A method and system for resource allocation pre-check and dynamic compensation of dual-state storage cooperation

Abstract

The application discloses a kind of dual-state storage collaborative resource allocation pre-inspection and dynamic filling method and system.Method includes: for each allocable resource, construct dual-state feature model containing steady-state feature and transient feature, transient feature is collected with first frequency and steady-state feature is updated with second frequency by probe;When receiving resource allocation request, the dual-state feature of candidate resource is mapped to different quadrant based on collaborative determination matrix, and target resource is selected according to quadrant priority;After resource allocation, transient feature is predicted using LSTM trend prediction model, and if it exceeds safety threshold, it is marked as sub-health state and generates filling instruction;Tasks on sub-health resources are divided into hot data tasks and warm data tasks, and stable area resources are selected as shadow resources, hot data tasks are transparently transferred, and warm data tasks are gradually migrated.The application realizes active risk avoidance and system high availability of resource allocation.

Description

Technical Field

[0001] This invention relates to the field of resource allocation and scheduling technology, and more specifically, to a method and system for resource allocation pre-detection and dynamic replacement with dual-state storage collaboration. Background Technology

[0002] With the rapid development of distributed systems and cloud computing, resource pooling and dynamic scheduling technologies have become key means to improve system resource utilization and service quality. In distributed storage systems, resource management servers are typically used to uniformly manage and schedule massive numbers of storage nodes. When a resource allocation request is received from a business system, a suitable storage node is selected from the resource pool for task allocation. To ensure the accuracy of resource allocation and the high availability of the system, existing technologies typically employ a threshold-based resource monitoring mechanism. That is, when a certain performance indicator of a resource exceeds a preset threshold, the resource is marked as unavailable or degraded, thereby avoiding the allocation of new tasks to abnormal resources.

[0003] However, existing technologies have the following technical shortcomings in practical applications:

[0004] First, existing resource monitoring mechanisms are primarily based on single-dimensional static threshold judgments, which cannot comprehensively reflect the true state of resources. For example, focusing only on the current load rate while ignoring long-term attributes such as hardware configuration and historical reliability, or focusing only on hardware specifications while ignoring real-time performance fluctuations, leads to biased resource assessment results. This can easily result in misjudging highly configured but momentarily congested resources as unavailable or prioritizing the allocation of low-configuration but currently idle resources, affecting the accuracy of resource allocation and overall system performance. Second, existing technologies typically employ a reactive response mechanism, migrating tasks only after resource failure or severe performance degradation. This passive approach often leads to business interruptions and sudden increases in data access latency. Especially in scenarios with slowly declining resource performance, existing technologies lack the ability to anticipate resource degradation trends and cannot provide preventative intervention before failures occur. Third, when task migration is required, existing technologies typically employ a uniform migration strategy, treating all tasks equally. This coarse-grained processing approach cannot distinguish the latency sensitivity of different tasks, which may cause interruptions in real-time transaction tasks that are highly sensitive to latency during the migration process, affecting the continuity of core business; while for batch processing tasks that can tolerate short-term latency, an overly aggressive migration strategy may lead to unnecessary resource consumption. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a dual-state storage collaborative resource allocation pre-detection and dynamic replacement method. This method achieves comprehensive perception of resource status by constructing a dual-state feature model, performs refined pre-detection allocation based on a collaborative judgment matrix, identifies sub-optimal resources in advance using a trend prediction model, and performs differentiated smooth replacement based on task sensitivity levels. This solves the technical problems of one-sided resource assessment, lack of trend prediction, and coarse migration strategies in existing technologies.

[0006] To achieve the above objectives, the present invention adopts the following technical solution;

[0007] A dual-state storage collaborative resource allocation pre-detection and dynamic compensation method is applied to a resource allocation system, which includes a resource management server and probes deployed at each allocatable resource endpoint, comprising the following steps:

[0008] The S100 resource management server constructs a dual-state feature model containing steady-state and transient features for each allocable resource, and collects and maintains the latest dual-state feature data of each resource in real time through probes deployed at each resource end;

[0009] S200. When a resource allocation request containing service quality requirements is received, the resource management server reads the current dual-state characteristics of each candidate resource from the candidate resource pool based on the latest dual-state feature data; then it calls the pre-stored collaborative decision matrix to map the dual-state characteristics of each candidate resource to the collaborative decision matrix to determine its quadrant, and selects the target resource that meets the service quality requirements from the candidate resource pool according to the preset quadrant priority order.

[0010] S300. After resource allocation, the resource management server uses a trend prediction model to predict the transient characteristic data of the allocated resources. If the prediction result exceeds the safety threshold, the resource is marked as sub-healthy and a dynamic replacement instruction is generated.

[0011] The S400 resource management server divides tasks running on sub-healthy resources into hot data tasks and warm data tasks. It selects resources in the optimal and stable zone from the candidate resource pool as shadow resources. Hot data tasks are transparently transferred through underlying network technology, while warm data tasks are migrated gradually until the task migration is complete.

[0012] As a further aspect of the present invention, S100 includes:

[0013] The resource management server loads the metadata definition file of the dual-state feature model. The metadata definition file specifies the steady-state feature field group and the transient feature field group. The steady-state feature field group includes at least one of the following: hardware model, total capacity, and nominal performance parameters. The transient feature field group includes at least one of the following: number of real-time input / output operations per second, current load rate, and read / write response latency.

