A hybrid storage method and system for server-agnostic computing

Abstract

The application discloses a hybrid storage method and system for server non-aware computing, and belongs to the technical field of computer storage. The method is as follows: key access features are extracted based on non-aware computing data access requests; access logs corresponding to the key access features are extracted; access burst intensity scores are obtained based on the access logs and the key access features; a set of heat survival probabilities is obtained based on the access logs and the key access features; associated prediction results are obtained based on the access logs, the key access features and a dynamic graph model; access mode prediction results are formed according to the access burst intensity scores, the set of heat survival probabilities and the associated prediction results; data management decisions are output according to current system state information, the access mode prediction results and a preset reinforcement learning decision engine; hybrid storage operations are performed based on the data management decisions and a preset heterogeneous storage structure, dynamic optimal distribution of data among heterogeneous storage layers is realized, and access delay and storage cost are effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of power systems, and more particularly to a hybrid storage method and system for server-insensitive computing. Background Technology

[0002] Serverless computing, as an emerging cloud computing execution model, enables elastic scaling, fine-grained billing, and simplified operation and maintenance, significantly improving computing flexibility and cost-effectiveness. However, this architecture faces significant challenges in intermediate data management. Function instances typically run in ephemeral and stateless containers or micro-virtual machines (such as Firecracker and gVisor), and their local storage is reclaimed when the instance ends. This results in intermediate data sharing between functions heavily relying on external storage systems, such as remote object storage (e.g., Amazon S3, Azure BlobStorage, Google Cloud Storage) or distributed file systems. While this separation of compute and storage brings elasticity and flexibility, it also introduces significant data access overhead.

[0003] In existing technologies, intermediate data storage in server-insensitive computing environments suffers from the following drawbacks: high access latency and network overhead. Its reliance on remote storage for data exchange leads to frequent network transmissions, significantly increasing data access latency, particularly for data-intensive or latency-sensitive applications (such as real-time analytics, machine learning inference, and online transaction processing). These frequent data transmissions also consume significant network bandwidth. Furthermore, it is difficult to balance storage cost and performance. High-performance storage media, such as Dynamic Random Access Memory (DRAM), are expensive and unsuitable for large-scale, long-term storage of all intermediate data. While low-cost storage media are economical, they are slow, such as Hard Disk Drives (HDDs) or low-speed Solid-State Drives (SSDs). Drives can easily become a bottleneck in system performance. Therefore, how to effectively control storage costs while ensuring low latency and high throughput is a core problem that urgently needs to be solved in this field. The intermediate data access patterns in server-insensitive applications exhibit high heterogeneity. Due to the extremely large range of data object sizes (from several KB to TB), diverse access patterns (including high-frequency reads, burst reads, periodic access, and related access), uncertain data lifecycles that are closely related to workflow dependencies and execution paths, and the dynamic changes in load over time (such as promotional activities) and application logic, the data access patterns are complex and dynamic, further exacerbating the complexity of storage management.

[0004] To address the aforementioned issues, existing solutions all have limitations to varying degrees. While pure memory-based remote storage solutions offer superior performance, they are extremely costly and have limited memory resources. Memory-based caching systems (such as Redis) can accelerate remote storage access, but traditional caching strategies are ineffective in the dynamic and transient environment of server-insensitive computing, prone to cache pollution or premature eviction, such as the Least Recently Used (LRU) algorithm. Some solutions also require platform modifications or suffer from cold start issues. While explorations of persistent memory (PMEM) offer performance and persistence between DRAM and SSDs, its write amplification and non-uniform memory access (NUMA) effects still require further investigation. Overall, existing technologies struggle to efficiently manage intermediate data in server-insensitive computing environments and cannot simultaneously meet the demands for low latency, high throughput, low cost, and adaptive dynamic complex access patterns. Summary of the Invention

[0005] This invention provides a hybrid storage method and system for server-insensitive computing, which can solve the technical problem that existing storage management solutions rely on fixed thresholds or manual weight scheduling and cannot adapt to the dynamic and complex access patterns of server-insensitive computing environments. It can achieve dynamic and optimal distribution of data among heterogeneous storage layers, while effectively reducing access latency and controlling storage costs.

[0006] This invention provides a hybrid storage method for server-insensitive computing, comprising: Real-time acquisition of seamless computing data access requests, and extraction of key access features based on the seamless computing data access requests; Extract the data object identifier corresponding to the key access feature, and extract the access log based on the data object identifier; Short-term predictions are made based on the access logs and key access features to obtain an access burst intensity score; Long-term predictions are made based on the access logs and key access features to obtain a set of survival probabilities based on popularity. The association prediction results are obtained based on the access logs, key access features, and pre-built dynamic graph models. The access pattern prediction result is formed based on the access burst intensity score, the heat survival probability set, and the correlation prediction result. Obtain current system status information, and output data management decisions based on the current system status information, access mode prediction results, and a preset reinforcement learning decision engine, so as to perform hybrid storage operations based on the data management decisions and a preset heterogeneous storage structure.

[0007] The above-mentioned solution achieves structured separation and prediction of short-term bursts, long-term lifecycle decay, and associated access patterns through short-term prediction, long-term prediction, and correlation prediction. After fusing the results of multi-dimensional prediction with the real-time system status, a decision is generated through a preset reinforcement learning decision engine. This avoids the limitations of single threshold or fixed weight scheduling, enabling storage decisions to simultaneously consider access latency, storage cost, and system load status. The deep coupling of multi-scale prediction and policy learning achieves dynamic optimal distribution of data across heterogeneous storage layers, while effectively reducing access latency and controlling storage costs. This solves the technical problem that existing storage management solutions rely on fixed thresholds or manual weight scheduling and cannot adapt to the dynamic and complex access patterns of server-insensitive computing environments.

[0008] Furthermore, the step of performing short-term prediction on the access logs and key access features to obtain an access burst intensity score includes: Based on the access logs and key access features, short-term extraction is performed to obtain the access frequency sequence, read / write ratio sequence, and request time interval sequence per unit time. A short-time feature sequence is constructed based on the unit time access frequency sequence, read-write ratio sequence, and request time interval sequence. The short-time feature sequence is input into a pre-constructed temporal convolutional network for convolution calculation to obtain a high-dimensional temporal feature vector. Calculate the time decay weight based on the requested time interval sequence; The time decay feature statistics are obtained by weighted summation based on the time decay weight and the access frequency sequence per unit time. The high-dimensional temporal feature vector and the time decay feature statistics are fused and mapped to obtain the access burst intensity score.

[0009] The above scheme achieves comprehensive quantification of short-term dependencies and recent sudden impacts through joint modeling of convolutional extraction and time decay weighting. This enables the sudden intensity score to respond quickly to second-level access fluctuations, providing an accurate quantitative basis for subsequently placing or retaining data in the highest performance storage layer. It solves the technical problem that traditional moving average or threshold judgment is difficult to accurately identify sudden access patterns.

[0010] Furthermore, in obtaining the association prediction result based on the access logs, key access features, and the pre-built dynamic graph model, the pre-building process of the dynamic graph model includes: Historical co-occurrence access data is obtained based on the access logs and key access features; For the current data object node corresponding to the data object identifier, obtain the long-term feature sequence corresponding to the current data object node; Based on the access log, obtain several associated data object nodes corresponding to the current data object node, and obtain the long-term feature sequences of the associated nodes corresponding to the several associated data object nodes respectively; For any of the associated data object nodes, the access time interval is extracted based on the long-term feature sequence and the long-term feature sequence of the associated nodes, and the corrected weight is obtained based on the access time interval and the preset incremental decay function. Obtain several corrected weights corresponding to the aforementioned associated data object nodes; The pre-constructed initial association graph model is updated based on several modified weights to obtain a dynamic graph model.

[0011] The above scheme enables the constructed original graph model to evolve dynamically with the access sequence, and reflects the changes in the strength of the relationship between data objects in real time according to the edge weights. This avoids the technical defect that static graph structures cannot adapt to the dynamic drift of access patterns, and provides accurate and timely graph structure data for subsequent association prediction.

[0012] Furthermore, obtaining the association prediction result based on the access logs, key access features, and the pre-built dynamic graph model includes: Based on the long-term feature sequence and several long-term feature sequences of the associated nodes, the association prediction result is obtained; The long-term feature sequence is modified based on the association prediction results and several modification weights to obtain a modified feature sequence, which is then used as the long-term feature sequence.

[0013] The above scheme enables the node representation to be dynamically updated according to the associated access pattern, so that even if the current data object has not been accessed, its feature vector can carry the potential popularity information triggered by the sudden access of associated objects, providing the subsequent reinforcement learning decision engine with accurate state features including the correction of association effects.

[0014] Further, the step of acquiring current system state information, and outputting data management decisions based on the current system state information, access pattern prediction results, and a preset reinforcement learning decision engine, to perform hybrid storage operations based on the data management decisions and a preset heterogeneous storage structure, includes: Obtain current system status information; Based on the current system status information and access mode prediction results, output the probability distribution of data management actions; The target data management action is determined based on the probability distribution of the data management actions. The target data management action and its corresponding probability value are determined based on the data management action probability distribution and preset selection conditions. Based on the target data management action, the system obtains operation judgment information from the current system status information, target action burst intensity score, target action heat survival probability, and target action association prediction result from the access mode prediction result; Based on the action probability value, target data management action, operation judgment information, target action burst intensity score, target action heat survival probability, target action association prediction result, and preset reinforcement learning decision engine, the data management decision is output. Hybrid storage operations are performed based on the data management decisions and the preset heterogeneous storage structure.

