the slope is greater than 0.8), it is judged as a high-risk sample and is eliminated. Only scheduling behaviors whose disturbance response converges near zero (-0.1 to +0.1) or remains stable are retained, significantly improving the robustness and stability of the final scheduling strategy.

[0102] The GAN perturbation generator for the S4-1 perturbation response training set includes the following steps:

[0103] S4-1-1. The conditional input vector is constructed by embedding the resource request frequency, resource history status window, and scheduling graph into three sets of concatenated representations. The concatenated vector dimension is [128, 256, 64] and is uniformly mapped to the generated input vector through a two-layer feedforward network.

[0104] S4-1-2. The discriminator uses a forward time residual network with 4 residual modules and a basic unit of Conv1D-ReLU-Dropout combination. The final output discrimination value is based on the weighted integral scoring function of the standard deviation of the time window fluctuation of the scheduling failure rate and the mean error of resource usage.

[0105] Specifically, in practical applications, the input vector of the GAN perturbation generator consists of three parts: one is the frequency sequence of tenant resource requests (such as the number of changes in CPU requests per second), with an embedding dimension of 128; the second is the resource history status window (including the status in the past 60 seconds), with an embedding dimension of 256; the third is the current scheduling graph context embedding (including tenant topological position, adjacent tenant load and other structural information), with an embedding dimension of 64. After splicing, the three groups of vectors are mapped into a unified input vector (dimension of 512) through a two-layer feedforward neural network and input into the GAN generator. In the discriminator design, a forward time residual structure is adopted, and the residual module adopts the Conv1D-ReLU-Dropout structural unit, where the convolution kernel size is 3 and the Dropout ratio is 0.2. After stacking through 4 layers, the output is the scheduling discriminant score. The scoring function comprehensively considers the standard deviation (indicating jitter) of the scheduling failure rate in multiple time windows and the mean deviation of resource utilization, with weights set to 0.6 and 0.4 respectively. This achieves high-precision judgment of the quality of generated samples, improves the authenticity and effectiveness of disturbance simulation, and thus effectively supports the robustness verification of subsequent scheduling strategies.

[0106] An AI-driven multi-tenant cloud computing resource optimization allocation system relates to any AI-driven multi-tenant cloud computing resource optimization allocation method in claims 1-8, and the AI-driven multi-tenant cloud computing resource optimization allocation system is designed based on the AI-driven multi-tenant cloud computing resource optimization allocation method.

[0107] Specifically, the basic principles, main features and advantages of the present invention are shown and described above; those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected; the scope of protection claimed in the present invention is defined by the attached claims and their equivalents.

Claims

1. An AI-driven multi-tenant cloud computing resource optimization allocation method, characterized in that: The AI-driven multi-tenant cloud computing resource optimization allocation method includes the following steps: S1. Build a dynamic profile of tenant resource usage. This is done by representing each tenant's resource request sequence, request-response latency sequence, and historical load jitter trajectory as a tensor. A three-dimensional residual attention network is used for representation compression, ultimately mapping each tenant into a three-dimensional dense tensor structure. This structure includes dimensions for resource type, time window, and perturbation density. S2. Extract mutual exclusion features from the compressed tenant tensor representation and use a gated graph neural encoder to topologically model resource competition conflicts between tenants. The graph neural network introduces multi-dimensional edge types. Node state updates are jointly influenced by historical conflict weights and resource isomorphism coefficients, and a conflict risk factor map is output. S3. Construct a hierarchical scheduling candidate space based on the conflict risk factor graph, use a distributed multi-objective deep policy network to learn a throughput maximization strategy and a value reward optimization strategy, respectively, and introduce a periodic switching controller between the throughput maximization strategy and the value reward optimization strategy. The periodic switching controller uses the cross-difference between the perturbation gradient and the reward stability as a period adjustment factor to achieve an exploration-convergence balance in resource scheduling. S4. Robustness screening is performed on the generated scheduling action set. An adaptive annealing mechanism is used to dynamically update the temperature factor based on the historical fluctuation window response curve. Scheduling action sequences that are sensitive to future load changes are eliminated, and a convergent scheduling strategy set is output. S5. Execute resource allocation under the final selected scheduling action and continuously collect micro-resource utilization traces during the allocation cycle, including CPU, memory, I / O, and GPU periodic occupancy rates as feedback signals to guide policy fine-tuning and updates during subsequent training.

2. The AI-driven multi-tenant cloud computing resource optimization allocation method according to claim 1 is characterized in that: The step of constructing a dynamic profile of tenant resource usage in S1 further includes the following steps: S1-1. Obtain the tenant's resource request frequency sequence, CPU utilization gradient sequence, and bandwidth time series density sequence as ternary time series input. Use a fixed 30-second window, extract overlapping data using a 5-second sliding window, and map them to a three-dimensional dense tensor structure. Each dimension of the three-dimensional dense tensor represents a resource item, including CPU, memory, bandwidth, GPU, time window, and load disturbance index. S1-2. The three-dimensional tensor structure is input into the three-dimensional residual attention network for embedding compression. The three-dimensional residual attention network adopts a three-layer convolution unit and attention fusion. The channel dimensions of each layer are 64, 128, and 64. The attention module introduces an anti-normalized scalar reinforcement mechanism with the maximum load perturbation point as the anchor, and embeds a cross-scale skip connection structure in the residual path to enhance the low-frequency perturbation retention rate.