[0014] The resource management server identifies allocable resources through a resource discovery mechanism and creates an independent model instance for each discovered resource.

[0015] The resource management server sends query commands to each resource through the standard hardware management interface protocol to obtain its steady-state characteristic data, and fills the obtained steady-state characteristic data into the steady-state characteristic field of the corresponding model instance to realize the initialization assignment of the model instance;

[0016] The resource management server selects a compatible probe version based on the hardware model and operating system type of each resource and distributes it to each allocable resource. It also sends a collection strategy configuration file to the probe, specifying that transient feature data is collected at a first frequency of hundreds of milliseconds and steady-state feature update data is collected at a second frequency of minutes.

[0017] The resource management server receives transient feature data reported by the probe, performs smoothing and denoising processing using an exponentially weighted moving average algorithm, and then updates the corresponding model instance in the high-speed memory database using an overwrite write method. At the same time, it receives steady-state feature update data reported by the probe, updates the corresponding steady-state feature fields in the model instance, and persists them to the backend relational database.

[0018] As a further aspect of the present invention, in the exponentially weighted moving average algorithm, the smoothed value at the current moment is obtained by weighting the measured value at the current moment and the smoothed value at the previous moment using a smoothing coefficient; the smoothing coefficient ranges from 0 to 1, and the closer the smoothing coefficient is to 1, the greater the weight of the current measured value, and the closer the smoothing coefficient is to 0, the greater the weight of the historical value.

[0019] As a further aspect of the present invention, S200 specifically includes:

[0020] After receiving a resource allocation request that includes quality of service requirements, the resource management server queries the dual-state feature database to form a candidate resource pool based on the resource type and basic capacity conditions specified in the request.

[0021] The current dual-state feature data of each resource in the candidate resource pool is read in batches from the high-speed memory database. The current dual-state feature data is the latest data maintained in real time by S100.

[0022] The read bi-state feature data is normalized to map feature values ​​of different dimensions to a standard interval;

[0023] The pre-stored collaborative decision matrix is ​​invoked. The collaborative decision matrix is ​​a two-dimensional coordinate system constructed with the normalized steady-state feature score as the horizontal axis and the transient feature score as the vertical axis. The two-dimensional coordinate system is divided into four quadrants by the preset steady-state feature threshold and transient feature threshold.

[0024] The normalized bi-state features of each candidate resource are used as coordinate points and mapped to the collaborative decision matrix to determine its quadrant.

[0025] Candidate resources are sorted according to a preset quadrant priority order, and the resource with the highest priority is selected as the target resource for this allocation.

[0026] As a further aspect of the present invention, the four quadrants include:

[0027] The first quadrant has both steady-state and transient characteristics that are higher than the corresponding thresholds, i.e., resources in the optimal and stable region;

[0028] The second quadrant, representing the potential resource area, consists of regions where steady-state characteristics are below a threshold but transient characteristics are above a threshold.

[0029] Resources in the third quadrant, where steady-state characteristics are above the threshold and transient characteristics are below the threshold, are considered risk zone resources.

[0030] Resources with both bi-state characteristics below the corresponding threshold in the fourth quadrant are considered to be in the elimination zone.

[0031] As a further aspect of the present invention, S300 includes:

[0032] After resource allocation, the resource management server will include the allocated resources in the continuous monitoring list, and its historical transient characteristic data will be organized into time series and stored in the time series database.

[0033] A pre-set trend prediction model based on a long short-term memory neural network is invoked, and key transient feature data of the resource at the most recent preset number of time points are input into the model at a preset frequency to obtain a sequence of predicted values ​​at the next preset number of time points.

[0034] The predicted value sequence is compared point by point with the preset safety threshold. If the prediction result shows that a certain key transient feature of the resource will reach or exceed the safety threshold within the preset time window in the future, the resource is judged to be in a sub-healthy state. Resources judged to be in a sub-healthy state will continue to be monitored, but will no longer be considered as candidate resources for new allocation.

[0035] Generate dynamic fill-in instructions, which include a unique identifier for the sub-health resource, a list of tasks currently running on the resource, and the sensitivity pre-grading results of each task.

[0036] As a further aspect of the present invention, the key transient characteristics include device temperature, read / write response latency, and the number of input / output operations per second.

[0037] As a further aspect of the present invention, S400 includes:

[0038] The resource management server parses the dynamic fill instruction to obtain the unique identifier of sub-health resources, the task list, and the sensitivity pre-classification results of each task;

[0039] Based on their sensitivity to changes in resource status, tasks are divided into hot data tasks and warm data tasks. Hot data tasks are those that are highly sensitive to latency and use long-lived connection communication; warm data tasks are those that have a higher tolerance for latency and use short-lived connection communication or belong to batch processing tasks.

[0040] Based on the collaborative decision matrix, all resources in the optimal and stable zone are selected from the current candidate resource pool as candidate shadow resources;

[0041] From the candidate shadow resource pool, taking into account factors such as current load rate, remaining capacity, and network distance from sub-healthy resources, the resource with the best overall conditions is selected as the target shadow resource.

[0042] For hot data tasks, flow table rules are issued by the software-defined network controller to transparently transfer business traffic that originally pointed to sub-healthy resources to the target shadow resources;

[0043] For warm data tasks, a gradual migration strategy is adopted. The resource configuration table is updated in the task scheduler, and newly initiated warm data tasks are allocated to the target shadow resource. Warm data tasks that are currently being executed on sub-healthy resources are allowed to continue running until they end naturally.