[0015] The above scheme constructs a unified state vector input policy network from short-term burst scores, long-term survival probabilities, and correlation prediction results. It selects actions through probability distribution rather than based on fixed thresholds or manual weight adjustment, enabling the decision-making process to comprehensively consider migration benefits, resource constraints, and prediction confidence, and generate optimal data management decisions that adapt to the current system state. This solves the technical problem in existing technologies where prediction results are directly used for rule judgment or weight allocation, which cannot achieve adaptive optimal decision-making.

[0016] Furthermore, the step of outputting the data management action probability distribution based on the current system state information and access mode prediction results includes: The current system state information and access mode prediction results are structurally integrated to obtain a unified state vector; The unified state vector is normalized and encoded to obtain a standard state vector; The standard state vector is input into a preset multi-layer neural network for forward propagation calculation to obtain the candidate action score vector; The candidate action scoring vector is subjected to exponential normalization to obtain several probability values ​​corresponding to several preset candidate actions. The data outputs the probability distribution of the action based on several probability values.

[0017] The above scheme realizes a computable and trainable reasoning process from multidimensional heterogeneous state input to normalized probability output, so that the action probability distribution can objectively reflect the relative merits of each candidate action in the current state.

[0018] Furthermore, the execution of hybrid storage operations based on the data management decision and the preset heterogeneous storage structure includes: When the data management decision is the first action, the hybrid storage operation includes: responding to the non-intrusive computing data access request by outputting the target access data; When the data management decision is the second action, the hybrid storage operation includes: reading a data block based on the data object identifier, writing the data block into the target storage layer located in the preset heterogeneous storage structure, and responding to the non-perceptible computing data access request to output the target access data; When the data management decision is the third action, the hybrid storage operation includes: extracting the target associated data object, preloading the target associated data object into the preload cache layer in the preset heterogeneous storage structure, and responding to the non-perceptible computing data access request to output the target access data.

[0019] The above solution is based on a pre-defined heterogeneous storage structure and is executed in a hierarchical manner through three types of actions: maintenance, migration, and prefetching. The prefetching operation decouples the data transfer delay from the foreground response through asynchronous background loading, so that the current request is not affected by the prefetching time, while preparing high-performance cached data in advance for future access.

[0020] Furthermore, after obtaining the current system state information, and outputting a data management decision based on the current system state information, access pattern prediction results, and a preset reinforcement learning decision engine, and then performing a hybrid storage operation based on the data management decision and a preset heterogeneous storage structure, the method further includes: Obtain data on operation execution overhead and performance impact; The preset reinforcement learning decision engine is optimized and updated based on the operation execution overhead and performance impact data to obtain the target reinforcement learning decision engine, and the target reinforcement learning decision engine is used as the preset reinforcement learning decision engine.

[0021] After performing hybrid storage operations, the above scheme further acquires data on operation execution overhead and performance impact. Then, based on this execution feedback data, it optimizes and updates the preset reinforcement learning decision engine. By calculating the difference between the actual reward value and the expected reward and updating the weights of the policy network and value network through backpropagation, the decision engine can learn from each execution result and continuously improve the decision quality. Finally, the target reinforcement learning decision engine is obtained and replaces the original preset engine, enabling the reinforcement learning decision engine to adapt to dynamic changes in workload online.

[0022] This invention also provides a hybrid storage system for server-insensitive computing, used to implement the above-mentioned hybrid storage method for server-insensitive computing. The system includes: The access monitoring module is used to acquire non-perceptible computing data access requests in real time and extract key access features based on the non-perceptible computing data access requests. The data extraction module is used to extract the data object identifier corresponding to the key access feature, and extract the access log based on the data object identifier; The access prediction module is used to perform short-term predictions on the access logs and key access features to obtain an access burst intensity score; perform long-term predictions on the access logs and key access features to obtain a set of survival probabilities based on popularity; obtain association prediction results based on the access logs, key access features, and a pre-built dynamic graph model; and form access pattern prediction results based on the access burst intensity score, the set of survival probabilities based on popularity, and the association prediction results. The hybrid storage execution module is used to acquire the current system status information, and output data management decisions based on the current system status information, access mode prediction results, and a preset reinforcement learning decision engine, so as to execute hybrid storage operations based on the data management decisions and the preset heterogeneous storage structure.

[0023] The above solution obtains non-perceptible computing data access requests in real time and extracts key access features through the access monitoring module. The data extraction module extracts access logs based on the data object identifiers corresponding to the key access features. The access prediction module performs short-term prediction to obtain a burst intensity score, long-term prediction to obtain a set of heat survival probabilities, and obtains correlation prediction results based on a dynamic graph model and integrates them into access pattern prediction results. The hybrid storage execution module combines the current system status information and access pattern prediction results, outputs data management decisions through a preset reinforcement learning decision engine, and executes hybrid storage operations. This enables multi-dimensional prediction, reinforcement decision-making, and hybrid storage execution to work together at the system level, realizing a closed loop of intelligent prediction and decision-making.

[0024] Furthermore, it also includes a dynamic update module, which is used to: acquire operation execution overhead and performance impact data; optimize and update the preset reinforcement learning decision engine according to the operation execution overhead and performance impact data to obtain a target reinforcement learning decision engine, and use the target reinforcement learning decision engine as the preset reinforcement learning decision engine.

[0025] The above scheme obtains data on operation execution overhead and performance impact after performing hybrid storage operations through a dynamic update module. Based on this feedback data, it optimizes and updates the preset reinforcement learning decision engine to obtain the target reinforcement learning decision engine, which is then used as the engine for subsequent decisions. This enables the system to have online learning and continuous policy improvement capabilities. The decision engine can continuously adjust policy parameters based on actual execution feedback to adapt to dynamically changing access loads and system states, allowing the system to maintain optimal decision-making and adaptability during long-term operation. Attached Figure Description

[0026] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of a hybrid storage method for server-insensitive computing provided in this embodiment; Figure 2 This is a schematic diagram of a short-term burst detection module provided in this embodiment; Figure 3 This is a schematic diagram of a long-term pattern prediction module provided in this embodiment; Figure 4 This is a schematic diagram of a correlation data prediction module provided in this embodiment. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0030] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0031] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0032] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0033] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0034] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0035] Example 1: This embodiment provides a hybrid storage method for server-insensitive computing, such as... Figure 1 As shown, it includes: S1. Real-time acquisition of seamless computing data access requests, and extraction of key access features based on the seamless computing data access requests; S2. Extract the data object identifier corresponding to the key access feature, and extract the access log based on the data object identifier; S3. Perform short-term prediction on the access logs and key access features to obtain an access burst intensity score; S4. Perform long-term prediction on the access logs and key access features to obtain a set of survival probabilities based on popularity. S5. Obtain association prediction results based on the access logs, key access features, and pre-built dynamic graph model; S6. Based on the access burst intensity score, the heat survival probability set, and the correlation prediction results, an access pattern prediction result is formed; S7. Obtain the current system status information, and output data management decisions based on the current system status information, access mode prediction results, and preset reinforcement learning decision engine, so as to perform hybrid storage operations based on the data management decisions and preset heterogeneous storage structure.

[0036] In its implementation, this embodiment provides a hybrid storage method for server-insensitive computing, implemented using a hybrid storage system. After the hybrid storage system starts, the access monitoring module enters a standby state, continuously listening for server-insensitive functions' access requests to intermediate data. When a new data access request arrives, the access monitoring module intercepts it and extracts key access feature information from the intercepted request. This key access feature information includes, but is not limited to: data object identifier, access type (read or write), request initiation timestamp, data size, source function identifier, and non-uniform memory access node information. The extracted access node information is then structured to form key access features. Next, the data object identifier is extracted from the access event records, and based on this identifier, the corresponding access log for that data object is extracted from the system log, including access time, access frequency, read / write operation type, and request time interval.

[0037] In the dynamic prediction phase of access patterns, based on the key access features extracted above, the dynamic prediction engine predicts the future access patterns of intermediate data processed without the server's awareness. These future access patterns include short-term access bursts, long-term access popularity trends, and the probability of related access between data. Specifically, the access event records generated after extracting key access features, along with relevant historical access records under the corresponding data object identifiers, are input into the dynamic prediction engine. Three independently modeled sub-modules within the engine then perform prediction calculations at different time scales and dimensions.

[0038] Specifically, the dynamic prediction engine calls the Short-Term Burst Detection (STBD) module for short-term prediction. Based on the recent access event sequence of the target data object, it uses a temporal convolutional network and impulse response filter model to perform short-term prediction calculations, outputting an access burst intensity score for the data object in the near future. The dynamic prediction engine calls the Long-Term Pattern Prediction (LTPP) module for long-term prediction. Based on the target data object's longer historical access sequence, it uses a hierarchical Transformer model and Kaplan-Meier survival analysis model to perform long-term prediction calculations, outputting a set of heat survival probabilities for the data object at multiple future time points. The dynamic prediction engine calls the Correlation-based Prediction (CorrBP) module for correlation prediction. Based on historical co-occurrence access data and a pre-built dynamic graph neural network model (i.e., the dynamic graph model), it performs correlation prediction, outputting a list of other data objects strongly correlated with the currently accessed object and their correlation hit reward potential.

