AI-enabled storage resource dynamic allocation method and system

Abstract

The invention relates to the technical field of computer storage, and discloses an AI enabling storage resource dynamic allocation method and system, and the method comprises the steps: constructing a multi-level Markov decision model, and representing the storage resource management as a local and global state conversion and decision process; a multi-agent reinforcement learning framework is constructed and deployed in an interface layer and a scheduling layer to realize distributed intelligent decision making; a bidirectional information flow mechanism between the interface layer and the scheduling layer is realized, a hierarchical barrier is broken, and efficient transmission of information is realized; based on a multi-scale collaborative optimization algorithm, local autonomy and global consistency are fused, and a self-adaptive storage resource allocation strategy is dynamically generated. Through a multi-level autonomous cooperation mechanism, the technical bottlenecks of hierarchical splitting, rigid boundary, single optimization, lack of autonomous evolution ability and the like of a traditional storage system are broken through, and the system performance is remarkably improved, the adaptive ability is enhanced, the cooperation efficiency is optimized, and decision-making intelligence is achieved.

Description

Technical Field

[0001] This invention relates to the field of computer storage technology, and more specifically, to an AI-enabled method and system for dynamic allocation of storage resources. Background Technology

[0002] With the rapid development of cloud computing, big data, and artificial intelligence technologies, data centers are expanding in scale, and storage systems are becoming increasingly complex. As a core component of data center operations and maintenance, the efficiency and performance of storage resource management directly impact the overall system's service quality and operating costs.

[0003] Existing storage resource management technologies mainly include static pre-allocation, dynamic threshold adjustment, and rule-based automated management. However, these technologies suffer from the following problems: First, traditional storage systems typically treat interface design and resource scheduling as independent technical components, optimizing them separately. This leads to a lack of effective information sharing and collaboration mechanisms between the interface and scheduling layers. The interface layer struggles to perceive the global resource status, while the scheduling layer cannot fully understand the actual application requirements, ultimately resulting in a mismatch between resource allocation decisions and actual application needs. Second, existing systems often employ rigid hierarchical management structures, with each layer interacting through predefined fixed interfaces, lacking dynamic adaptability. When system load or application requirements change, this rigid structure struggles to respond quickly, resulting in inefficient resource allocation. Furthermore, existing technologies often focus only on local or global optimization, failing to achieve unified coordination across multiple optimization dimensions and levels, making it difficult to effectively resolve the contradiction between local and global optima. Finally, traditional storage resource management systems rely on fixed strategies and rules, making it difficult to self-adjust and optimize according to environmental changes. Systems lacking learning capabilities cannot extract knowledge from historical experience, leading to insufficient adaptability in dynamic environments.

[0004] In distributed environments, collaborative decision-making among multiple management entities presents an even more complex challenge. Existing technologies typically employ centralized control or simple negotiation mechanisms, which struggle to balance autonomous decision-making with collaborative consistency, particularly in large-scale systems. With the rise of edge computing and multi-cloud environments, storage resources are becoming increasingly dispersed and heterogeneous, further exacerbating the complexity of resource management. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an AI-enabled method and system for dynamic allocation of storage resources.

[0006] An AI-enabled method for dynamic allocation of storage resources includes the following steps: A multi-level Markov decision model is constructed, which includes a local state space, a global state space, a local action space, a global action space, a local state transition function, a global state transition function, a local reward function, a global reward function, and a discount factor. A multi-agent reinforcement learning framework is constructed, in which local agents are deployed at the interface layer and global agents are deployed at the scheduling layer; the local agents adopt a deep neural network structure and the global agents adopt a graph neural network structure. A bidirectional information flow mechanism is implemented, including bottom-up state aggregation and top-down policy guidance. Local state information is aggregated into global state through a hierarchical aggregation network, and global policy is decomposed into local constraints through a policy decomposition function. A multi-scale collaborative optimization algorithm is executed to construct a joint optimization objective function, dynamically adjust the weight coefficients of local and global rewards, balance local autonomy and global consistency, and generate an adaptive storage resource allocation strategy.

[0007] Preferably, the local agent uses the Proximal Policy Optimization algorithm for policy optimization, and the deep neural network structure includes an input layer, a feature extraction layer, an attention mechanism layer, and a policy output layer; the global agent uses the Multi-Agent Deep Deterministic Policy Gradient algorithm for policy optimization.

[0008] Preferably, the hierarchical aggregation network includes: A local encoder is used to encode the state of each interface layer in a reduced dimension. The hierarchical aggregation module employs a hierarchical aggregation strategy based on the system topology. The attention weight calculation module assigns dynamic weights to each local state based on a similarity function; The weighted aggregation module aggregates local states into global states based on attention weights and a transformation function.

[0009] Preferably, the attention weight calculation is performed through the following steps: Calculate the similarity score between the query vector and the local state vector; The similarity scores are indexed. The processed scores are normalized to ensure that the sum of all weights is 1; the attention given to information at different interface layers is adaptively adjusted according to the current state and objectives of the system.

[0010] Preferably, the strategy decomposition function includes: The context-aware module is used to analyze the meaning of the global policy in the specific context of the current global state and local state; the constraint generator is used to generate constraints such as resource usage limits, performance targets and cooperation instructions based on the global policy. Soft constraint converters are used to transform hard constraints into reward function correction terms, allowing local agents to maintain a certain degree of autonomous decision-making space while satisfying global constraints.

[0011] Preferably, the joint optimization objective function is achieved through the following formula: Maximize the expected cumulative discount reward of policy π, where the reward includes weighted local rewards and global rewards; Dynamically adjust the weight coefficients of local rewards and global rewards based on the global state and local gradients using a balancing function.

[0012] Preferably, the balance function includes: The local urgency assessment module is used to analyze the local state change trend and current load status, and calculate the local urgency index; The global coordination requirement assessment module is used to analyze the global status and resource distribution, and calculate the global coordination requirement index. The balance calculation module is used to comprehensively consider local urgency and global coordination needs to calculate a balance value, which determines whether the system tends to make local autonomous decisions or global coordination.

[0013] Preferably, the multi-level Markov decision model and multi-agent reinforcement learning framework are applied to one of the following scenarios: Large-scale distributed storage environment; Multi-tenant cloud storage platform; Edge-cloud collaborative storage system; Highly dynamic load environment; Heterogeneous storage system.