[0044] Once all tasks on the sub-health resource platform have been migrated, the record replacement is complete.

[0045] As a further aspect of the present invention, the method for selecting the resource with the best comprehensive conditions as the target shadow resource includes: using a multi-factor weighted scoring method to calculate a comprehensive score based on the current load rate, remaining capacity ratio, and network distance to the sub-healthy resource of the candidate shadow resource, and selecting the resource with the highest comprehensive score as the target shadow resource; wherein, the current load rate has the highest weight in the scoring.

[0046] The second aspect of this application discloses a dual-state storage collaborative resource allocation pre-detection and dynamic compensation system, the system including a resource management server and probes deployed at each allocatable resource end;

[0047] The resource management server includes:

[0048] The dual-state feature modeling module is used to load the metadata definition file of the dual-state feature model, create an independent model instance for each allocable resource, and obtain steady-state feature data through the standard hardware management interface to initialize the model instance.

[0049] The probe management module is used to deploy probes to each allocable resource terminal and issue acquisition strategies that include the acquisition frequency of dual-state features.

[0050] The data receiving and preprocessing module is used to receive transient feature data and steady-state feature update data reported by the probe, and to perform smoothing and noise reduction processing on the transient feature data.

[0051] The dual-state feature database includes a high-speed in-memory database for storing real-time transient feature data and a back-end relational database for persistently storing steady-state feature data.

[0052] The resource allocation pre-detection module is used to receive resource allocation requests containing service quality requirements, and to determine the target resource based on the latest bi-state feature data maintained in the high-speed memory database by calling the collaborative decision matrix and quadrant priority.

[0053] The trend prediction module is used to load a trend prediction model based on a long short-term memory neural network, identify the sub-health status of allocated resources, and trigger replacement commands.

[0054] The dynamic fill module is used to respond to fill commands and perform differentiated smooth migration of hot data tasks and warm data tasks according to the task sensitivity level.

[0055] The probes deployed at each allocable resource terminal are used to collect and report transient characteristic data and steady-state characteristic update data of the resources at different frequencies according to the collection strategy.

[0056] In summary, due to the adoption of the above technical solution, the beneficial technical effects of the invention are as follows:

[0057] First, by constructing a dual-state feature model and continuously collecting two types of feature data at different frequencies, and by comprehensively mapping and prioritizing the dual-state features through a collaborative decision matrix, resource allocation decisions can be based on both the long-term stable attributes and real-time dynamic performance of resources. This avoids assigning tasks to resources in poor condition from the source during the allocation process, directly improving the accuracy of resource allocation and the overall performance of the system.

[0058] Secondly, by continuously using trend prediction models to predict transient characteristics after resource allocation and triggering dynamic replacement commands when a sub-healthy state is identified, the system can intervene in advance when resource performance deteriorates but has not yet reached the fault threshold, thus avoiding business interruption caused by sudden failures and realizing the transformation from passive response to proactive prevention.

[0059] Furthermore, by implementing differentiated migration strategies for hot data tasks and warm data tasks based on task sensitivity during the replacement process, transparent transfer is carried out for latency-sensitive hot data tasks, while gradual migration is carried out for warm data tasks with tolerable latency. This ensures that the entire replacement process can be completed while guaranteeing the continuity of core business, thus ensuring a seamless switch of critical business and avoiding resource waste caused by excessive intervention in non-critical tasks.

[0060] As can be seen, this invention, through a complete technical chain of comprehensive perception → pre-detection decision → trend prediction → differentiated supplementation, has directly improved the accuracy of resource allocation, the timeliness of risk identification, and the smoothness of business migration in multiple dimensions, forming a logically coherent and synergistic complete technical solution that effectively overcomes many defects in the existing technology. Attached Figure Description

[0061] Figure 1 A flowchart illustrating a dual-state storage collaborative resource allocation pre-check and dynamic fill-in method;

[0062] Figure 2 A flowchart of a dual-state storage collaborative resource allocation pre-detection and dynamic compensation method S100;

[0063] Figure 3 A flowchart of a dual-state storage collaborative resource allocation pre-detection and dynamic compensation method S200;

[0064] Figure 4 S300 flowchart of a resource allocation pre-detection and dynamic compensation method for dual-state storage collaboration;

[0065] Figure 5 S400 flowchart is a resource allocation pre-check and dynamic fill method for dual-state storage collaboration. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0068] The core objective of this invention is to provide a resource allocation pre-detection and dynamic compensation method based on dual-state storage collaboration. By performing dual-state feature modeling, pre-detection decision-making, trend prediction, and smooth compensation on resources, proactive risk avoidance and high system availability can be achieved in resource allocation.

[0069] The core idea of ​​this method is to combine the long-term stable attributes (steady-state characteristics) of resources with short-term dynamic changes (transient characteristics) and perform refined resource matching through a collaborative judgment matrix. At the same time, it identifies resource degradation risks in advance through trend prediction and performs differentiated smooth replacement based on task sensitivity levels, thereby constructing a proactive resource management system that spans the entire chain from perception and decision-making to execution.