[0039] It should be noted that the dynamic graph model in this embodiment is built based on historical access logs. Each data object corresponds to a node in the graph. When two data objects are accessed within a similar time window, a weighted edge is established between the corresponding nodes. The initial weight of the edge can be set to zero or initialized according to the number of historical co-occurrences. Each time a co-occurrence access occurs, the edge weight is updated according to the basic increment coefficient and the time decay function, so that the edge weight corresponding to the recently frequently co-occurring objects is rapidly increased, and the influence of historical co-occurrence gradually weakens over time, thus forming a graph structure that dynamically changes with the access sequence. In the completed dynamic graph input graph attention network, the neighbor node representations of the target node are linearly transformed and then concatenated into the feedforward neural network. After normalization, the contribution weights of each neighbor node to the future access patterns of the target node are obtained. The representation of the target node is then updated by weighted summation of the neighbor node representations. The association hit reward potential of each neighbor node to the target node is determined by the contribution weight of the corresponding neighbor node and the edge weight. The higher the value, the higher the contribution of the neighbor node to the prediction of the future access behavior of the target node. Based on this, the system generates a list of data objects that are most likely to have strong associations and their reward potentials, providing a decision-making basis for subsequent data management actions.

[0040] After completing the prediction calculations for the above three dimensions, the access burst intensity score, the heat survival probability set, and the correlation prediction results are integrated to form a comprehensive prediction view of the future access patterns of the data object. The access pattern prediction results are output as the input basis for subsequent reinforcement learning decisions.

[0041] In specific implementation, the preset heterogeneous storage structure in this embodiment is a three-layer cache structure implemented through a hybrid storage execution module, including an instantaneous cache (IC-DRAM, Instant Cache in DRAM) for handling burst access, a main cache (MC-DRAM, Main Cache in DRAM), and a persistent cache (PC-PMEM, Persistent Cache in PMEM) utilizing the second medium-speed persistent storage layer. When performing operations involving the second medium-speed persistent storage layer (e.g., PMEM), direct access is performed using memory mapping (mmap), and / or atomic write operations are performed using the Persistent Memory Development Kit (PMDK). When performing operations that degrade data from a higher performance layer to the third low-speed storage layer (e.g., SSD), asynchronous batch writes are performed using a batch data degradation pipeline.

[0042] Optionally, the specific execution flow of the batch data degradation pipeline is as follows: After the process starts, the data set to be degraded, coolingData, and the preset batch size threshold, BATCH_SIZE, are first input, and the batch set used to temporarily store the data blocks to be written is initialized to an empty set; then, the data set to be degraded is traversed in a loop to determine whether there are still unprocessed data blocks d in the data set to be degraded. If there are unprocessed data blocks, the data block d is added to the current batch set; after the data block is added, it is determined whether the length of the current batch set is equal to the preset BATCH_SIZE. If the batch threshold is not reached, the process returns to continue to determine the remaining data blocks in the data set to be degraded and continues to add data blocks to the batch set. The process involves accumulating data to be downgraded within a batch. If the batch size reaches the BATCH_SIZE threshold, the batch of data blocks is written to the solid-state drive (SSD) storage layer via an asynchronous input / output interface, and the corresponding metadata information is updated synchronously. After completing the downgrade writing for this batch, the process returns to continue traversing the remaining data blocks in the data set to be downgraded. When all data blocks in the data set to be downgraded have been traversed and processed, it is determined whether there is any remaining data in the current batch that has not reached the batch threshold. If there is no remaining data, the downgrade process ends. If there is remaining data, all remaining data blocks in the batch are written to the SSD storage layer and the corresponding metadata is updated. The process ends after all the data to be downgraded is migrated.

[0043] It should be noted that the above batch degradation process is executed as an asynchronous background task. The solid-state drive write operation is decoupled from the front-end data read and write operation, and will not block normal business access requests. By using batch aggregated write, the operation overhead of a single small data migration can be amortized, making full use of the performance advantages of sequential write of solid-state drives, and effectively reducing the interference of data degradation operation on the latency of front-end business access.

[0044] Optionally, the step of performing short-term prediction on the access logs and key access features to obtain an access burst intensity score includes: Based on the access logs and key access features, short-term extraction is performed to obtain the access frequency sequence, read / write ratio sequence, and request time interval sequence per unit time. A short-time feature sequence is constructed based on the unit time access frequency sequence, read-write ratio sequence, and request time interval sequence. The short-time feature sequence is input into a pre-constructed temporal convolutional network for convolution calculation to obtain a high-dimensional temporal feature vector. Calculate the time decay weight based on the requested time interval sequence; The time decay feature statistics are obtained by weighted summation based on the time decay weight and the access frequency sequence per unit time. The high-dimensional temporal feature vector and the time decay feature statistics are fused and mapped to obtain the access burst intensity score.

[0045] In its implementation, this embodiment uses a short-term burst detection module to quickly identify and quantify the short-term burstiness of data access. Preferably, a model structure combining a temporal convolutional network and an impulse response filter is adopted. The temporal convolutional network processes time-series data in parallel through residual blocks with dilated convolutions, which can effectively capture short-term dependencies and has high computational efficiency. The impulse response filter assigns higher weights to access events that are more recent in time, highlighting the impact of recent accesses and enabling a rapid response to bursty access behavior.

[0046] In one specific embodiment, short-term features are first extracted based on access logs and key access characteristics. Statistics such as the number of accesses within a unit time window (included in the unit time access frequency sequence), read / write ratio (included in the read / write ratio sequence), and average request interval (included in the request interval sequence) are statistically analyzed. These three types of sequences form a short-term feature sequence, which is then normalized to eliminate dimensional differences. Subsequently, the short-term feature sequence is input into a residual block structure with dilated convolutions. Each convolutional layer performs a one-dimensional convolution operation on the input sequence, i.e., a weighted summation of a k-length convolutional kernel with the corresponding time step data. The dilation coefficient controls the sampling interval of the convolutional kernel, expanding the receptive field without increasing computational depth and capturing short-term dependencies over longer spans. The residual connection adds the input and convolutional output and maps it through a non-linear activation function, avoiding gradient decay problems in deep networks and improving the stability of feature representation. After multi-layer residual block stacking and computation, a high-dimensional temporal feature vector is output. Meanwhile, the impulse response filtering module calculates the time decay weights for each historical time step based on the request time interval sequence. The weight coefficients can be calculated using an exponential decay function or an inverse distance function, giving greater weight to data at time steps closer to the current time. The calculated time decay weights are then weighted and summed with the access frequency sequence per unit time, and the weight coefficients are normalized to maintain numerical stability, ultimately yielding the time decay feature statistics. After completing the two feature calculations, the high-dimensional temporal feature vector output by the temporal convolutional network is concatenated or weighted and fused with the time decay feature statistics obtained from the impulse response filtering. This is then mapped through a fully connected layer to a single access burst intensity score. This score can be compressed to the 0-1 range using a nonlinear function to characterize the probability intensity of a sudden increase in access to the current data object within a short period.

[0047] After inputting the access event sequence of the target data object within a recent time window, the system uses the complete calculation process of convolutional feature extraction, time decay weighting, and linear mapping to jointly model short-term dependencies and recent sudden impacts. Unlike traditional simple moving average or threshold judgment methods, it outputs a quantitative access burst intensity score. When the access burst intensity score is at a high level, it will trigger the subsequent reinforcement learning decision engine to urgently place or retain the corresponding data object in the highest-performance dynamic random access memory layer to cope with the upcoming access peak and minimize data access latency.

[0048] It should be noted that in this embodiment, the short-term prediction is set to the second level, that is, the time window for short-term prediction is set to within a few seconds or tens of seconds.

[0049] In another specific embodiment, this embodiment provides a method such as Figure 2 The short-term burst detection module shown includes two computational branches: an impulse response filter and a temporal convolutional network (TCN). The outputs of both branches are ultimately fed into a burst intensity quantizer, which outputs an access burst intensity score. The impulse response filter calculates time decay weights based on the request time interval sequence in the access history (i.e., the access log). The temporal convolutional network receives the preprocessed short-term feature sequence as an input tensor. Internally, the network consists of multiple cascaded residual blocks, sequentially including residual blocks (k, b) along the data flow direction. 0), residual block (k,b) 1) Until the residual block (k,b) (n-1) is used to extract deep features of the temporal dimension layer by layer. A single residual block (k,d) receives the output of the previous residual block as input and contains two basic layers (k,d). Each basic layer is configured with Weightnorm (weight normalization), ReLU (corrected linear unit), and Dropout (random deactivation regularization) components to implement weight normalization, non-linear activation, and network regularization, respectively. The residual block also has an optional 1 One convolution is used to align channels when the input and output dimensions do not match. The convolution output and the residual input are fused by an addition node and then passed down as the input to the next residual block. The core computational unit of the residual block is a causal convolutional layer (k,d), which adopts an dilated convolution structure with a receptive field of 1 + (k-1). d, can expand the temporal receptive field without increasing network depth, and outputs a high-dimensional temporal feature vector in the form of an output tensor after computation. The burst intensity quantizer is used to receive and fuse two features, namely the high-dimensional temporal feature vector and the time decay feature statistics, and outputs a normalized access burst intensity score after fully connected mapping and nonlinear compression.

[0050] In another specific embodiment, this embodiment provides a method such as Figure 3 The long-term pattern prediction module shown includes a Transformer unit and a Kaplan-Meier survival analysis engine. It takes historical access data (i.e., long-sequence data extracted from access logs and key access features) corresponding to the target data object as input and outputs the survival probability of the data object at multiple future time points. The hierarchical Transformer unit adopts a hierarchical architecture of segmented encoding and cross-segment aggregation. Along the data flow direction, it sequentially sets up an embedding layer, a positional encoding layer, N sets of interleaved Transformer Encoders, a pooling layer, and a fully connected layer. The interleaved Transformer Encoder internally consists of multiple sets of segment-wise encoders (SWE) and cross-segment encoders (CSE) stacked alternately. Each level contains multiple parallel segment encoders, each processing a segment of the access sequence. The cross-segment encoder performs global information aggregation on the outputs of all segment encoders at the same level, outputting a global feature representation of the long sequence after multi-level alternating encoding. The global features output by the hierarchical Transformer unit are fed into the downstream Kaplan-Meier survival analysis engine, which then performs the final survival probability calculation and outputs the results.