[0014] Preferably, the local agent is based on an experience replay learning method, which improves sample utilization efficiency by storing and reusing historical experience; the global agent is based on a hierarchical reinforcement learning structure, which reduces learning difficulty and improves generalization ability by decomposing complex decision problems into multiple levels of sub-problems.

[0015] An AI-enabled dynamic storage resource allocation system for the above method includes: The interface layer is equipped with local intelligent agents that use a deep neural network structure to handle application requests and optimize local resource allocation. The scheduling layer, equipped with a global intelligent agent and employing a graph neural network structure, is responsible for the macro-level scheduling and optimization of system resources. A two-way information flow channel is used to achieve bottom-up state aggregation and top-down policy guidance; The multi-scale collaborative optimization module is used to execute joint optimization algorithms and dynamically balance local autonomy and global consistency. The beneficial effects of this invention are as follows: The AI-enabled dynamic allocation method for storage resources proposed in this invention achieves significant technical results by addressing the problems of traditional storage systems, such as hierarchical fragmentation, rigid boundaries, single optimization perspective, and lack of autonomous evolution capabilities, through a multi-level autonomous collaborative mechanism. These results include: improved overall system performance (throughput increased by 40%-60%, response latency reduced by 35%-50%, and resource utilization increased by 25%-45%); improved dynamic environment adaptability (resource adjustment response time shortened by 85%, and performance stability improved by 70% under load fluctuations); reduced cross-layer communication latency by 60% and coordination overhead reduced by 50%; improved decision-making intelligence (resource allocation accuracy increased by 65%, and anomaly handling capability enhanced by 80%); and excellent scalability (performance degradation of no more than 10% and coordination overhead increase of only 2 times when the system size increases tenfold). Attached Figure Description

[0016] Figure 1 This is the present invention; Figure 2 It is a bar chart comparing system performance improvements; Figure 3 It is a line graph comparing system response latency under sudden load scenarios; Figure 4 This is a bar chart showing the degree of improvement (%) in inter-layer collaboration efficiency; Figure 5 It is a line graph comparing the long-term learning effects of the system; Figure 6 It is a bar chart comparing intelligent decision-making capabilities; Figure 7 It is a line graph comparing system scalability. Detailed Implementation

[0017] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0018] To aid in understanding the technical solution of this application, the terms used in this application will first be explained below: Multi-level autonomous collaboration: refers to a resource management approach in a distributed system where components at different levels maintain a certain degree of autonomy while achieving overall optimal resource management through information exchange and collaborative decision-making.

[0019] Storage resources: refers to storage-related computing resources in the system, such as storage space, I / O bandwidth, throughput, and IOPS, which are available for allocation and management.

[0020] Interface layer: This refers to the system layer that interacts directly with the application and is responsible for receiving and processing the application's storage resource requests.

[0021] Scheduling layer: This refers to the system layer responsible for global resource allocation decisions, coordinating resource requests from multiple interface layers to achieve reasonable allocation of global resources.

[0022] Multi-agent system: A system composed of multiple agents that can perceive the environment, make decisions and execute actions. Each agent has a certain degree of autonomy and achieves the overall goal of the system through cooperation.

[0023] Markov decision processes: a mathematical framework for modeling decision-making scenarios where the outcome is partly random and partly controlled by the decision-maker, applicable to sequential decision optimization problems.

[0024] It should be noted that the above terms are used for the purpose of understanding the technical solution of this application and should not be construed as limiting the scope of protection of this application.

[0025] Furthermore, all different types of parameters involved in this application require appropriate data preprocessing before calculation to ensure data comparability and the rationality of the calculations. Data preprocessing mainly includes the following aspects: Standardization / normalization of data with different dimensions: For parameters of different units or magnitudes (such as time, percentage, count value, etc.), standardization is performed before they are included in the same calculation to make them comparable.

[0026] Time series data smoothing: For time series data with fluctuations, methods such as sliding window averaging and exponentially weighted moving average are used to smooth the data and reduce the impact of noise.

[0027] Categorical feature encoding: Converting non-numerical data (such as application type, request type, etc.) into numerical representations. Common methods include one-hot encoding and label encoding.

[0028] Outlier handling: Detect and process data points that deviate significantly from the normal distribution to avoid extreme values ​​from adversely affecting the model.

[0029] The specific preprocessing methods will be further explained in the subsequent descriptions of the relevant calculations.

[0030] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, illustrates an AI-enabled dynamic allocation method for storage resources provided by this application. Those skilled in the art should understand that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application.

[0031] Step 1: Construct a multi-level Markov decision model According to an embodiment of this application, a hierarchical Markov decision model is first constructed to represent the storage resource management system as a state transition and decision-making process at two levels: local and global. The model is represented as follows: in: It represents the local state space, which includes state information such as interface layer resource usage and application request characteristics. The dimension is the number of locally observable parameters. Represents the global state space, which includes state information such as the overall resource distribution of the system and the load status of nodes. Its dimension is the number of system-level parameters. It represents the local action space, which includes executable operations such as interface parameter adjustment and local resource allocation, and defines the set of all actions that can be executed at the interface layer; It represents the global action space, which includes system-level operations such as cross-node resource scheduling and load balancing, and defines the set of all actions that the scheduling layer can execute; This represents a local state transition function, defining the probability distribution of state changes caused by local operations; This represents the global state transition function, defining the probability distribution of state changes caused by global operations; This represents a local reward function, used to evaluate the impact of local operations on performance. This represents the global reward function, which evaluates the impact of global operations on the overall system efficiency. This represents the discount factor, with a value range of [0,1]. It balances the weight of current decisions and future returns, and the closer it is to 1, the more emphasis is placed on long-term returns.

[0032] This model provides a mathematical framework for achieving multi-level collaborative optimization by uniformly formalizing the decision-making problem at both the local and global levels. The local state space describes the local operating state of the system, while the global state space describes the overall operating state of the system. The local action space defines the specific actions that the interface layer can take, while the global action space defines the macroscopic operations that the scheduling layer can execute. The state transition function describes the impact of actions on the system state. The reward function evaluates the effectiveness of actions, and the discount factor balances short-term and long-term gains.