[0070] like Figure 1 As shown, this application illustrates an exemplary method for resource allocation pre-detection and dynamic compensation in dual-state storage collaboration, specifically including the following steps:

[0071] S100. Construction and Real-time Perception of Dual-State Feature Model;

[0072] S200. Resource allocation pre-detection based on collaborative decision-making;

[0073] S300. Sub-health status identification and replacement triggering based on trend prediction;

[0074] S400. Smooth dynamic fill based on task sensitivity classification.

[0075] Please refer to Figure 2 The flowchart illustrates an exemplary dual-state storage collaborative resource allocation pre-detection and dynamic filling method S100 of this application;

[0076] In a dual-state storage collaborative resource allocation pre-detection and dynamic replacement method, the purpose of S100 is to establish a dual-state feature model that can comprehensively describe the state of each allocatable resource in the system, and to realize high-frequency acquisition of transient features and low-frequency update of steady-state features, so as to provide a data basis for subsequent decision-making.

[0077] The specific steps include:

[0078] S110. Resource Discovery and Initialization:

[0079] After the resource management server starts up, it discovers and identifies all manageable resources in the system by reading a pre-configured resource list file or by using a network broadcast probe protocol. Allocable resources include storage nodes, compute nodes, or network devices. For each discovered resource, the server creates a unique resource identifier entry in the internally maintained resource management table and allocates corresponding memory storage space for storing the resource's bi-state characteristic data.

[0080] S120. Prototype definition of the bi-state feature model:

[0081] The resource management server loads the metadata definition file of the bi-state feature model from the system configuration center. The metadata definition file clearly defines the two main categories of feature fields included in the bi-state feature model and their data structure requirements:

[0082] The first category is the steady-state characteristic field group, which defines the long-term stable attributes of the resource, including hardware model, total capacity, nominal performance parameters, historical average failure rate, and physical location information. The steady-state characteristic fields adopt a low-frequency update strategy.

[0083] The second category is the transient feature field group, which defines the real-time dynamic indicators of resources, including the number of real-time input / output operations per second, current load rate, temperature, read / write response latency, and remaining lifetime percentage. The transient feature field group adopts a high-frequency acquisition strategy.

[0084] S130. Model instantiation and initialization assignment:

[0085] Based on the dual-state feature prototype defined in step S120, the resource management server creates an independent model instance for each allocatable resource discovered in step S110. The model instance is logically represented as a set of state records that uniquely corresponds to the resource.

[0086] The resource management server sends query commands to each resource through the standard hardware management interface protocol to obtain its steady-state characteristic data. The steady-state characteristic data includes the CPU model, total memory capacity, total disk capacity, network interface bandwidth, device firmware version, and rack location information. The steady-state characteristic data is then filled into the steady-state characteristic field of the corresponding model instance.

[0087] For transient feature fields, they are initialized to empty values ​​or system default values ​​at this time, and are filled with real-time data reported by subsequent probes.

[0088] S140. Probe Adaptation and Deployment:

[0089] Based on the hardware model and operating system type of each resource obtained by S130, the resource management server selects a compatible probe version from the pre-set probe program library and distributes the probe program to each allocable resource terminal through a secure file transfer protocol.

[0090] The probe program is a lightweight background daemon process designed to consume less than one percent of system resources. It can directly access and obtain real-time running status data of resources by calling the underlying interfaces of the operating system and hardware drivers.

[0091] S150. Probe activation and dual-frequency acquisition configuration:

[0092] Each allocable resource terminal automatically runs the probe program after receiving it. The probe establishes an encrypted communication link with the resource management server and reports its own running status.

[0093] The resource management server sends a personalized data collection strategy configuration file to each probe via a communication link, specifying the following parameters:

[0094] The first frequency, namely the transient feature acquisition frequency, is set to the level of hundreds of milliseconds. The probe needs to continuously acquire and report transient feature data such as the number of input / output operations per second, read / write response latency, device temperature, and central processing unit utilization at this interval.

[0095] The second frequency, namely the steady-state feature update frequency, is set to the minute level. The probe needs to check and report the steady-state features that may change slowly at this interval, including remaining storage capacity, remaining lifetime percentage, and firmware version.

[0096] Data reporting methods include: transient feature data is reported at high speed and with low latency using the User Datagram Protocol (UDP), and steady-state feature update data is reported reliably using the Transmission Control Protocol (TCP).

[0097] S160. Data Smoothing and Preprocessing:

[0098] After receiving the transient feature data reported by the probe, the resource management server uses an exponentially weighted moving average algorithm to smooth and denoise the data, removing transient spikes and interference.

[0099] The formula for calculating the exponentially weighted moving average algorithm is as follows:

[0100] ;

[0101] For the present The smoothing value at any given time; For the present The measured value at that moment; For the previous moment The smoothed value; This is the smoothing coefficient, and its value range is... ;

[0102] Smoothing coefficient The closer the value is to 1, the greater its weight, resulting in a faster response but a weaker smoothing effect; smoothing coefficient The closer to 0, the greater the weight of historical values, resulting in a stronger smoothing effect but a slower response speed.

[0103] To balance sensitivity and stability, the smoothing coefficient Set to 0.2. The smoothing coefficient. Adjustments can be made based on the fluctuation characteristics of different resource features: for resource features with large fluctuations, the amount can be appropriately reduced. To enhance the smoothing effect; for resource features that require rapid response to mutations, the value can be appropriately increased. To improve response speed.