[0051] It should be noted that the long-term pattern prediction module is used to predict the trend of access popularity of data objects and the approximate cooling time nodes in the medium and long term (minute-level, hour-level) based on long-sequence data. Its input is the access records of the target data object in a long historical period, and the output is a set of popularity survival probability covering multiple future time points, which is used to guide subsequent medium and long-term data migration decisions.

[0052] In practice, the long-sequence historical access data of the target data object is first divided into multiple equally long segmented access sequences according to fixed time periods. All segmented access sequences are input into hierarchical Transformer units. The discrete access statistical features are first converted into high-dimensional vector representations by an embedding layer, and then temporal position information is injected into the features at each time step by a positional encoding layer, fully preserving the sequential dependencies of the sequences. The sequence features after embedding and positional encoding are then fed into N sets of interleaved stacked Transformer Encoder structures. Within each encoder, multiple parallel segmented encoders encode and compute the access sequences for each segment, obtaining segment-level feature representations for each segment. Then, a cross-segment encoder performs global aggregation of all segment-level features, capturing long-span access dependencies across segments. After multiple iterations of segmented encoding and cross-segment aggregation, the deep temporal access features inherent in the long sequence are fully extracted. The feature sequence output by the encoder is compressed in dimension and aggregated globally by a pooling layer, and then mapped through a fully connected layer to obtain a global feature vector representing the long-term access patterns of the data object.

[0053] Meanwhile, based on the statistical results of the full historical access data, the Kaplan-Meier survival analysis method is used to count the number of data objects that remain active (access frequency higher than a preset threshold) and the number of data objects that have entered the failure state at each time point, and calculate the baseline survival function. Then, the global feature vector output by the hierarchical Transformer is mapped to the potential failure risk corresponding to each time period through a feedforward network. The baseline survival function is conditionally corrected with this risk value, and finally the probability that the target data object will remain active at each time point in the future is output, which is the popularity survival probability. The survival probabilities corresponding to all time points together constitute the popularity survival probability set.

[0054] This set of survival probabilities based on popularity will serve as one of the inputs to the subsequent preset reinforcement learning decision engine to guide medium- to long-term data migration decisions. When the predicted data object maintains a high survival probability and remains active for a long period of time, the decision engine tends to retain it in high-performance storage layers such as persistent memory or dynamic random access memory. When the predicted data object's survival probability drops rapidly and it is about to enter the cooling-off phase, the decision engine plans to downgrade it to a low-speed, high-capacity storage layer such as a solid-state drive to free up valuable high-performance storage space.

[0055] Optionally, in obtaining the association prediction result based on the access logs, key access features, and the pre-built dynamic graph model, the pre-building process of the dynamic graph model includes: Historical co-occurrence access data is obtained based on the access logs and key access features; For the current data object node corresponding to the data object identifier, obtain the long-term feature sequence corresponding to the current data object node; Based on the access log, obtain several associated data object nodes corresponding to the current data object node, and obtain the long-term feature sequences of the associated nodes corresponding to the several associated data object nodes respectively; For any of the associated data object nodes, the access time interval is extracted based on the long-term feature sequence and the long-term feature sequence of the associated nodes, and the corrected weight is obtained based on the access time interval and the preset incremental decay function. Obtain several corrected weights corresponding to the aforementioned associated data object nodes; The pre-constructed initial association graph model is updated based on several modified weights to obtain a dynamic graph model.

[0056] In the specific implementation process, this embodiment first constructs a dynamic graph model based on the co-occurrence access data of all collected data object nodes. Each data object in the system is represented as a data object node in the dynamic graph model. The feature sequence corresponding to the data object node (i.e., the long-term feature sequence, the long-term feature sequence of associated nodes, etc.) includes its access frequency, type, etc. A connection edge is established between two corresponding nodes. The initial weight of the edge can be set to zero or initialized according to the historical co-occurrence count.

[0057] In the specific implementation of the correlation data prediction module, the real-time dynamic graph model is constructed and updated based on access logs and historical co-occurrence access data extracted from key access features. First, each data object identifier is mapped to a node in the graph structure, with the data object to be predicted as the current data object node, and the long-term feature sequence corresponding to this current data object node is obtained. Simultaneously, based on the access logs, several associated data object nodes that have a co-occurrence access relationship with the current data object are identified, and the long-term feature sequence of each associated data object node is obtained. Node features include attribute information such as the access frequency and access type of the corresponding data object.

[0058] When two data objects are accessed within a preset contiguous time window, for any associated data object node, the access time interval is extracted from the long-term feature sequence of the current data object node and the long-term feature sequence of the associated node. Based on the access time interval and a preset incremental decay function, the corrected weight of the corresponding edge is calculated. In this embodiment, the edge weight update adopts a mechanism of incremental superposition combined with time decay. Each time a co-occurrence access occurs, the edge weight is updated according to the preset incremental decay function shown in the following formula: ; In the formula, The updated corrected weights are defined by α as the basic increment coefficient, Δt as the access time interval, and τ as the decay constant. This update mechanism allows the edge weights of objects that frequently co-occur in recent times to be rapidly increased, while the influence of historical co-occurrence gradually decays over time, ensuring that the graph structure can reflect the changes in the strength of associations between data objects in real time.

[0059] After obtaining the corrected weights corresponding to all associated data object nodes, the edge weights of the pre-constructed initial association graph model are updated based on each corrected weight, and finally a dynamic graph model that changes dynamically with the access sequence is obtained, providing a basic graph structure for the association prediction of the subsequent graph attention network.

[0060] Optionally, obtaining the association prediction result based on the access logs, key access features, and pre-built dynamic graph model includes: Based on the long-term feature sequence and several long-term feature sequences of the associated nodes, the association prediction result is obtained; The long-term feature sequence is modified based on the association prediction results and several modification weights to obtain a modified feature sequence, which is then used as the long-term feature sequence.

[0061] In one specific embodiment, this embodiment provides as follows: Figure 4 The illustrated relational data prediction module includes an input layer, a graph attention network (GAT), two linear layers, and an output layer. The graph attention network (GAT) comprises the following two core computational processes: First, a linear transformation is performed on the features of each node to obtain the transformed feature vector. , The association prediction results corresponding to each neighbor node are calculated using the LeakyReLU activation function; the attention coefficients corresponding to each neighbor node are also calculated using the LeakyReLU activation function. The process includes: processing the long-term feature sequence of the current data object node i respectively. And the long-term feature sequence of each associated data object node. Multiplying by a shared weight matrix W maps the features to a higher-dimensional space more suitable for calculating correlations, resulting in the transformed feature map. , The transformed , Concatenate them together to form a joint vector. The joint vector is passed into the LeakyReLU activation function and then normalized using Softmax to obtain the attention coefficients. Based on this attention coefficient The correlation hit reward potential is calculated by multiplying the modified weights obtained above, and the correlation hit reward potential and attention coefficient are output as the correlation prediction result.

[0062] Second, based on the attention coefficient, a weighted summation is performed on the long-term feature sequences of several associated nodes to update the final representation of the target node, resulting in a corrected feature sequence. The specific calculation process is shown in the following formula: .

[0063] For example, if the initial edge weight w=0, and α=0.1 and τ=10 are set, then the corrected weight is calculated at the access time interval Δt=2 minutes. The weight is approximately 0.082; after another co-occurrence visit at a 5-minute interval, the weight is adjusted. Approximately 0.143, This embodiment uses the above mechanism to rapidly increase the edge weights of objects that frequently co-occur in the near future, while the influence of historical co-occurrence gradually diminishes over time, ensuring that the graph structure can reflect changes in the strength of associations between data objects in real time.

[0064] Optionally, the step of obtaining the current system state information, and outputting a data management decision based on the current system state information, access pattern prediction results, and a preset reinforcement learning decision engine, to perform hybrid storage operations based on the data management decision and a preset heterogeneous storage structure, includes: Obtain current system status information; Based on the current system status information and access mode prediction results, output the probability distribution of data management actions; The target data management action is determined based on the probability distribution of the data management actions. The target data management action and its corresponding probability value are determined based on the data management action probability distribution and preset selection conditions. Based on the target data management action, the system obtains operation judgment information from the current system status information, target action burst intensity score, target action heat survival probability, and target action association prediction result from the access mode prediction result; Based on the action probability value, target data management action, operation judgment information, target action burst intensity score, target action heat survival probability, target action association prediction result, and preset reinforcement learning decision engine, the data management decision is output. Hybrid storage operations are performed based on the data management decisions and the preset heterogeneous storage structure.

[0065] In one specific embodiment, the preset heterogeneous storage structure includes a first high-speed storage layer (e.g., DRAM), a second medium-speed persistent storage layer (e.g., PMEM), and a third low-speed storage layer (e.g., SSD). The transient cache (IC-DRAM) is located in the DRAM layer. The DRAM layer has a small capacity and uses an extremely short Time-To-Live (TTL), such as 1 second. It is mainly used to handle burst access data predicted by STBD, providing the lowest latency response. Data is evicted to the next layer after the TTL expires or when preempted by higher-priority burst data. The main cache (MC-DRAM) is located in the DRAM layer, has a larger capacity, and is the main cache area. It handles data evicted from IC-DRAM and non-burst hot data that the decision engine decides to upgrade to DRAM, employing a more complex cache replacement strategy (such as LFU, LRU, or a strategy based on LTPP prediction). The persistent cache (PC-PMEM) is located in the PMEM layer, has a larger capacity, and utilizes the persistence of PMEM and its lower latency compared to SSD to handle data degraded from MC-DRAM and data that the decision engine decides to place in the PMEM layer. It also employs an appropriate cache replacement strategy. Data flows between these caching layers based on TTL, access frequency prediction, and decision instructions, with the SSD layer serving as the final archive storage.