[0033] Regarding model parameter processing, the system performs unified preprocessing for different types of state variables, specifically including: 1. Time-related parameters (such as response time, latency, etc.): These are obtained through logarithmic transformation. Reduce the magnitude difference, and then normalize to the [0,1] interval using Min-Max; 2. Percentage parameters (such as resource utilization rate, completion rate, etc.): directly divide by 100 to map to the [0,1] interval; 3. Counting parameters (such as queue length, number of requests, etc.): Use the Sigmoid function. Normalization is performed, where and These are scaling and offset parameters determined based on the distribution of historical data; 4. Classification parameters (such as request type, resource type, etc.): Converted into numerical vectors using one-hot encoding.

[0034] This model serves as the theoretical foundation for resource allocation decisions in the system. By providing a unified formal description of decision-making problems at both the local and global levels, it offers a mathematical framework for achieving multi-level collaborative optimization.

[0035] Step 1.1: Construction of Local and Global State Spaces In one embodiment, a local state space and a global state space are constructed based on system monitoring data: The local state space includes the following dimensions: application request queue length, request type distribution, interface resource utilization, response time, and other locally observable parameters.

[0036] The global state space includes the following dimensions: resource utilization of each node, overall system throughput, cross-node data transmission volume, global service quality indicators, and other system-level parameters.

[0037] Optionally, the local state space can also include features such as application type identification results, historical resource usage patterns, and periodic load predictions to further improve the comprehensiveness of the state representation.

[0038] Step 1.2: Defining the Action Space According to one embodiment of this application, the system's executable local action space and global action space are defined as follows: The local action space includes operations such as adjusting the interface buffer size, modifying the request processing priority, and configuring the local resource allocation ratio.

[0039] The global action space includes operations such as cross-node data migration, adjusting node resource quotas, and modifying global scheduling policy parameters.

[0040] In another embodiment, the action space can be further subdivided into discrete actions and continuous actions. For example, the interface type can be selected as a discrete action, while the buffer size can be adjusted to a continuous action, thus more accurately representing the set of operations that the system can perform by mixing action spaces.

[0041] In addition, constraints can be defined for actions to ensure that the system always operates safely. For example, upper limits can be set for resource allocation operations to prevent a single application from consuming too many resources and causing system instability.

[0042] Step 1.3: Design of Transfer Function and Reward Function In this embodiment, a state transition function and a reward function are designed based on historical operating data and expert knowledge: Local transition function: represents the probability distribution of the system transitioning to a new local state after performing a certain local action in the current local state, and describes the uncertainty of the local action.

[0043] Global transition function: Represents the probability distribution of the system transitioning to a new global state after performing a certain global action in the current global state, describing the system-level impact of the global action.

[0044] Local reward function: Calculates the immediate benefits of local operations based on metrics such as interface response time and resource utilization, and evaluates the effectiveness of local behaviors.

[0045] Global reward function: Calculates the system benefits of global operations based on metrics such as system throughput and load balancing, and evaluates the synergistic effect of global scheduling.

[0046] Optionally, the reward function can include multiple evaluation metrics, such as: Performance Bonus: Evaluate the impact of operations on system performance, such as reduced response time, increased throughput, etc. Efficiency rewards: assessing the degree of improvement in resource utilization; Balance incentive: Assess the degree of improvement in load balancing; Stability bonus: Assess the degree of reduction in system volatility.

[0047] These different types of rewards need to be properly standardized before comprehensive calculation to eliminate differences in dimensions and numerical ranges between different indicators. The standardization process is as follows: Performance metric standardization: For metrics such as response time and latency, first calculate the relative improvement rate. (The larger the value, the more significant the improvement), and then through the Tanh function. Map it to the range [-1, 1], where It is a coefficient that adjusts sensitivity; Efficiency indicator standardization: Calculating relative improvements in response to changes in resource utilization. ,in It is a preset optimal utilization target (usually 70%-80%), and then mapped to a standardized reward value through a similar method; Standardization of Balance Indicators: Based on the Jain Fairness Index Calculate the degree of improvement in load balancing, among which This represents the load of the i-th node, and the index value is in Within the range, the closer the value is to 1, the more balanced it is; Stability index standardization: This is quantified by calculating the degree of reduction in the coefficient of variation (standard deviation divided by mean) of key indicators and mapping it to a standard range.

[0048] These standardized reward values ​​can be comprehensively considered through a weighted summation, with the weights dynamically adjusted based on the current system state and optimization objectives. The weighting coefficients are satisfied to ensure the interpretability and consistency of the total reward value.

[0049] Step 2: Construction of a Multi-Agent Reinforcement Learning Framework According to one embodiment of this application, a multi-agent reinforcement learning framework is constructed based on the Markov decision model built in step 1, and deployed at the interface layer and scheduling layer to achieve distributed intelligent decision-making. This framework includes local agents and global agents, and optimizes the overall decision-making strategy through collaborative learning.

[0050] Step 2.1: Local Agent Architecture In this embodiment, a local intelligent agent is deployed at the interface layer to handle application requests and optimize local resource allocation: State input: Obtain the current interface state and request characteristics from the local state space as the basis for the agent's decision-making.

[0051] Policy network: It adopts a deep neural network structure to map the state to the action probability distribution, realizing the mapping from state to behavior.

[0052] Value network: Estimates the value function of the current state, assists in policy optimization, and provides long-term return evaluation.

[0053] Action output: Select the optimal action from the local action space and execute it to achieve intelligent adjustment of resource allocation.

[0054] The local agent uses the Proximal Policy Optimization (PPO) algorithm for policy optimization, with the objective function being: in This represents the probability ratio between the old and new strategies. It is the estimation of the advantage function. This is the trimming parameter (usually set to 0.1 or 0.2). These are the policy network parameters.

[0055] The objective function first calculates the ratio of the probabilities of the current policy and the old policy under the same state and action, reflecting the magnitude of the policy update. Then, it multiplies this ratio by an estimate of the dominance function, which represents the value gain of choosing a specific action relative to the average level. Next, it constrains this product by using a pre-defined upper or lower bound if the ratio exceeds it, preventing excessively large policy updates from causing training instability. Finally, it takes the expected value of the samples from all time steps to obtain the final optimization objective. This objective function, by limiting the policy update step size, achieves continuous policy optimization while ensuring training stability.

[0056] Specifically, according to embodiments of this application, the deep neural network structure of the local agent includes: 1. Input layer: Receives state vectors with dimensions consistent with the local state space, containing application request features and resource usage, providing raw input data for subsequent processing.

[0057] 2. Feature Extraction Layer: Composed of 3 fully connected layers, each containing 128 neurons, using the ReLU activation function, responsible for extracting key features from the original state data.