[0104] S170. Update of the dual-state feature database:

[0105] For the smoothed transient feature data, the resource management server locates the corresponding model instance based on the resource identifier and updates the corresponding transient feature field in the instance by overwriting; to ensure millisecond-level access latency, the transient feature field data is directly stored in a high-speed memory database directly connected to the server and is not persistently written.

[0106] For low-frequency steady-state feature update data streams, the probe checks the fields in the steady-state features that may change once per minute; if a change is detected, it is reliably reported through the transmission control protocol; after receiving the data, the server updates the corresponding steady-state feature fields in the model instance and persists the updated data to the backend relational database for long-term querying and historical trend analysis.

[0107] S180. High-speed storage and access interface:

[0108] After all bi-state feature data is updated, it is stored in a high-speed in-memory database in the form of key-value pairs. The key is a combination of resource identifier and feature field name, and the value is the current feature data. The storage structure ensures that subsequent steps have microsecond-level access latency when reading feature data, which meets the performance requirements of real-time decision-making.

[0109] Through the above steps, the resource management server successfully builds and continuously maintains a dual-state feature model for each allocatable resource. This model reflects the instantaneous operating state of the resource in high-speed memory in real time, while its stable attribute information is persistently stored in the backend database, providing a data input basis for subsequent resource allocation pre-detection based on collaborative judgment.

[0110] Please refer to Figure 3 The flowchart illustrates an exemplary dual-state storage collaborative resource allocation pre-detection and dynamic filling method S200 of this application;

[0111] In a dual-state storage collaborative resource allocation pre-detection and dynamic replacement method, the purpose of S200 is to perform pre-detection based on the dual-state characteristics of the resource when the system receives a new resource allocation request, and to finely screen candidate resources through a collaborative decision matrix to avoid assigning tasks to resources that are about to have performance problems.

[0112] The specific steps include:

[0113] S210. Receive resource allocation request:

[0114] The resource management server receives a resource allocation request from the business system; the resource allocation request includes constraints such as the required resource capacity, the minimum number of input / output operations per second, the maximum allowed read / write response latency, and the required number of CPU cores.

[0115] S220. Construct a candidate resource pool:

[0116] The resource management server queries the dual-state feature database for all resources that meet the basic capacity conditions according to the resource type specified in the resource allocation request, forming a candidate resource pool. The query conditions include: remaining storage capacity is greater than the requested capacity, online status is normal, and not marked as sub-healthy or offline.

[0117] S230. Batch reading of bi-state feature data:

[0118] The resource management server reads the current dual-state characteristic data of each resource in the candidate resource pool in batches from the high-speed memory database. The data read includes the remaining capacity ratio and hardware performance level in the steady-state characteristics, and the current number of input / output operations per second, current read / write response latency, current load rate, and device temperature in the transient characteristics.

[0119] S240. Feature data normalization processing:

[0120] The resource management server normalizes the read dual-state feature data, mapping feature values ​​of different dimensions to a standard range of 0 to 1. The steady-state feature normalization is based on the ratio of the current remaining capacity to the total capacity; the transient feature normalization is based on the ratio of the current measured value to the nominal maximum value of the device; and for inverse indicators such as latency, the conversion is performed by subtracting the ratio of the current value to the maximum allowable value from 1.

[0121] S250. Call the cooperative decision matrix:

[0122] The resource management server calls the pre-stored collaborative decision matrix. The collaborative decision matrix is ​​a two-dimensional coordinate system, with the horizontal axis representing the normalized steady-state feature score and the vertical axis representing the normalized transient feature score. The matrix space is divided into four quadrants: the first quadrant corresponds to high steady-state and high transient, the second quadrant corresponds to low steady-state and high transient, the third quadrant corresponds to high steady-state and low transient, and the fourth quadrant corresponds to low steady-state and low transient.

[0123] In the specific implementation process:

[0124] Any candidate resource The normalized steady-state characteristic score is denoted as Normalized transient feature scores are denoted as Then the position coordinates of the resource in the collaborative decision matrix. Represented as:

[0125] ;

[0126] Let the steady-state feature segmentation threshold be... The threshold for transient feature segmentation is The rule for determining the quadrant to which a resource belongs is as follows:

[0127] ;

[0128] S260. Resource Quadrant Mapping and Determination:

[0129] The resource management server uses the normalized bi-state features of each candidate resource as coordinate points, maps them to the collaborative decision matrix, and determines the quadrant to which the resource belongs based on the position of the coordinate points.

[0130] The quadrant division threshold of the collaborative judgment matrix is ​​determined through statistical analysis of historical operating data. Specifically, the resource management server collects steady-state and transient characteristic scores of all resources within historical operating cycles to form a sample set. The K-means clustering algorithm is used to divide the sample points into four categories, corresponding to high steady-state and high transient, low steady-state and high transient, high steady-state and low transient, and low steady-state and low transient, respectively. The cluster centers of each category are calculated, and the boundary values ​​of adjacent cluster centers are used as the quadrant division threshold.

[0131] Based on cluster analysis, the threshold for dividing steady-state features was calculated. The threshold value was set at 0.6 for transient feature segmentation. The threshold is set at 0.6. Based on the above threshold setting, the resource quadrant determination rule is as follows:

[0132] If the resource falls in the first quadrant, then it satisfies... and If so, it is determined to be a resource in a superior and stable area;

[0133] If the resource falls in the second quadrant, then it satisfies... and If so, it is determined to be a resource in a potential area;

[0134] If the resource falls in the third quadrant, then it satisfies... and If so, it is determined to be a resource in a risk area;

[0135] If the resource falls in the fourth quadrant, then it satisfies... and If so, it is determined to be a resource in the elimination zone.