[0066] In the specific implementation process, this embodiment utilizes a preset reinforcement learning decision engine to generate data management decisions for the intermediate data, so as to specify the storage location of the data corresponding to the current data object in the preset heterogeneous storage structure or to perform migration operations according to the data management decisions. In this embodiment, the preset reinforcement learning decision engine adopts the Proximal Policy Optimization (PPO) algorithm, which ensures the stability of the learning process by limiting the policy update magnitude. For each current data object that needs to be decided, its state space includes the data object's own characteristics and the current system state information: the data object's own characteristics include action probability value, target data management action, operation judgment information, target action burst intensity score, target action heat survival probability, and target action association prediction result, as well as the storage location corresponding to the current data object, etc.; the current system state information includes the recent average read latency, tail latency, average read / write throughput, recent migration overhead, and the current utilization rate of the dynamic random access memory layer, persistent memory layer, and solid-state drive layer.

[0067] In this embodiment, a preset reinforcement learning decision engine selects an action for each data object or a batch of data objects. The action space includes a first action, which is to keep the data object unchanged in the current storage layer; a second action, which is to upgrade and migrate the data object from the persistent memory layer or solid-state drive layer to the dynamic random access memory layer, downgrade and migrate the data object from the dynamic random access memory layer to the persistent memory layer, and migrate the data object bidirectionally between the persistent memory layer and the solid-state drive layer; and a third action, which is to prefetch the strongly correlated data objects identified by the correlation data prediction module to the dynamic random access memory layer or the persistent memory layer.

[0068] The current system status information also includes the real-time storage utilization of the dynamic random access memory layer, persistent memory layer, and solid-state drive layer, the system's recent average read latency and tail latency, the system's recent read / write throughput, and the recent migration operation overhead. Access pattern prediction results and current system status information are provided as input to the policy network. The policy network first structurally integrates monitoring metrics and various prediction results to construct a unified state vector. This vector includes the current storage level, access frequency, recent access interval, burst score, survival probability, association strength index, resource utilization of each layer, and migration overhead estimation features. Numerical features are normalized, and categorical features are encoded and transformed to obtain a standard state vector. Then, the standard state vector is input into a multi-layer neural network for forward propagation calculation. Each hidden layer is mapped layer by layer through matrix multiplication, bias superposition, and nonlinear activation functions, outputting an action score vector with dimensions equal to the number of candidate data management actions. Finally, the action score vector is exponentially normalized to obtain the probability value corresponding to each candidate action, ensuring that the sum of all probability values ​​is 1, forming a complete action probability distribution. During the execution phase, the action with the highest probability is selected or the final decision is determined by sampling according to probability. During the strategy optimization phase, the difference in revenue is calculated based on the performance indicators obtained after execution, and the network weights are updated through backpropagation.

[0069] The specific logic for action selection is as follows: When the probability of maintaining the action is the highest and exceeds the preset stability threshold, the data object is kept in the original storage layer; when the probability of upgrading the action is the highest, and the object is currently located in the persistent memory layer or solid-state drive layer, and the burst score is high, the survival probability is high, the access frequency is continuously increasing, and the utilization rate of the high-performance storage layer has not reached the upper limit, the upgrade migration is executed; when the probability of downgrading the action is the highest, and the object access frequency decreases, the survival probability is in the low to medium range, the short-term burst score is low, and the utilization rate of the high-performance storage layer is close to the threshold, the downgrading is executed to release high-performance resources; when the probability of inter-layer migration is the highest, and the object's life cycle is predicted to be in the end of the decay stage or the access frequency is extremely low but still has some retention value, the corresponding inter-layer migration is executed; when the probability of prefetching the action is the highest, and the association strength index exceeds the set threshold, the target object has not yet resided in the high-performance layer, and there is still spare resources, the preloading of strongly associated data objects is triggered.

[0070] The reward function guides the reinforcement learning agent to maximize long-term cumulative rewards and consists of a weighted sum of several sub-items: latency reward is negatively correlated with average latency, tail latency, and burst latency peaks; throughput reward is positively correlated with read / write throughput; transfer penalty is negatively correlated with transfer overhead; utilization penalty is negatively correlated with storage utilization at each layer; and association hit reward is positively correlated with the reward value of successfully prefetched and accessed associated data. Each weight factor is adjusted experimentally. The actor network and evaluator network are optimized using a proximal policy optimization algorithm. The actor network outputs action probabilities, the evaluator network evaluates state values, and the decision engine learns a policy to make near-optimal decisions in complex states, dynamically balancing performance and cost.

[0071] Optionally, the step of outputting the data management action probability distribution based on the current system state information and access mode prediction results includes: The current system state information and access mode prediction results are structurally integrated to obtain a unified state vector; The unified state vector is normalized and encoded to obtain a standard state vector; The standard state vector is input into a preset multi-layer neural network for forward propagation calculation to obtain the candidate action score vector; The candidate action scoring vector is subjected to exponential normalization to obtain several probability values ​​corresponding to several preset candidate actions. The data outputs the probability distribution of the action based on several probability values.

[0072] In the specific implementation process, the reinforcement learning decision engine first acquires the current global system state information. This current system state information includes at least the real-time storage utilization rates of the dynamic random access memory layer, persistent memory layer, and solid-state drive layer in the multi-layer heterogeneous storage pool, as well as one or more of the following: recent average access latency, system tail latency, system throughput, and recent migration overhead. Then, based on the current system state information and access pattern prediction results, the internal policy network infers and outputs a data management action probability distribution. Based on this data management action probability distribution, candidate target data management actions are initially determined. Finally, combining the data management action probability distribution with preset selection conditions, the target data management action is determined. The system performs a data management operation and extracts the corresponding probability value of the action. After determining the target data management action, it extracts matching operation judgment information from the current system status information based on the target data management action. At the same time, it extracts the target action burst intensity score, target action heat survival probability, and target action association prediction results from the access mode prediction results. Finally, it combines the action probability value, target data management action, operation judgment information, target action burst intensity score, target action heat survival probability, and target action association prediction results to generate the final data management decision through a reinforcement learning decision engine. Based on the data management decision, it performs the corresponding hybrid storage operation in the preset multi-layer heterogeneous storage structure.

[0073] When outputting the probability distribution of data management actions based on the current system state information and access pattern prediction results, the current system state information and access pattern prediction results are first structurally integrated to construct a unified state vector containing features such as the current storage level of the data object, access frequency, recent access interval, burst intensity score, multi-time point heat survival probability, association strength index, storage resource utilization of each layer, and migration overhead estimation. Then, the numerical features in the unified state vector are normalized, and the categorical features are encoded and converted to eliminate the dimensional differences of different features to obtain a standard state vector. Subsequently, the standard state vector is input into a preset multi-layer neural network for forward propagation calculation. Each hidden layer completes layer-by-layer feature mapping through matrix operations, bias superposition, and nonlinear activation functions, outputting a candidate action score vector with a dimension consistent with the preset number of candidate data management actions. Then, the candidate action score vector is subjected to exponential normalization to obtain the probability values ​​corresponding to several preset candidate actions, and the sum of the probability values ​​of all candidate actions is 1. Based on the probability values ​​corresponding to each candidate action, a complete probability distribution of data management actions is output.

[0074] It should be noted that the reinforcement learning decision engine in this embodiment preferably adopts the Proximal Policy Optimization (PPO) algorithm. This algorithm belongs to the actor-critic reinforcement learning algorithm, which ensures the stability of the learning process by limiting the magnitude of policy updates in each round, and is suitable for the complex decision-making needs of large-scale discrete state spaces and action spaces in this scenario. The candidate data management actions preset in this embodiment include five categories: maintaining the data object unchanged in the current storage layer, upgrading and migrating the data object from the persistent memory layer or solid-state disk layer to the dynamic random access memory layer, downgrading and migrating the data object from the dynamic random access memory layer to the persistent memory layer, performing bidirectional migration between the persistent memory layer and the solid-state disk layer, and prefetching strongly correlated data objects to the high-performance storage layer. In the actual execution stage, a deterministic strategy can be used to directly select the action with the highest probability, or an exploration mechanism can be used to randomly sample according to the probability distribution. All action selections must meet the storage layer legality and system resource capacity constraints. During the strategy optimization phase, the engine calculates the revenue difference based on the performance metrics feedback after data operation execution, updates the network weights through backpropagation, and achieves continuous iterative optimization of the decision-making strategy. The reward function used to guide strategy learning comprehensively considers access latency, system throughput, migration overhead, storage utilization, and the prefetching hit rate of related data. It consists of a weighted sum of latency reward, throughput reward, migration penalty, utilization penalty, and association hit reward. The latency reward is negatively correlated with system access latency to drive the minimization of access latency; the throughput reward is positively correlated with system read and write throughput to improve the overall system throughput; the migration penalty is negatively correlated with data migration overhead to suppress meaningless frequent migrations; the utilization penalty is negatively correlated with the utilization of high-cost storage layers to control overall storage costs; and the association hit reward is positively correlated with the actual hit rate of prefetched data to encourage effective association data prefetching strategies. The weight coefficients of each reward sub-item can be adjusted through experiments according to the performance and cost preferences of the actual business scenario.