[0058] 3. Attention Mechanism Layer: Employs a multi-head self-attention mechanism to handle the correlation between input features, with particular attention to the mutual influence between application request patterns and resource states.

[0059] 4. Policy Output Layer: Contains two branches: Policy branch: Outputs the probability distribution in the local action space, using the softmax activation function.

[0060] Value branch: Output state value estimate, using a linear activation function.

[0061] Optionally, in some embodiments, the local agent can employ an experience-replay-based learning method to improve sample utilization efficiency by storing and reusing historical experience. Specific implementations of this experience-replay mechanism include: (1) Preprocessing of empirical data: First, the collected raw interaction data is preprocessed, including numerical standardization (mapping feature values ​​of different scales to the same range), missing value handling (filling by interpolation or mean) and outlier filtering (removing extreme samples that may affect the learning quality). (2) Priority Sampling: Assign a priority weight to each empirical sample. The priority calculation formula is as follows: ,in It is the temporal difference error (reflecting the degree of surprise or information content of the sample). It is a small positive number (ensuring that all samples have a non-zero probability of being sampled); (3) Experience storage structure: A segmented tree data structure is used to store experience and its corresponding priority value, supporting... Priority updates and sampling operations with varying complexity; (4) Importance Sampling: When sampling according to priority, importance weights are used. Revise the learning objectives, among which It is the size of the playback buffer. It is a sample The sampling probability, It is a parameter that gradually increases from 0 to 1.

[0062] In addition, a distributed training approach can be adopted, in which multiple interface layer agents collect experience in parallel and share model parameters, thereby accelerating the learning process.

[0063] In e-commerce platform storage system applications, local intelligent agents can sense sudden order processing requests and allocate more I / O resources in advance during peak periods by adjusting buffer size and data prefetching strategies to ensure rapid response of critical transaction data.

[0064] Step 2.2: Global Intelligent Agent Architecture According to one embodiment of this application, a global intelligent agent is deployed at the scheduling layer to be responsible for the macro-scheduling and optimization of system resources: State aggregation: Collect and aggregate state information from each node to form a global view.

[0065] Global Policy Network: Employs a graph neural network structure to process system topology information and generate a global scheduling policy.

[0066] Global Value Network: Evaluates the value of the global state and guides the optimization of global strategies.

[0067] Scheduling execution: Decompose global decisions into specific scheduling instructions and issue them to relevant nodes.

[0068] The global agent employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm for policy optimization, which can effectively coordinate the actions of multiple agents in partially observable environments.

[0069] According to a preferred embodiment of this application, the graph neural network structure of the global agent is specifically implemented as follows: 1. Graph Construction Layer: The storage system is represented as a graph structure G(V,E), where V is the set of nodes (storage nodes, computing nodes, etc.) and E is the set of edges (network connections, data dependencies, etc.).

[0070] 2. Node Feature Embedding Layer: The initial features of each node include the resource state of the node and the state information of the local agent.

[0071] The preprocessing and embedding transformation of node features are as follows: (1) Feature cleaning: First, the original node features are cleaned, including removing noisy data, handling missing values ​​(using temporal correlation interpolation method), and outlier detection (identifying and correcting outliers based on statistical distribution characteristics). (2) Feature normalization: Normalize the node attributes of different dimensions. For bounded features, use the Min-Max normalization method. The values ​​are mapped to the [0,1] interval, and Z-score normalization is used for unbounded features. ; (3) Feature fusion: The node features from different sources (such as resource status, historical performance data, topology information, etc.) are fused and a unified representation is generated through feature connection and nonlinear transformation; (4) Dimension Transformation: A fully connected neural network is used to map the original features of different dimensions to embedding vectors of fixed dimensions, represented as follows: ,in It is a node The original feature vector, and It consists of a learnable weight matrix and a bias vector. It is a non-linear activation function (such as ReLU).

[0072] 3. Graph Convolutional Layer: A 3-layer Graph Convolutional Network (GCN) is used, with each layer containing 64 channels and a message passing radius of 2 hops, to capture the topological relationships and state propagation between nodes.

[0073] 4. Graph Pooling Layer: An attention-based graph pooling method is used to aggregate node information and generate a graph-level representation.

[0074] 5. Global Decision Layer: Connects two fully connected layers, outputting global action and state value estimates respectively.

[0075] It should be noted that in other embodiments, the global agent may also adopt a hierarchical reinforcement learning structure, which reduces the learning difficulty and improves the generalization ability by decomposing the complex decision problem into multiple levels of sub-problems.

[0076] In multi-tenant cloud storage platform applications, the global agent can coordinate cross-node resource allocation based on the service level agreements (SLAs) and resource usage of different tenants. For example, when an increase in I / O latency is detected for a high-priority tenant, the global agent will comprehensively consider the overall system load and network conditions, and ensure the performance goals of that tenant through data migration or resource reallocation, while minimizing the impact on other tenants.

[0077] Step 3: Implementation of the bidirectional information flow mechanism According to one embodiment of this application, a bidirectional information flow mechanism is constructed between the interface layer and the scheduling layer to break down hierarchical barriers and achieve efficient information transmission and collaborative optimization of decision-making.

[0078] Step 3.1: Bottom-up state aggregation In this embodiment, the state information aggregation process from the interface layer to the scheduling layer is implemented: Local feature extraction: The local agent extracts key features from the raw monitoring data, including application request patterns, interface performance indicators, resource utilization status, etc.

[0079] Information compression and encoding: The extracted feature information is compressed and encoded to reduce communication overhead. Dimensionality reduction techniques such as Principal Component Analysis (PCA) or autoencoders are used. The specific data preprocessing and dimensionality reduction process includes the following steps: Data standardization: First, the original feature data is standardized, transforming each feature dimension into a standard normal distribution with zero mean and unit variance. The calculation formula is as follows: ,in It is a mean vector. It is the standard deviation vector; Dimensional importance assessment: Principal component analysis is used to calculate the eigenvalues ​​and eigenvectors of the covariance matrix. The magnitude of the eigenvalues ​​reflects the data variance in the direction of the corresponding eigenvector. The larger the variance, the greater the information content. Dimension selection: Choose the number of dimensions to retain based on the cumulative variance contribution rate (usually retaining 95% of the information). Specifically, choose the smallest one. Make Established, among which It is the first 1 eigenvalue, It is the original dimension; Projection transformation: By calculating the original data and the selected... The inner product of eigenvectors projects high-dimensional data into a low-dimensional space, thus achieving data compression.