[0136] S270. Sort by priority:

[0137] The resource management server sorts all candidate resources according to a preset quadrant priority order. The quadrant priority order is as follows: the stable zone has the highest priority, followed by the potential zone, then the risk zone, and the elimination zone has the lowest priority. Resources within the same quadrant are sorted from high to low according to their comprehensive score, which is calculated by weighting the steady-state feature score and the transient feature score.

[0138] S280. Determine the resource allocation strategy:

[0139] The resource management server selects the highest priority resource from the sorted candidate resource list as the target resource for this allocation. If there are enough resources in the priority and stable zone, it will prioritize random selection or allocation in a round-robin manner. If there are not enough resources in the priority and stable zone, it will select resources from the potential zone for allocation and add a low-priority task tag to the task during allocation for subsequent differentiated monitoring and early warning.

[0140] S290. Recording and Feedback of Allocation Results:

[0141] The resource management server records the target resource identifier, allocation time, and requested service quality parameters for this allocation, and feeds the allocation result back to the business system. Simultaneously, it increments the current load count of the resource by 1 for subsequent load balancing adjustments.

[0142] Through the above steps, the resource management server implements resource allocation pre-detection based on dual-state characteristics, ensuring that high-priority tasks are allocated to resources in the best state first, thereby reducing the risk of task failure due to resource performance degradation from the source.

[0143] Please refer to Figure 4 The flowchart illustrates an exemplary dual-state storage collaborative resource allocation pre-detection and dynamic filling method S300 of this application;

[0144] In a dual-state storage collaborative resource allocation pre-detection and dynamic replacement method, the purpose of S300 is to continuously monitor and predict the operating status of allocated resources, identify sub-healthy resources that are about to fail in advance, and trigger a preventive replacement process.

[0145] The specific steps include:

[0146] S310. Establish a continuous monitoring list:

[0147] After resource allocation is completed, the resource management server adds the allocated resources to the continuous monitoring list; for each resource in the monitoring list, its historical transient characteristic data is organized into a time series and stored in the time series database according to the resource identifier; each time series records the transient characteristic sampling data within the most recent preset duration, and the sampling data includes timestamps and corresponding characteristic values.

[0148] S320. Load the trend prediction model:

[0149] The resource management server calls a pre-built trend prediction model, which is a simplified version of a long short-term memory neural network model. This model has been trained offline using historical operating data and can learn typical patterns of performance degradation of various resources.

[0150] The mathematical expression of the model is as follows:

[0151] ;

[0152] ;

[0153] For LSTM networks in The hidden state at any given moment; for The input features at time t represent the transient feature data input into the LSTM network at time t, specifically including key transient features such as device temperature, read / write response latency, and the number of input / output operations per second; for The hidden state of the previous moment; Indicates the future The predicted value at any given time; The prediction step size represents the predicted future time point index, with a value ranging from 1 to... ; To output the predicted sequence length; and Here are the weight matrix and bias terms of the fully connected layer; the values ​​of these terms can be adjusted according to the rate of change of resource characteristics and the system response time requirements: for resource performance indicators that change rapidly, they can be appropriately reduced to improve the response speed; for scenarios that require earlier warnings, they can be appropriately increased to extend the prediction time window.

[0154] S330. Periodic Feature Prediction:

[0155] For each allocated resource in the monitoring list, the resource management server updates the most recently allocated resource at a preset frequency. The key transient feature data at each time point is input into the trend prediction model; the key transient features include device temperature, read / write response latency, and the number of input / output operations per second, which can most directly reflect the changing trend of the health status of resources.

[0156] S340. Comparison of prediction results with threshold:

[0157] The resource management server will output the future trend prediction model. The predicted value sequence at each time point is compared point by point with the preset security thresholds. The security thresholds are dynamically set according to the hardware specifications of the resources and the service level agreement, and specifically include: device temperature threshold, read / write response latency threshold, and minimum number of input / output operations per second.

[0158] S350. Determination of Sub-health Status:

[0159] If the trend prediction results indicate that a key transient characteristic of a resource will reach or exceed a safety threshold within a preset future time window, the resource management server determines that the resource faces an impending performance risk. The risk determination criteria are: the predicted values ​​at any consecutive time points in the predicted value sequence all exceed the threshold, or the predicted value at a single time point exceeds the threshold by a margin greater than a preset percentage.

[0160] S360. Status flag update:

[0161] The resource management server immediately updates the marker of resources that meet the sub-health criteria in the system status table from healthy to sub-healthy, and synchronizes the updated status to the corresponding field in the dual-state feature database; resources marked as sub-healthy will continue to be monitored, but will no longer participate in new allocation as candidate resources.

[0162] S370. Generate dynamic fill instruction:

[0163] The resource management server automatically generates a dynamic fill-in instruction based on the identifier of the sub-healthy resource. The dynamic fill-in instruction includes the following information: the unique identifier of the sub-healthy resource, the list of tasks currently running on the resource, the sensitivity pre-grading results of each task, and the expected performance degradation time given by the trend prediction model.