[0075] Optionally, the step of performing hybrid storage operations based on the data management decision and the preset heterogeneous storage structure includes: When the data management decision is the first action, the hybrid storage operation includes: responding to the non-intrusive computing data access request by outputting the target access data; When the data management decision is the second action, the hybrid storage operation includes: reading a data block based on the data object identifier, writing the data block into the target storage layer located in the preset heterogeneous storage structure, and responding to the non-perceptible computing data access request to output the target access data; When the data management decision is the third action, the hybrid storage operation includes: extracting the target associated data object, preloading the target associated data object into the preload cache layer in the preset heterogeneous storage structure, and responding to the non-perceptible computing data access request to output the target access data.

[0076] In the specific implementation process, this embodiment adopts a hybrid storage execution module that receives the data management action instruction selected by the reinforcement learning decision engine, and performs the corresponding data placement or migration operation in the multi-layer heterogeneous storage pool according to the specific type of the action instruction.

[0077] During the action determination phase, the action probability distribution results are combined with the current system state information for joint constraint judgment. When the action probability of keeping the data object unchanged in the current storage layer is the highest among all candidate actions, and this probability exceeds the preset stability threshold, it is determined that the migration benefit in the current state is insufficient to cover the migration cost or the data access popularity is in a stable range, and the corresponding maintenance operation is executed.

[0078] When the probability of upgrading and migrating a data object to the dynamic random access memory layer is the highest, and the data object is currently located in the persistent memory layer or solid-state drive layer, and at the same time, the access burst strength score is high, the hot survival probability is high, the access frequency continues to rise, and the utilization rate of the high-performance storage layer has not reached the upper limit, the corresponding upgrade and migration operation is performed.

[0079] When the probability of downgrading and migrating data objects to the persistent memory layer is highest, and the data object access frequency decreases, the hot survival probability is in the low to medium range, the short-term burst intensity score is low, and the utilization rate of the high-performance storage layer is close to the threshold, the downgrading and migration operation should be performed to release high-performance storage resources.

[0080] When the probability of performing a migration operation between the persistent memory layer and the solid-state drive layer is highest, and the data object's lifecycle is predicted to be entering the end of its decay phase or its access frequency is extremely low but it still has some retention value, an upgrade or downgrade migration operation in the corresponding direction is performed according to the current layer of the data object.

[0081] It should be noted that in this embodiment, the survival probability of the heat at multiple consecutive time points output by the long-term pattern prediction module is used to determine whether the life cycle of the data object has entered the decay tail stage. When the survival probability at a future preset time point is lower than the preset decay threshold and the subsequent multiple consecutive predicted time points are all lower than the threshold, or when the decrease in the survival probability at multiple consecutive time points exceeds the preset decay rate threshold, the life cycle of the data object is determined to have entered the decay tail stage. In the determination process, features such as short-term burst intensity score and actual access interval can also be combined for auxiliary verification.

[0082] When the probability of prefetching strongly correlated data objects is the highest, and the correlation strength index exceeds the set threshold, the target correlated data objects have not yet resided in the high-performance storage layer, and the current system storage resources have spare capacity, the preloading operation of the predicted strongly correlated data objects is triggered.

[0083] This embodiment can use a deterministic strategy to directly select the action with the highest probability, or it can use an exploratory mechanism to randomly sample according to the probability distribution. All action selections must satisfy the storage level legality and system resource capacity constraints. Through the joint judgment of the above probability distribution interval and system state conditions, an interpretable mapping relationship between action selection and prediction results can be achieved.

[0084] When performing maintenance operations, there is no need to carry out inter-layer data migration. The hybrid storage execution module directly locates the storage location of the target data object and responds to the original, unobtrusive computational data access request. If the received request is a read request, the target access data is read from the current storage layer and returned. If the received request is a write request, the response result is returned after the data is written in the current storage layer.

[0085] When performing the upgrade migration to the dynamic random access memory layer, the hybrid storage execution module reads the corresponding data block from the source storage layer based on the data object identifier, writes the data block to the target dynamic random access memory layer, and synchronously updates the metadata to record the latest storage location of the data object. After the data migration is completed, it responds to the original access request and outputs the target access data.

[0086] When performing a downgrade migration to the persistent memory layer, the hybrid storage execution module reads the corresponding data block from the dynamic random access memory layer, writes the data to the persistent memory layer through an optimized persistent write method, updates the metadata records synchronously, and responds to the original access request after the migration is completed.

[0087] When performing migration operations between the persistent memory layer and the solid-state drive layer, the hybrid storage execution module coordinates the bidirectional migration of data between the two storage layers. For degradation operations from the persistent memory layer to the solid-state drive layer, it can trigger the background batch data degradation pipeline to perform asynchronous batch writing. After the migration is completed, the metadata records are updated and access requests are responded to.

[0088] When performing the operation of prefetching associated data, the hybrid storage execution module asynchronously extracts the target associated data object and preloads the target associated data object from its current storage location into the preset high-performance cache layer. This prefetching operation does not block the response flow of the current access request, and synchronously responds to the original access request and outputs the target access data.

[0089] It should be noted that during the aforementioned data migration or write operations, multiple storage media adaptation and optimization techniques can be applied to improve execution efficiency and reduce operational overhead. For read and write operations at the persistent memory layer, memory mapping can be used for direct access, and / or the Persistent Memory Development Kit (PMDK) can be used to perform atomic write operations. This kit can internally handle CPU cache line flushing and memory barrier logic, ensuring consistency and reliability during data persistence while avoiding the additional overhead caused by traditional file system input / output paths. For write operations at the solid-state drive (SSD) layer, batch and asynchronous processing mechanisms can be adopted to fully utilize the performance advantages of sequential writes on SSDs and amortize the overhead of writing small amounts of data at a time. At the same time, a three-layer cache structure consisting of transient cache, main cache, and persistent cache can be managed in a unified manner, executing cache data eviction and promotion logic to adapt to access needs at different time scales. After all data operations are completed, the actual overhead and performance impact data of this operation are recorded. For migration operations, the execution completion node is when the data can be accessed normally in the new location. For prefetch operations, the execution completion node is when the prefetch task is successfully started. The recorded overhead data includes migration time, bandwidth consumption, etc., and the performance impact data includes cache hit status, actual access latency, etc. The relevant data will be used as feedback information for subsequent policy iteration and optimization of the reinforcement learning decision engine.

[0090] Optionally, after obtaining the current system state information, and outputting a data management decision based on the current system state information, access pattern prediction results, and a preset reinforcement learning decision engine, and then performing a hybrid storage operation based on the data management decision and a preset heterogeneous storage structure, the method further includes: Obtain data on operation execution overhead and performance impact; The preset reinforcement learning decision engine is optimized and updated based on the operation execution overhead and performance impact data to obtain the target reinforcement learning decision engine, and the target reinforcement learning decision engine is used as the preset reinforcement learning decision engine.

[0091] In the specific implementation process, after completing the hybrid storage operation based on the data management decision, this embodiment also continuously optimizes and updates the reinforcement learning decision engine through a feedback mechanism. Specifically, it collects the execution overhead and performance impact data of this data operation. The execution overhead includes data migration time, storage bandwidth usage, input and output resource consumption, etc., while the performance impact data includes cache hit status, actual access latency, system real-time throughput, changes in the utilization rate of each storage layer, and the hit status of associated prefetched data, etc. Then, the above execution data is input into the reinforcement learning decision engine, and the actual reward value corresponding to this decision is calculated in combination with the preset reward function. The deviation between the actual value and the expected value of the current state is evaluated by the evaluator network, and then the weight parameters of the policy network are updated by backpropagation based on the near-end policy optimization algorithm to obtain the optimized target reinforcement learning decision engine. This optimized decision engine is then used as the preset reinforcement learning decision engine in the subsequent data management decision process, forming a complete closed-loop control structure of access monitoring, pattern prediction, intelligent decision-making, execution implementation, and feedback optimization.

[0092] It should be noted that the above optimization and updates can be carried out by combining online incremental updates with periodic batch training. During the off-peak period, centralized training is carried out based on accumulated historical execution data to ensure the stability of the strategy. During the operation of the business, small incremental updates are carried out to adapt to the dynamic changes in load, avoiding decision oscillation caused by frequent and large-scale adjustment of the strategy, and taking into account the adaptability of the system decision and the stability of operation.

[0093] Example 2: This embodiment provides a hybrid storage system for server-insensitive computing, used to implement the aforementioned hybrid storage method for server-insensitive computing. The system includes: The access monitoring module is used to acquire non-perceptible computing data access requests in real time and extract key access features based on the non-perceptible computing data access requests. The data extraction module is used to extract the data object identifier corresponding to the key access feature, and extract the access log based on the data object identifier; The access prediction module is used to perform short-term predictions on the access logs and key access features to obtain an access burst intensity score; perform long-term predictions on the access logs and key access features to obtain a set of survival probabilities based on popularity; obtain association prediction results based on the access logs, key access features, and a pre-built dynamic graph model; and form access pattern prediction results based on the access burst intensity score, the set of survival probabilities based on popularity, and the association prediction results. The hybrid storage execution module is used to acquire the current system status information, and output data management decisions based on the current system status information, access mode prediction results, and a preset reinforcement learning decision engine, so as to execute hybrid storage operations based on the data management decisions and the preset heterogeneous storage structure.