[0080] Layered aggregation and transmission: Local information is transmitted layer by layer to the global scheduling layer through multi-level aggregation to form a global view of the system.

[0081] The state aggregation process can be represented as: in Represents the global state. Indicates the first Local state of an interface layer It is an aggregation function used to integrate multiple local state information into a global state representation. This indicates the total number of interface layers. This function generates a global state description that represents the overall system status through a series of information extraction, filtering, and fusion operations. This process ensures that local information scattered across different interface layers can be effectively transmitted to the global scheduling layer, providing the necessary information foundation for global decision-making.

[0082] According to a preferred embodiment of this application, the aggregation function is specifically implemented as a hierarchical aggregation network, including the following components: 1. Local Encoder: Performs dimensionality reduction encoding on the state of each interface layer to generate fixed-dimensional feature vectors, reducing data redundancy and extracting key information.

[0083] 2. Hierarchical Aggregation: Based on the system topology, a hierarchical aggregation strategy is adopted. First, aggregation is performed between interface layers within the same physical node; Then, aggregation is performed between nodes in adjacent regions; Finally, aggregation is performed globally. 3. Attention Weight Calculation: Dynamic weights are assigned to each local state. The calculation method is as follows: First, the relevance of each local state to the current global focus is evaluated. The higher the relevance, the more important the local state is to the current system goal. Then, these relevance scores are exponentialized to enhance the difference and ensure that all weights are positive. Finally, normalization is performed so that the sum of all weights is 1, thus obtaining the importance weight of each local state. This dynamic weight allocation mechanism can adaptively adjust the attention to information at different interface layers according to the current system state and goal.

[0084] 4. Weighted Aggregation: Weighted aggregation is performed based on the calculated attention weights. The calculation method is as follows: First, each local state is transformed to the global state representation space through a specific mapping function to ensure information compatibility; then, each transformed local state is multiplied by its corresponding attention weight to obtain a weighted local state representation; finally, all weighted local state representations are summed to obtain a comprehensive global state representation. Through this weighted aggregation method, the system can highlight important information and weaken secondary information, forming a more accurate global system state representation.

[0085] Similarity function The specific implementation involves calculating the correlation between the query vector and the local state. The calculation method is as follows: First, the query vector is processed separately... and local state vector Through different transformation matrices and A linear transformation is performed, which maps the two types of vectors to the same feature space. Then, the dot product of the transformed query vector and the transformed local state vector is calculated to obtain their original similarity score. Finally, this score is divided by the feature dimension. The square root of the similarity is normalized, which prevents the gradient vanishing problem in high-dimensional spaces. Through this similarity calculation method, the system can accurately assess the relevance between different local states and the current global focus, thus influencing the attention weights. The allocation is based on the similarity function. The specific mathematical expression is: in It is the transformation matrix of the query vector. It is the transformation matrix of the state vector. It is the dimension of the query vector. It is the dimension of the state vector. It is the transformed feature dimension.

[0086] Transformation function The specific implementation involves mapping the local state from the local feature space to the global feature space. The calculation method is as follows: first, the local state vector... With a specially designed transformation matrix Matrix multiplication is performed, and the dimension of the transformation matrix is ​​designed to ensure that the dimension of the output vector is consistent with the global state space; then a bias vector is added to the multiplication result. The mapping results are then further adjusted. Through this linear transformation operation, the system can convert local state information of different forms and dimensions into a unified global state representation format, ensuring that information from different interface layers can be effectively represented and aggregated in the global state space, providing standardized input for subsequent global decision-making. Transformation function The mathematical expression is: in It is a transformation matrix that transforms the dimension into a matrix of 1 / 2. The local state vector is mapped to a dimension of The global state space, It is the bias vector, used to adjust the position of the mapping result.

[0087] Optionally, in some embodiments, the aggregation process can be carried out asynchronously, with different nodes reporting status information at different frequencies. The system dynamically adjusts the aggregation cycle according to the importance and timeliness of the information, ensuring the timely transmission of key information while reducing communication overhead.

[0088] In edge-cloud collaborative storage system scenarios, this aggregation mechanism can effectively handle the heterogeneity and unstable connections of edge nodes, adaptively adjust the information weights of different nodes, ensure that the status of key nodes is fully considered, and reduce communication overhead and latency.

[0089] Step 3.2: Top-down strategy guidance According to one embodiment of this application, a policy guidance process from the scheduling layer to the interface layer is implemented: Global constraint generation: The global agent generates a series of constraints and guidance signals based on the overall system state and optimization objectives.

[0090] Personalized decomposition: The global strategy is decomposed into specific guidance information for each node, taking into account node characteristics and local environment differences.

[0091] Soft constraint transformation: Converting guidance information into soft constraints or reward function modifications for local agents, guiding rather than forcing local decisions.

[0092] The policy guidance process can be described as follows: For each interface layer, the system generates specific constraints applicable to that interface layer based on the global policy, global state, and the local state of that interface layer. This process is implemented through a dedicated policy decomposition function, which transforms the global-level decision into specific guidance information for each interface layer, ensuring that the local decisions of the interface layer maintain a certain degree of autonomy while remaining consistent with the global policy, thereby achieving multi-level collaborative optimization.

[0093] According to a preferred embodiment of this application, the implementation of the strategy decomposition function includes the following components: 1. Context-aware module: Analyzes the meaning of global policies in the specific context of the current global and local states, and understands the specific impact of global decisions on specific interface layers.

[0094] 2. Constraint Generator: Generates constraints applicable to specific interface layers based on the global strategy, including: Resource usage limit: Limits the maximum amount of resources that the interface layer can use; Performance targets: Specify the performance metrics that the interface layer should achieve; Collaboration instructions: Guidelines for collaborative actions with other interface layers.

[0095] 3. Soft Constraint Converter: This converts hard constraints into soft constraints, i.e., a modification of the reward function. The calculation method is as follows: First, retain the original local reward function's evaluation values ​​for local states and actions, reflecting the direct impact of actions on local performance. Then, calculate the degree to which actions conform to global constraints, controlling the strength of the constraint's influence through a balancing parameter. Finally, add the two parts to obtain the modified reward function. This soft constraint mechanism allows local agents to maintain a certain degree of autonomous decision-making space while satisfying global constraints, achieving guided rather than forced policy coordination.