3. The AI-driven multi-tenant cloud computing resource optimization allocation method according to claim 2 is characterized in that: Said S1-1 further comprises the following steps: S1-1-1. Perform nonlinear registration on the unified time axis for resource request frequency sequences from different tenants. Use a fifth-order B-spline kernel for time domain alignment. Interpolate to fill in missing segments of low-activity tenants to ensure that the input tensor is evenly distributed across all tenants. S1-1-2. Introduce the disturbance index based on weighted information entropy as the third-dimensional data source of the tensor. The disturbance index based on weighted information entropy is calculated by weighted sliding entropy of the tenant's resource jitter variance in the past hour. The disturbance entropy uses a 0.3-second period window and performs bidirectional exponential smoothing, and finally is normalized to the range of [0,1].

4. The AI-driven multi-tenant cloud computing resource optimization allocation method according to claim 1, characterized in that: The S2 further performs mutual exclusion feature extraction on the compressed tenant tensor, including the following steps: S2-1. Use the principal components of tensor embedding to construct a resource competition graph with the structure G = (V, E, W) V is the tenant set, E is the competitive edge set, and W is the edge weight. The edge weight is constructed by the cross-utilization function between resource types; the form is: wherein u i is the resource occupancy main vector of tenant i, is the square of the F norm of the tensor globally; S2-2. A heterogeneous gated graph neural network is used to encode the constructed resource competition graph. In the resource competition graph, an independent edge type is introduced for each type of resource competition. The node update function integrates the time dimension gradient and the resource coupling factor. The edge update introduces a type penalty weight. During training, a residual aggregation method with attention is used to realize the dynamic feature fusion of multiple resource conflicts.

5. The AI-driven multi-tenant cloud computing resource optimization allocation method according to claim 4 is characterized in that: Said S2-2 further comprises the following steps: S2-2-1. Introduce a resource coupling factor during the update process of the heterogeneous gated graph neural network. The resource coupling factor is expressed as the correlation coefficient between the current resource main vector and the moving average main vector of the past 24 hours, and perform softmax smoothing activation during forward propagation; S2-2-2. Add the type penalty weight in the edge update, the value of which is the ratio of the average tenant occupancy rate of the current resource type to the resource surplus rate, and use it as an explicit regularization term in the loss function to suppress the formation of misleading coupling paths on scarce resources.

6. The AI-driven multi-tenant cloud computing resource optimization allocation method according to claim 1, characterized in that: The construction of the hierarchical scheduling candidate space in S3 further includes the following steps: S3-1. Based on the conflict risk factor map, tenants are divided into different resource levels. Resource levels are unsupervisedly sorted by the cosine similarity of the principal components of the tenant and resource tensors. Asymmetric hash clustering is used, and tenants are mapped to the scheduling subspace using the locality-sensitive hashing + cluster center enforcement constraint method. Cross-tenants are eliminated by boundary distance projection; S3-2. A multi-objective policy network training process is constructed in parallel for each scheduling layer. The policy network consists of a dual-branch structure, which trains the optimal throughput path and the value return path respectively. The action space uses an enhanced pruning mechanism to remove low-contribution actions, and finally combines them to obtain a set of candidate scheduling actions.

7. The AI-driven multi-tenant cloud computing resource optimization allocation method according to claim 6, characterized in that: The S3-2 further comprises the following steps: S3-2-1. The optimal throughput path search prioritizes the most concurrent tenants, initializes the policy queue by sorting tenant node-level load expansion rates, and employs a gradient-guided Monte Carlo tree search. Action combinations in high-density conflict areas in the resource risk graph are eliminated. S3-2-2. A value guidance factor is introduced in the search for the optimal return path. The value guidance factor is evaluated based on the partial derivative of the resource allocation action and the average return surface of the historical stable interval, and a priority scheduling table based on the time-weighted lag factor is constructed.

8. The AI-driven multi-tenant cloud computing resource optimization allocation method according to claim 1, characterized in that: The step of performing robustness screening on the scheduling action set in S4 further includes: S4-1. Construct a disturbance response training set, including all burst load events, resource scheduling failures, and sampling points with sudden response latency exceeding 50ms in the past three months. Use a conditional generative adversarial network to generate disturbance-augmented samples, using the concatenation of the tenant's historical state matrix and the scheduling graph context. S4-2. Input the candidate scheduling action sequence into the disturbance response discrimination module. The disturbance response discrimination module internally constructs a multi-channel variation gradient detector to detect the disturbance response gradient trend of the scheduling behavior under each type of resource dimension. Samples with significantly positive slope growth are eliminated, and scheduling actions with a disturbance response curve slope converging near 0 are retained.

9. The AI-driven multi-tenant cloud computing resource optimization allocation method according to claim 8, characterized in that: The GAN perturbation generator for the S4-1 perturbation response training set includes the following steps: S4-1-1. The conditional input vector is constructed by embedding the resource request frequency, resource history status window, and scheduling graph into three sets of concatenated representations. The concatenated vector dimension is [128, 256, 64] and is uniformly mapped to the generated input vector through a two-layer feedforward network. S4-1-2. The discriminator uses a forward time residual network with 4 residual modules and a basic unit of Conv1D-ReLU-Dropout combination. The final output discrimination value is based on the weighted integral scoring function of the standard deviation of the time window fluctuation of the scheduling failure rate and the mean error of resource usage.

10. An AI-driven multi-tenant cloud computing resource optimization allocation system, relating to any AI-driven multi-tenant cloud computing resource optimization allocation method according to claims 1-8, characterized in that: The AI-driven multi-tenant cloud computing resource optimization allocation system is designed based on the AI-driven multi-tenant cloud computing resource optimization allocation method.

Images