[0094] In specific implementation, this embodiment provides a hybrid storage system for server-aware computing. This system is used to implement the aforementioned hybrid storage method for server-aware computing. The overall system adopts a layered architecture design, mainly including an access monitoring module, a data extraction module, an access prediction module including a dynamic prediction engine, a hybrid storage execution module including a preset reinforcement learning decision engine and a multi-layer heterogeneous storage pool. It aims to solve the technical problems of performance bottlenecks, high costs, and difficulty in adapting to dynamic and complex access patterns in the intermediate data storage management of existing server-aware computing environments. While ensuring low data access latency and high throughput, it effectively controls storage costs and dynamically adapts to the diverse and time-varying data access characteristics of server-aware applications.

[0095] The multi-layer heterogeneous storage pool is the preset heterogeneous storage structure, which consists of at least two types of storage media with different performance and cost characteristics. In this embodiment, it preferably includes a high-speed volatile dynamic random access memory layer, a medium-speed persistent memory layer, and a low-speed large-capacity solid-state drive layer. By forming storage layers with performance and cost gradients, it supports dynamic migration and optimization management based on data access characteristics, and realizes a layered storage strategy for hot data and cold data.

[0096] The access monitoring module is used to intercept server-side functions that are unaware of access requests to intermediate data, acquire unaware computing data access requests in real time, and extract key access features. The extracted feature information includes, but is not limited to, data object identifier, access type, data size, access timestamp, source function identifier, and non-uniform memory access (NUMA) node information.

[0097] The data extraction module is used to extract data object identifiers corresponding to key access features, and retrieve the corresponding historical access logs based on the data object identifiers, providing a complete data foundation for subsequent access pattern prediction.

[0098] The access prediction module receives access features and access logs output by the data extraction module through a dynamic prediction engine. Based on machine learning or deep learning models, it predicts future access patterns for data objects. This engine includes a short-term burst detection module, a long-term pattern prediction module, and a correlation data prediction module. The short-term burst detection module performs short-term predictions on access logs and key access features to obtain an access burst intensity score, predicting the burst intensity of data access within the next few seconds. The long-term pattern prediction module performs long-term predictions on access logs and key access features to obtain a set of heat survival probabilities, predicting the access heat decay trend and lifespan of data objects over longer timescales such as minutes and hours. The correlation data prediction module obtains correlation prediction results based on access logs, key access features, and a pre-built dynamic graph model, predicting the likelihood of related access between different data objects. Finally, the dynamic prediction engine 102 integrates the access burst intensity score, the heat survival probability set, and the correlation prediction results to form a complete access pattern prediction result, which is then output to the reinforcement learning decision engine of the hybrid storage execution module.

[0099] The hybrid storage execution module receives access pattern prediction results from the reinforcement learning decision engine and combines them with the current system state information. It then uses a reinforcement learning algorithm to dynamically generate optimal data placement and migration decisions, specifying the target storage level or inter-level migration operation for the data object. The current system state information includes one or more of the following: storage utilization of each layer in the multi-layer heterogeneous storage pool, average system access latency, system tail latency, system throughput, and recent migration overhead. In this embodiment, the reinforcement learning algorithm preferably employs a near-end policy optimization algorithm. By limiting the magnitude of each policy update, the stability of the learning process is ensured, adapting to the complex decision-making requirements of the large-scale discrete state space and action space in this scenario.

[0100] The hybrid storage execution module is also responsible for executing data management decisions generated by the reinforcement learning decision engine, managing the placement, migration, caching, and eviction of data across different storage layers. To improve data operation efficiency and reduce the interference of migration overhead on front-end business, the hybrid storage execution module incorporates several optimization mechanisms, including a multi-level adaptive caching structure, specifically divided into transient cache, main cache, and persistent cache, each adapting to access needs at different time scales; a persistent memory direct access optimization mechanism, which reduces persistent memory access overhead through memory mapping and dedicated development toolkits; and a batch data degradation pipeline mechanism, which improves the execution efficiency of migrating cold data to slower storage layers through asynchronous batch writes in the background.

[0101] Optionally, it also includes a dynamic update module, which is used to: acquire operation execution overhead and performance impact data; optimize and update the preset reinforcement learning decision engine according to the operation execution overhead and performance impact data to obtain a target reinforcement learning decision engine, and use the target reinforcement learning decision engine as the preset reinforcement learning decision engine.

[0102] In the specific implementation process, the dynamic update module is used to obtain the execution overhead and performance impact data of data operations, and optimize and update the reinforcement learning decision engine based on the execution overhead and performance impact data. The optimized decision engine is then put into the subsequent decision-making process, forming a closed-loop control structure of prediction, decision-making, execution, and feedback, thereby realizing the continuous iterative improvement of the system's decision-making capabilities.

[0103] In its implementation, this embodiment also incorporates various storage access and data migration optimization mechanisms based on the hybrid storage execution module to improve data operation efficiency, fully leverage the hardware characteristics of each storage medium, and support the implementation of upper-layer intelligent decisions. For read / write operations on Persistent Memory (PMEM), the hybrid storage execution module employs a direct access optimization scheme, using memory mapping, development toolkit adaptation, access granularity optimization, and concurrency control to jointly ensure access performance and data reliability. Specifically, memory mapping directly maps persistent memory regions to the application's virtual address space via the memory map (mmap) system call, allowing direct access to persistent memory data via pointer read / write, bypassing traditional file system input / output paths and page caching, and significantly reducing data access latency. The Persistent Memory Development Kit (PMDK) adaptation utilizes its library functions, such as `pmemobj_tx_add_range` and `pmemobj_persist`, to perform atomic persistent write operations. The PMDK automatically handles CPU cache line flushing and memory barrier logic, ensuring data persistence and consistency even during system interruptions, while shielding the complexity of underlying hardware operations. Access granularity optimization leverages the internal access granularity characteristics of persistent memory, such as a 256-byte access granularity. During write operations, write operations are aligned to this granularity boundary, and multiple small-scale write operations are merged to reduce write amplification and improve persistent memory write efficiency. A concurrency control mechanism manages the number of concurrent write threads to persistent memory, preventing performance degradation caused by excessive concurrency and ensuring the stability of write operations.

[0104] For scenarios involving the migration of cold data from a high-performance storage layer to a low-speed storage layer, the hybrid storage execution module is equipped with a batch data degradation pipeline to efficiently complete the migration of large batches of cold data while minimizing the impact of data migration on front-end business access. When the reinforcement learning decision engine issues a batch degradation command, the hybrid storage execution module starts a background asynchronous task to group and aggregate the data blocks to be degraded. Once each batch of data reaches a preset batch threshold, such as an integer multiple of 4KB, it is written to the solid state drive (SSD) layer through an asynchronous input / output interface. The SSD write operation is decoupled from the front-end read / write operation, and they do not block each other. After each batch of data is written, the metadata information of the corresponding data block in the persistent memory layer is updated, marking the corresponding data as cold data or in a migrated state. This batch degradation mechanism can amortize the operation overhead of a single small data migration, fully utilize the performance advantages of sequential writes on SSDs, and effectively reduce the interference of data degradation operations on front-end business access latency.

[0105] It should be noted that the hybrid storage method provided in this embodiment can be implemented through computer program instructions. The computer program can be written in one or more programming languages, including object-oriented programming languages ​​and conventional procedural programming languages. The computer program code can be executed entirely on the user terminal device, partially on the user terminal device, executed as an independent software package, partially on the user terminal device and partially on a remote device, or executed entirely on a remote server device. In application scenarios involving remote devices, the remote device can be connected to the user terminal device through a local area network (LAN) or a wide area network (WAN), or it can establish a connection with external devices through a public network.

[0106] This embodiment provides a hybrid storage method and system for serverless computing, which relates to data storage management in cloud computing and serverless computing environments. In particular, it relates to a hybrid storage system and its data management method that utilizes multi-layer heterogeneous storage media and combines intelligent prediction and decision-making mechanisms. The aim is to solve the technical problem that intermediate data storage in serverless computing environments is difficult to balance performance and cost, and to balance the low latency and high throughput performance requirements of data access with the storage cost control requirements in serverless computing applications.

[0107] The hybrid storage method provided in this embodiment includes the following steps: monitoring data access requests and extracting access features; predicting future access patterns of data using a dynamic prediction engine, which includes a short-term burst detection module, a long-term pattern prediction module, and a correlation prediction module; generating optimal data placement or migration decisions using a reinforcement learning decision engine based on the prediction results and the real-time system status, where the reinforcement learning decision engine employs a near-end policy optimization algorithm; and executing corresponding decisions in a multi-layer heterogeneous storage pool through a hybrid storage execution module, wherein the multi-layer heterogeneous storage pool consists of three types of storage media with different performance and cost characteristics: high-speed dynamic random access memory, medium-speed persistent memory, and low-speed solid-state drives. Correspondingly, the hybrid storage system of this embodiment includes an access monitoring module, a dynamic prediction engine, a reinforcement learning decision engine, a hybrid storage execution module, and a multi-layer heterogeneous storage pool. Through intelligent prediction and decision-making, it dynamically optimizes the distribution of data across different storage layers, significantly reducing access latency, effectively controlling storage costs, improving system throughput and adaptability to dynamic loads, and achieving a high-efficiency balance between performance and cost.