[0164] S380. Instruction Distribution and Process Triggering:

[0165] The resource management server sends the dynamic fill instruction to the fill execution module and records the event in the system log. After receiving the instruction, the fill execution module immediately starts the smooth dynamic fill process of step S400.

[0166] Through the above steps, the resource management server achieves early detection of resource degradation trends and accurate identification of sub-health states, providing a basis for proactive risk avoidance and preventing business interruptions caused by sudden failures.

[0167] Please refer to Figure 5 The flowchart illustrates an exemplary method for resource allocation pre-detection and dynamic filling in of dual-state storage collaboration according to this application.

[0168] In a dual-state storage collaborative resource allocation pre-detection and dynamic compensation method, the purpose of S400 is to respond to the dynamic compensation instruction and smoothly migrate tasks on sub-healthy resources to healthy shadow resources in a way that is imperceptible or minimally imperceptible to the business, thereby ensuring the continuity and stability of system services.

[0169] The specific steps include:

[0170] S410. Parsing the padding instruction:

[0171] The resource management server receives and parses the dynamic fill instruction to obtain the unique identifier of the sub-health resource, the list of currently running tasks on the resource, and the sensitivity pre-classification results of each task;

[0172] S420. Task Sensitivity Classification:

[0173] The resource management server calls the task classification module to divide tasks into hot data tasks and warm data tasks based on their sensitivity to changes in resource status; the classification criteria include:

[0174] Hot data tasks are characterized by: high sensitivity to latency, use of long-connection communication, strict requirements for data consistency, and belonging to real-time online transaction business. Interruption of such tasks will directly lead to a decline in user experience or business failure.

[0175] Warm data tasks are characterized by: high tolerance for latency, use of short connection communication, and belonging to batch processing or offline computing business. Short delays or interruptions of such tasks will not have a substantial impact on core business.

[0176] S430. Filter the shadow resource pool:

[0177] Based on the collaborative decision matrix in step S200, the resource management server selects all resources in the optimal and stable zone from the current candidate resource pool as candidate shadow resources; furthermore, it selects resources with the same steady-state characteristics as the sub-healthy resources from the optimal and stable zone resources, including those located in the same data center, the same rack, and with the same hardware configuration, to ensure performance consistency after replacement.

[0178] S440. Determine the target shadow resource:

[0179] The resource management server selects the resource with the best overall conditions from the candidate shadow resource pool, taking into account factors such as current load rate, remaining capacity, and network distance from sub-healthy resources. The selection algorithm uses a multi-factor weighted scoring, with the current load rate having the highest weight, to ensure that the replacement process does not lead to shadow resource overload.

[0180] The scoring formula for selecting the algorithm is expressed as follows:

[0181] ;

[0182] Candidate shadow resources The overall score; Candidate shadow resources The current load rate; Candidate shadow resources The remaining capacity percentage; Candidate shadow resources The network distance between sub-health resources; The maximum network distance between all candidate shadow resources and sub-health resources; These are the weighting coefficients for each indicator;

[0183] S450. Transparent Transfer of Hot Data Tasks:

[0184] For hot data tasks, the resource management server issues flow table rules through the software-defined network controller to transparently transfer business traffic that was originally directed to unhealthy resources to the target shadow resource within milliseconds.

[0185] The transfer process includes:

[0186] For new connection requests, redirect directly to the shadow resource;

[0187] For established long-lived Transmission Control Protocol (TCP) connections, the connection context is migrated to a shadow resource through connection tracking and seamless transfer techniques to ensure uninterrupted connection.

[0188] By switching the border gateway protocol routing, the service Internet Protocol address of sub-health resources is migrated to shadow resources, achieving seamless failover for the client.

[0189] S460. Gradual Migration of Warm Data Tasks:

[0190] For warm data tasks, the resource management server does not immediately interrupt the executing task, but instead adopts a gradual migration strategy.

[0191] The gradual migration strategy specifically includes:

[0192] Update the resource configuration table in the task scheduler to allocate newly initiated tasks of this type to shadow resources;

[0193] For the warm data tasks currently being performed on sub-health resources, allow them to continue running until they end naturally.

[0194] Monitor the progress of tasks related to monitoring sub-health resources and record the estimated completion time for each task.

[0195] S470. Migration process monitoring:

[0196] The resource management server continuously monitors the status of unhealthy resources and shadow resources. For unhealthy resources, it monitors their transient characteristic changes to determine whether they have recovered on their own; for shadow resources, it monitors their load changes to ensure that no performance bottlenecks occur after replacement.

[0197] S480. Confirmation of completion of fill-in:

[0198] Once all tasks on the sub-health resource have been migrated, including hot data tasks that have been transferred and warm data tasks that have ended naturally, the resource management server records the completion log and marks the event as complete.

[0199] S490. Resource Status Post-Processing:

[0200] For sub-healthy resources that have been successfully restored, if their transient characteristics subsequently return to normal and they are determined to be risk-free by the trend prediction model, the resource management server will re-mark their status as healthy and they can participate in subsequent allocation as candidate resources; for resources that are completely offline or cannot be repaired, they will be permanently removed from the resource pool.

[0201] Through the above steps, the resource management server achieves smooth dynamic replacement based on task sensitivity classification, completing the task migration on sub-optimal resources while minimizing business impact, thus ensuring the overall service continuity of the system.