[0096] Constraint compliance evaluation function The specific implementation is to evaluate local actions. Global constraints The degree of satisfaction is calculated as follows: first, the overall constraint is decomposed into... Specific constraints Examples of constraints include resource usage limits, performance targets, and collaboration instructions; then, for each specific constraint, the degree of compliance of local actions is calculated. Next, different weighting coefficients are assigned based on the importance of each constraint. Finally, all weighted compliance scores are summed to obtain the comprehensive constraint compliance assessment result. (Constraint compliance assessment function) The mathematical expression is: in It is aimed at the first Evaluation function for each specific constraint. These are the corresponding weighting coefficients. Ensure weight normalization. Evaluate the function for different types of constraints. The implementation method is as follows: Regarding the upper limit constraint on resource usage ,in Indicates action The resulting resource usage It is the upper limit of resource usage. It is the penalty coefficient.

[0097] For performance target constraints ,in It is an action The expected performance values ​​produced This is the target performance value.

[0098] Constraints on collaborative instructions ,in It is the angle between the action vector and the instruction vector. The closer the cosine value is to 1, the more consistent the directions are.

[0099] This multi-dimensional constraint evaluation mechanism enables the system to comprehensively consider the impact of local actions on global coordination, achieving more refined strategy guidance.

[0100] It should be noted that in other embodiments, policy guidance can also adopt a hierarchical reinforcement learning method, where the global agent does not directly generate specific constraints, but generates sub-goals, and the local agents autonomously determine the specific methods to achieve the sub-goals, thereby enhancing the system's flexibility.

[0101] Alternatively, policy guidance can be combined with predictive models to issue guidance information in advance based on predictions of the system's future state, enabling local agents to adjust early and avoid potential problems.

[0102] In heterogeneous storage system applications, the global agent can issue targeted policy guidance to each interface layer based on data access patterns and the characteristics of different storage media. For example, for applications that handle sequential read and write of large files, it can guide them to prioritize the use of HDD resources; while for transaction database applications that require random access, it can guide them to use SSD or memory storage resources, achieving the best match between resource types and application requirements.

[0103] Step 4: Implementation of Multi-Scale Cooperative Optimization Algorithm According to one embodiment of this application, a multi-scale collaborative optimization algorithm is implemented, which integrates local autonomy and global consistency to achieve dynamic optimization allocation of storage resources.

[0104] Step 4.1: Construction of the Joint Optimization Objective Function In this embodiment, a joint optimization objective function that considers both local performance and global efficiency is constructed: in This represents a local reward function. This represents the global reward function. and These are dynamically adjusted weighting coefficients. It is a discount factor. and They are time steps state and action, This represents the global expectation operator.

[0105] The function first defines the immediate reward of the system at each time step, which consists of two parts: a local reward reflecting local performance and a global reward reflecting global efficiency. Then, it assigns dynamically adjusted weights to these two parts of the reward to balance local optimization and global synergy. Next, it introduces a discount factor so that the system considers long-term gains while optimizing current decisions. Finally, it calculates the expected value of the sum of discounted rewards across all time steps, which serves as the overall optimization objective of the system. The weights are dynamically adjusted based on the current system state, achieved through a balancing function and a sigmoid mapping, ensuring that the system can adaptively balance local response and global synergy in different scenarios.

[0106] Where the equilibrium function Responsible for determining the global state and local gradient The evaluation of the current system should focus on whether local autonomy or global coordination is important, and the sigmoid function... The evaluation results are then converted into weight values ​​between 0 and 1 to ensure that the weights are allocated within a reasonable range.

[0107] According to a preferred embodiment of this application, the specific implementation of the balance function includes the following components: 1. Local urgency assessment: Analyzing the local state change trend and current load situation, and calculating the local urgency index. The method is as follows: First, monitor Calculate the rate of change of key performance indicators such as response time, queue length, and resource utilization over time. The rate of change reflects the dynamic characteristics of the system state; then, the rate of change of each indicator is normalized by dividing the rate of change of that indicator by its historical maximum rate of change to obtain the relative change intensity; finally, different weight coefficients are assigned to each indicator according to its importance. Finally, the weighted relative change intensities are summed to obtain the comprehensive local urgency index. The complete mathematical expression is: in It is the first Performance metrics at time intervals Rate of change within, This indicates that the indicator is in time. The value, This is the historical maximum rate of change for this indicator (continuously updated using a sliding window method). It is an importance weighting coefficient that satisfies The selection of performance metrics depends on the actual system characteristics and typically includes: response time. (milliseconds), queue length (Number of requests), CPU utilization (Percentage) Memory Usage (Percentage), I / O latency (milliseconds), etc. The larger the index value, the more drastic the changes in the local system state, and the more room for autonomous adjustment is required.

[0108] Global Coordination Needs Assessment: Analyze the global status and resource distribution, and calculate the global coordination needs index. The method is as follows: First, calculate the resource utilization rate of each node in the system. variance This variance reflects the degree of unevenness in resource distribution; a larger variance indicates a more uneven resource distribution, requiring stronger global coordination. Then, the interdependence between components in the system is assessed. A higher degree of interdependence indicates a greater impact between components, requiring more careful global coordination; next, the overall health index of the system is calculated. The value is then subtracted from 1 to represent the level of potential risk in the system; higher risks require stronger global intervention. Finally, the evaluation results of these three aspects are weighted according to preset importance. , and Weighted summation yields a comprehensive global coordination demand index. Global Coordination Demand Index The calculation formula is: The calculation and preprocessing methods for each indicator are as follows: - Resource distribution imbalance First, the resource utilization rate of each node is standardized to the [0,1] interval. Then, the standard deviation is calculated and transformed using a nonlinear method. Map it to the range [0,1]. It is a parameter that adjusts sensitivity; - interdependence Based on information entropy and mutual information theory, the mutual information matrix between components in the system is calculated, and its normalized spectral norm is taken as the overall mutual dependence metric, with a numerical range of [0,1]; - Health index Taking into account system stability, performance compliance rate, and resource surplus, the geometric mean was used. The calculations are performed, where each sub-indicator has been standardized to the [0,1] interval.

[0109] The higher the index, the more the system needs to strengthen global coordination and reduce the weight of local autonomous decision-making.