[0108] Compared to existing technologies, this embodiment significantly reduces data access latency. By intelligently placing hot data and bursty access data in high-speed storage layers such as dynamic random access memory or persistent memory, and utilizing correlation prediction for data prefetching, it greatly reduces access to slow remote storage, thus lowering end-to-end access latency. This embodiment effectively controls storage costs by dynamically identifying cold data or long-lifecycle data and migrating it to lower-cost solid-state drive storage layers, avoiding the waste of high-performance storage resources. Compared to pure memory storage solutions, it significantly reduces overall storage costs. Through dynamic prediction and reinforcement learning decision-making, it enables intelligent data flow between different layers, allowing the system to achieve near-pure memory system performance while maintaining a cost far lower than pure memory systems, achieving a balance between high performance and low cost. Furthermore, this embodiment has strong environmental adaptability. The dynamic prediction engine and reinforcement learning decision engine can continuously learn and adapt to the complex and ever-changing dynamic access patterns of server-insensitive applications, automatically optimizing storage strategies without manual intervention. Optimized execution layer mechanisms, such as direct access to persistent memory and batch degradation, improve data operation efficiency. Combined with the parallel processing capabilities of multi-layered storage, the system possesses good throughput and scalability in high-concurrency scenarios. Furthermore, this embodiment is transparent to applications on the upper-layer server, allowing integration and use without modifying the application code.

[0109] It should be noted that existing related technical solutions typically use access prediction results directly for rule judgment or weight allocation, which belongs to the scheduling method based on thresholds or weighting functions, and does not involve the structured separation of prediction tasks at different time scales and deep coupling with the policy learning mechanism. This embodiment, on the one hand, independently models short-cycle burst behavior detection, long-term lifecycle survival probability estimation, and graph structure prediction based on association relationships, solving the technical problem of mutual interference between second-level access fluctuations and minute-level lifecycle decay; on the other hand, it constructs the above multi-dimensional prediction results into a unified state vector, inputs it into an actor-evaluator structured reinforcement learning policy network, and performs policy function search through a proximal constraint policy optimization algorithm, realizing data migration decisions across multiple heterogeneous physical storage layers such as dynamic random access memory, persistent memory, and solid-state drives, without relying on fixed thresholds or manual weight adjustment methods. Meanwhile, this embodiment introduces an instantaneous cache lifetime control mechanism, direct mapping access to persistent memory, and a batch asynchronous degradation pipeline mechanism at the execution layer to reduce migration oscillations and interference between front-end and back-end services, forming a closed-loop collaborative control structure for prediction, decision-making, and physical migration. The aforementioned multi-scale model partitioning method, state construction method, and policy learning-driven migration mechanism constitute the substantial technical features that distinguish this embodiment from the prior art.

[0110] Example 3: Based on the above embodiment of a hybrid storage method for server-insensitive computing, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a hybrid storage method for server-insensitive computing according to any embodiment of the present invention.

[0111] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0112] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0113] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0114] Example 4: Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute a hybrid storage method for server-insensitive computing as described in any of the above-described method embodiments of the present invention.

[0115] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0116] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A hybrid storage method for server-insensitive computing, characterized in that, include: Real-time acquisition of seamless computing data access requests, and extraction of key access features based on the seamless computing data access requests; Extract the data object identifier corresponding to the key access feature, and extract the access log based on the data object identifier; Short-term predictions are made based on the access logs and key access features to obtain an access burst intensity score; Long-term predictions are made based on the access logs and key access features to obtain a set of survival probabilities based on popularity. The association prediction results are obtained based on the access logs, key access features, and pre-built dynamic graph models. The access pattern prediction result is formed based on the access burst intensity score, the heat survival probability set, and the correlation prediction result. Obtain current system status information, and output data management decisions based on the current system status information, access mode prediction results, and a preset reinforcement learning decision engine, so as to perform hybrid storage operations based on the data management decisions and a preset heterogeneous storage structure.

2. The hybrid storage method for server-insensitive computing as described in claim 1, characterized in that, The step of performing short-term prediction on the access logs and key access features to obtain an access burst intensity score includes: Based on the access logs and key access features, short-term extraction is performed to obtain the access frequency sequence, read / write ratio sequence, and request time interval sequence per unit time. A short-time feature sequence is constructed based on the unit time access frequency sequence, read-write ratio sequence, and request time interval sequence. The short-time feature sequence is input into a pre-constructed temporal convolutional network for convolution calculation to obtain a high-dimensional temporal feature vector. Calculate the time decay weight based on the requested time interval sequence; The time decay feature statistics are obtained by weighted summation based on the time decay weight and the access frequency sequence per unit time. The high-dimensional temporal feature vector and the time decay feature statistics are fused and mapped to obtain the access burst intensity score.

3. The hybrid storage method for server-insensitive computing as described in claim 1, characterized in that, In obtaining the association prediction result based on the access logs, key access features, and the pre-built dynamic graph model, the pre-building process of the dynamic graph model includes: Historical co-occurrence access data is obtained based on the access logs and key access features; For the current data object node corresponding to the data object identifier, obtain the long-term feature sequence corresponding to the current data object node; Based on the access log, obtain several associated data object nodes corresponding to the current data object node, and obtain the long-term feature sequences of the associated nodes corresponding to the several associated data object nodes respectively; For any of the associated data object nodes, the access time interval is extracted based on the long-term feature sequence and the long-term feature sequence of the associated nodes, and the corrected weight is obtained based on the access time interval and the preset incremental decay function. Obtain several corrected weights corresponding to the aforementioned associated data object nodes; The pre-constructed initial association graph model is updated based on several modified weights to obtain a dynamic graph model.

4. The hybrid storage method for server-insensitive computing as described in claim 3, characterized in that, The association prediction results obtained based on the access logs, key access features, and pre-built dynamic graph model include: Based on the long-term feature sequence and several long-term feature sequences of the associated nodes, the association prediction result is obtained; The long-term feature sequence is modified based on the association prediction results and several modification weights to obtain a modified feature sequence, which is then used as the long-term feature sequence.

5. A hybrid storage method for server-insensitive computing as described in claim 1, characterized in that, The step of acquiring current system state information, and outputting data management decisions based on the current system state information, access pattern prediction results, and a preset reinforcement learning decision engine, to perform hybrid storage operations based on the data management decisions and a preset heterogeneous storage structure, includes: Obtain current system status information; Based on the current system status information and access mode prediction results, output the probability distribution of data management actions; The target data management action is determined based on the probability distribution of the data management actions. The target data management action and its corresponding probability value are determined based on the data management action probability distribution and preset selection conditions. Based on the target data management action, the system obtains operation judgment information from the current system status information, target action burst intensity score, target action heat survival probability, and target action association prediction result from the access mode prediction result; Based on the action probability value, target data management action, operation judgment information, target action burst intensity score, target action heat survival probability, target action association prediction result, and preset reinforcement learning decision engine, the data management decision is output. Hybrid storage operations are performed based on the data management decisions and the preset heterogeneous storage structure.

6. The hybrid storage method for server-insensitive computing as described in claim 5, characterized in that, The output data management action probability distribution based on the current system state information and access mode prediction results includes: The current system state information and access mode prediction results are structurally integrated to obtain a unified state vector; The unified state vector is normalized and encoded to obtain a standard state vector; The standard state vector is input into a preset multi-layer neural network for forward propagation calculation to obtain the candidate action score vector; The candidate action scoring vector is subjected to exponential normalization to obtain several probability values ​​corresponding to several preset candidate actions. The data outputs the probability distribution of the action based on several probability values.

7. A hybrid storage method for server-insensitive computing as described in claim 6, characterized in that, The process of performing hybrid storage operations based on the data management decisions and the preset heterogeneous storage structure includes: When the data management decision is the first action, the hybrid storage operation includes: responding to the non-intrusive computing data access request by outputting the target access data; When the data management decision is the second action, the hybrid storage operation includes: reading a data block based on the data object identifier, writing the data block into the target storage layer located in the preset heterogeneous storage structure, and responding to the non-perceptible computing data access request to output the target access data; When the data management decision is the third action, the hybrid storage operation includes: extracting the target associated data object, preloading the target associated data object into the preload cache layer in the preset heterogeneous storage structure, and responding to the non-perceptible computing data access request to output the target access data.

8. A hybrid storage method for server-insensitive computing as described in claim 1, characterized in that, After obtaining the current system state information, and outputting a data management decision based on the current system state information, access pattern prediction results, and a preset reinforcement learning decision engine, and then performing a hybrid storage operation based on the data management decision and a preset heterogeneous storage structure, the process further includes: Obtain data on operation execution overhead and performance impact; The preset reinforcement learning decision engine is optimized and updated based on the operation execution overhead and performance impact data to obtain the target reinforcement learning decision engine, and the target reinforcement learning decision engine is used as the preset reinforcement learning decision engine.

9. A hybrid storage system for server-insensitive computing, characterized in that, For implementing a hybrid storage method for server-insensitive computing as described in any one of claims 1 to 8, the system comprises: The access monitoring module is used to acquire non-perceptible computing data access requests in real time and extract key access features based on the non-perceptible computing data access requests. The data extraction module is used to extract the data object identifier corresponding to the key access feature, and extract the access log based on the data object identifier; The access prediction module is used to perform short-term predictions on the access logs and key access features to obtain an access burst intensity score; perform long-term predictions on the access logs and key access features to obtain a set of survival probabilities based on popularity; obtain association prediction results based on the access logs, key access features, and a pre-built dynamic graph model; and form access pattern prediction results based on the access burst intensity score, the set of survival probabilities based on popularity, and the association prediction results. The hybrid storage execution module is used to acquire the current system status information, and output data management decisions based on the current system status information, access mode prediction results, and a preset reinforcement learning decision engine, so as to execute hybrid storage operations based on the data management decisions and the preset heterogeneous storage structure.

10. A hybrid storage system for server-insensitive computing as described in claim 9, characterized in that, It also includes a dynamic update module, which is used to: acquire operation execution overhead and performance impact data; optimize and update the preset reinforcement learning decision engine according to the operation execution overhead and performance impact data to obtain a target reinforcement learning decision engine, and use the target reinforcement learning decision engine as the preset reinforcement learning decision engine.

Images