[0110] Balancing calculation: Taking into account both local urgency and global coordination needs, a balance value is calculated. The method is as follows: First, the local urgency index is... Multiply by its importance coefficient This indicates the system's demand for local autonomy; then, the global coordination demand index is... Multiply by its importance coefficient And subtract it from the former to reflect the constraint effect of global coordination on local autonomy; then add a weighted... Reference bias term This parameter is adjusted based on the system's preset strategy preference (leaning towards local autonomy or global coordination); the final result is a comprehensive balance value. A higher value indicates that the system should favor local autonomous decision-making, while a lower value indicates that it should favor global coordination. Balance value The calculation formula is: because and Since these two index values ​​originate from different calculation processes and have different numerical ranges, the system standardizes them before applying the above formula to ensure comparability in calculations. in and It is the historical average and standard deviation of the local urgency index. and These are the historical average and standard deviation of the global coordination demand index. Through this standardization process, both index values ​​are converted into a standard normal distribution with a mean of 0 and a standard deviation of 1, making it possible to compare and calculate them on the same scale.

[0111] Through this balancing mechanism, the system can dynamically adjust the relationship between local autonomy and global consistency based on real-time status, achieving adaptive optimization through multi-level collaboration.

[0112] Optionally, in some embodiments, the balancing function can further consider the historical state of the system and the changing trends of performance indicators, predict possible changes in future resource requirements, and achieve forward-looking weight adjustment.

[0113] In high-dynamic load environments, when a sudden increase in the load of a node is detected and the local urgency assessment value rises, the system will temporarily increase the local optimization weight, allowing the node to prioritize its own needs; while when the system load is stable, the global coordination weight is increased, focusing on overall resource efficiency optimization.

[0114] Step 4.2: Implementation of the hierarchical learning algorithm According to one embodiment of this application, a hierarchical reinforcement learning algorithm that takes into account inter-layer interactions is implemented: Parallel learning: Local agents and global agents learn their respective optimal strategies in parallel.

[0115] Strategy Coordination: The two-way information flow mechanism implemented in step 3 enables the coordination and consistency of strategies at different levels.

[0116] Experience sharing: Through the experience pool sharing mechanism, experience transfer and knowledge reuse are allowed among different intelligent agents.

[0117] Alternatively, in another embodiment, the hierarchical learning process can further employ a course-based learning approach, starting with simple tasks and gradually increasing the difficulty to accelerate convergence and improve final performance. Furthermore, knowledge distillation techniques can be used to transfer knowledge learned by complex models to lightweight models, reducing deployment overhead.

[0118] Step 4.3: Generate Adaptive Resource Allocation Strategy According to embodiments of this application, an adaptive storage resource allocation strategy is generated based on a multi-scale collaborative optimization algorithm: Real-time response: The interface layer agent adjusts interface parameters and local resource allocation in real time based on the characteristics of application requests and local states.

[0119] Global coordination: The scheduling layer agent optimizes cross-node resource allocation and task scheduling based on the overall system state.

[0120] Strategy balancing: Through a dynamic weight adjustment mechanism, local response speed and global resource efficiency are adaptively balanced in different scenarios.

[0121] It should be noted that in some embodiments, the adaptive resource allocation strategy can be combined with business insights and optimized based on business processes and user behavior patterns. For example, in an e-commerce platform, the system can detect product promotion activities and prepare and schedule resources in advance.

[0122] The present invention has the following advantages: System performance improvement: Through multi-level collaborative optimization, the overall system throughput is increased by 40%-60%, the average response latency is reduced by 35%-50%, and the resource utilization rate is increased by 25%-45%, significantly improving the overall performance of the storage system.

[0123] Figure 2 The paper demonstrates a percentage improvement in three key performance indicators between the AI-enabled dynamic allocation method and the traditional static allocation method and the simple adaptive method, clearly showcasing the significant advantages of the method of this invention in terms of improved system throughput, reduced response latency, and increased resource utilization.

[0124] Enhanced Adaptability: The system can quickly adapt to load changes and environmental fluctuations. In sudden load scenarios, the resource adjustment response time is reduced by 85%, and the performance stability under load fluctuation conditions is improved by 70%, which greatly enhances the system's resilience and adaptability.

[0125] Figure 3 The system response latency changes of three methods were simulated under sudden load scenarios such as e-commerce promotional activities, and the trend of load level changes was also displayed. The charts show that the method of this invention can maintain a low and stable response latency under the condition of a surge in load, demonstrating the system's strong adaptability to sudden loads.

[0126] Collaborative efficiency optimization: Breaking the rigid boundaries between layers enables information sharing and collaborative decision-making between the interface layer and the scheduling layer, reducing cross-layer communication latency by 60%, reducing coordination overhead by 50%, and achieving more efficient multi-layer collaboration.

[0127] Figure 4 The percentage improvement of the three methods in three indicators—reduced cross-layer communication latency, reduced coordination overhead, and improved inter-layer collaboration efficiency—was compared, highlighting the advantages of the method of this invention in breaking down layer barriers and achieving efficient collaboration between the interface layer and the scheduling layer.

[0128] Intelligent decision-making: Through a multi-agent reinforcement learning framework, the system can learn from historical experience and continuously optimize decision-making strategies, improving the accuracy of resource allocation decisions by 65% ​​and enhancing the ability to handle abnormal situations by 80%, thus realizing intelligent storage resource management.

[0129] Figure 5 The study compares the learning capabilities of the three methods over a long period of time, using relative throughput as an indicator. It shows that the AI-enabled dynamic allocation method can continuously optimize system performance over time, demonstrating the system's autonomous evolution capability and long-term value.

[0130] Figure 6 The performance of the three methods was compared across three key intelligent decision-making metrics: accuracy in resource allocation decisions, ability to handle anomalies, and prediction accuracy. The results show that the method presented in this invention, achieved through a multi-agent reinforcement learning framework, demonstrates a significantly higher level of decision-making intelligence than other methods, particularly in anomaly handling and prediction capabilities.

[0131] Enhanced scalability: This method is applicable to various scenarios from small systems to large-scale distributed environments. As the system size increases, the performance degradation does not exceed 10%, and when the number of nodes increases 10 times, the coordination overhead only increases 2 times, demonstrating excellent scalability.

[0132] Figure 7 This paper demonstrates the scalability of three methods in terms of performance maintenance and coordination overhead as the number of system nodes increases. The proposed method exhibits excellent scalability, with performance degradation controlled within 10% as the number of nodes increases, while coordination overhead grows slowly. This proves that the method is applicable to various scenarios, from small systems to large-scale distributed environments.

[0133] The AI-enabled dynamic allocation method for storage resources proposed in this invention effectively solves the problems of hierarchical fragmentation, rigid boundaries, single optimization, lack of autonomous evolution and difficulty in decision-making collaboration in existing technologies, and provides an innovative solution for storage resource management in complex and ever-changing environments.

[0134] Application scenario description: According to embodiments of this application, the AI-enabled dynamic allocation method for storage resources provided is applicable to the following scenarios: Large-scale distributed storage environment: In distributed storage systems that span multiple regions and data centers, the number of storage nodes is huge and widely distributed, and the resource status and network environment are complex and changeable. It is necessary to achieve efficient coordination of global resources while ensuring local response speed.

[0135] Multi-tenant cloud storage platform: In a multi-tenant cloud storage platform, different tenants’ applications have different resource requirements and service quality requirements. The system needs to maximize the overall resource utilization while meeting the service level agreements (SLAs) of each tenant.

[0136] Edge-cloud collaborative storage system: In a hybrid architecture that combines edge computing and cloud computing, storage resources are distributed across edge nodes and cloud centers. It is necessary to dynamically adjust resource allocation strategies based on data access patterns and computing needs to balance the advantages of edge response speed and cloud storage capacity.

[0137] Highly dynamic load environments: In application scenarios where load patterns change rapidly, such as promotional activities on e-commerce platforms, trending events on social media, and sudden tasks in scientific computing, the system needs to quickly detect load changes and make corresponding resource adjustments.

[0138] Heterogeneous storage systems: In heterogeneous storage systems that include multiple storage media and technologies, such as hybrid systems that use HDD, SSD, NVMe, and memory-level storage simultaneously, different types of storage resources need to be intelligently allocated according to data characteristics and access patterns.

[0139] In these application scenarios, the technical solution provided in this application can effectively cope with complex and ever-changing environmental challenges, realize intelligent dynamic allocation of storage resources through a multi-level autonomous and collaborative mechanism, and significantly improve system performance and resource utilization efficiency.

[0140] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An AI-enabled method for dynamic allocation of storage resources, characterized in that, Includes the following steps: A multi-level Markov decision model is constructed, which includes a local state space, a global state space, a local action space, a global action space, a local state transition function, a global state transition function, a local reward function, a global reward function, and a discount factor. A multi-agent reinforcement learning framework is constructed, in which local agents are deployed at the interface layer and global agents are deployed at the scheduling layer; the local agents adopt a deep neural network structure and the global agents adopt a graph neural network structure. A bidirectional information flow mechanism is implemented, including bottom-up state aggregation and top-down policy guidance. Local state information is aggregated into global state through a hierarchical aggregation network, and global policy is decomposed into local constraints through a policy decomposition function. A multi-scale collaborative optimization algorithm is executed to construct a joint optimization objective function, dynamically adjust the weight coefficients of local and global rewards, balance local autonomy and global consistency, and generate an adaptive storage resource allocation strategy.

2. The method according to claim 1, characterized in that, The local agent uses the Proximal Policy Optimization algorithm for policy optimization, and the deep neural network structure includes an input layer, a feature extraction layer, an attention mechanism layer, and a policy output layer; the global agent uses the Multi-Agent Deep Deterministic Policy Gradient algorithm for policy optimization.

3. The method according to claim 1, characterized in that, The hierarchical aggregation network includes: A local encoder is used to encode the state of each interface layer in a reduced dimension. The hierarchical aggregation module employs a hierarchical aggregation strategy based on the system topology. The attention weight calculation module assigns dynamic weights to each local state based on a similarity function; The weighted aggregation module aggregates local states into global states based on attention weights and a transformation function.

4. The method according to claim 3, characterized in that, The attention weight calculation is achieved through the following steps: Calculate the similarity score between the query vector and the local state vector; The similarity scores are indexed. The processed scores are normalized to ensure that the sum of all weights is 1; Based on the current state and objectives of the system, the focus on information from different interface layers is adaptively adjusted.

5. The method according to claim 1, characterized in that, The strategy decomposition function includes: The context-aware module is used to analyze the meaning of the global policy in the specific context of the current global state and local state; the constraint generator is used to generate constraints such as resource usage limits, performance targets and cooperation instructions based on the global policy. Soft constraint converters are used to transform hard constraints into reward function correction terms, allowing local agents to maintain a certain degree of autonomous decision-making space while satisfying global constraints.

6. The method according to claim 1, characterized in that, The joint optimization objective function is achieved through the following formula: Maximize the expected cumulative discount reward of policy π, where the reward includes weighted local and global rewards; and dynamically adjust the weight coefficients of local and global rewards based on the global state and local gradients using a balancing function.

7. The method according to claim 6, characterized in that, The balance function includes: The local urgency assessment module is used to analyze the local state change trend and current load status, and calculate the local urgency index; The global coordination requirement assessment module is used to analyze the global status and resource distribution, and calculate the global coordination requirement index. The balance calculation module is used to comprehensively consider local urgency and global coordination needs to calculate a balance value, which determines whether the system tends to make local autonomous decisions or global coordination.

8. The method according to claim 1, characterized in that, The multi-level Markov decision model and multi-agent reinforcement learning framework are applied to one of the following scenarios: Large-scale distributed storage environment; Multi-tenant cloud storage platform; Edge-cloud collaborative storage system; Highly dynamic load environment; Heterogeneous storage system.

9. The method according to claim 1, characterized in that, The local agent is based on an experience replay learning method, which improves sample utilization efficiency by storing and reusing historical experience; the global agent is based on a hierarchical reinforcement learning structure, which reduces learning difficulty and improves generalization ability by decomposing complex decision problems into multiple levels of sub-problems.

10. An AI-enabled dynamic storage resource allocation system, used to execute the method according to any one of claims 1-9, characterized in that, include: The interface layer is equipped with local intelligent agents that use a deep neural network structure to handle application requests and optimize local resource allocation. The scheduling layer, equipped with a global intelligent agent and employing a graph neural network structure, is responsible for the macro-level scheduling and optimization of system resources. A two-way information flow channel is used to achieve bottom-up state aggregation and top-down policy guidance; The multi-scale collaborative optimization module is used to execute joint optimization algorithms and dynamically balance local autonomy and global consistency.

Images