[0202] Example 2:

[0203] A dual-state storage collaborative resource allocation pre-detection and dynamic compensation method system, the system includes a resource management server and probes deployed on each allocatable resource end;

[0204] The resource management server includes:

[0205] Dual-state feature modeling module: used to construct a dual-state feature model containing steady-state and transient features for each allocable resource; the steady-state features include hardware model, total capacity, nominal performance parameters, historical average failure rate, and physical location information; the transient features include real-time input / output operations per second, current load rate, temperature, read / write response latency, and remaining lifetime percentage;

[0206] Probe management module: used to deploy probes to each allocable resource end and to send acquisition strategy configuration files to the probes. The acquisition strategy configuration files specify the first acquisition frequency of transient characteristics and the second acquisition frequency of steady-state characteristics.

[0207] Data receiving and preprocessing module: used to receive transient feature data and steady-state feature update data reported by the probe, and to perform smoothing and noise reduction processing on the transient feature data using an exponential weighted moving average algorithm;

[0208] Dual-state feature database: includes a high-speed in-memory database and a back-end relational database; the high-speed in-memory database is used to store real-time updated transient feature data, providing microsecond-level access latency; the back-end relational database is used to persistently store steady-state feature data and historical data;

[0209] Resource allocation pre-detection module: It is used to receive resource allocation requests, read the dual-state features of each candidate resource from the candidate resource pool according to the service quality requirements in the request, call the collaborative judgment matrix to map the dual-state features of each candidate resource to the matrix quadrant, and determine the target resource according to the preset quadrant priority order;

[0210] Trend prediction module: Used to load a pre-built trend prediction model based on long short-term memory neural network, predict the trend of transient characteristics of allocated resources, and trigger a sub-healthy state marker when the prediction result exceeds the safety threshold.

[0211] Dynamic fill module: In response to dynamic fill commands, based on the task sensitivity classification results, it performs transparent transfer for hot data tasks and progressive migration for warm data tasks until all tasks have been migrated.

[0212] The probes deployed on each of the allocable resource terminals are lightweight background daemons that consume less than one percent of system resources. They are used to collect transient feature data at a first frequency and report it using the User Datagram Protocol (UDP) according to the collection strategy issued by the resource management server, and to check and report steady-state feature update data at a second frequency and report it using the Transmission Control Protocol (TCP).

[0213] Example 3:

[0214] An application of a dual-state storage collaborative resource allocation pre-detection and dynamic compensation method in a large-scale distributed storage system, using the following system configuration and testing scheme:

[0215] I. System Configuration:

[0216] Hardware environment:

[0217] Resource management server: configured with 2 Intel Xeon Gold 6330 processors, 512GB memory, 2TB NVMe SSD for high-speed memory database, and 10TB HDD array for back-end relational database;

[0218] Allocable resource nodes: a total of 200 storage nodes, each node is configured with 2 Intel Xeon Silver 4314 processors, 128GB of memory, 8 6TB SAS hard drives, and dual 10Gbps network interfaces;

[0219] Network environment: Spine-Leaf architecture is adopted, with core switches at 100Gbps and access switches at 10Gbps / 25Gbps.

[0220] Software environment:

[0221] Operating systems: The resource management server runs CentOS 8.4, and the storage nodes run RockyLinux 8.4;

[0222] Database: The high-speed in-memory database uses a Redis 6.2 cluster, and the backend relational database uses PostgreSQL 13;

[0223] Probe program: A self-developed probe program based on the C language, supporting multiple hardware monitoring interfaces such as iostat, smartctl, and ipmitool;

[0224] Trend prediction model: Based on the LSTM model of TensorFlowLite, the model size is about 15MB and the inference time is less than 50ms.

[0225] II. Parameter Configuration:

[0226] Transient feature acquisition frequency: 100 milliseconds;

[0227] Steady-state feature update frequency: 1 minute;

[0228] Smoothing coefficient : 0.2;

[0229] Collaborative Decision Matrix Threshold 0.6, 0.6;

[0230] LSTM model input sequence length 30, Output predicted sequence length :10;

[0231] Safety thresholds: Device temperature 65℃, read / write response latency 50ms, IOPS below 5000;

[0232] Shadow resource selection weights: , , .

[0233] Test plan:

[0234] Test scenario design:

[0235] Scenario 1: Resource allocation test under normal load, simulating 1000 concurrent task requests to verify the resource filtering effect of the collaborative judgment matrix;

[0236] Scenario 2: Resource degradation simulation test, gradually increasing the load pressure on 10 storage nodes to verify the early warning accuracy of the trend prediction model;

[0237] Scenario 3: Fault injection test, randomly select 5 nodes to simulate hardware failure, and verify the integrity and migration efficiency of the dynamic replacement process;

[0238] Scenario 4: Extreme load test, increasing the system load to over 80% to verify the system's stability under high load.

[0239] Test metrics:

[0240] Resource allocation accuracy: Proportion of resources allocated to the priority and stable zones;

[0241] Prediction accuracy: accuracy and recall rate of sub-health status assessment;

[0242] Migration efficiency: hot data task migration time, warm data task migration completion time;

[0243] Business impact: The rate of task interruption and the increase in latency during the migration process.

[0244] A leading cloud computing service provider deployed this method in its object storage system, managing over 5,000 storage nodes and serving millions of users.

[0245] Implementation results:

[0246] The collaborative decision matrix improves resource allocation efficiency by 35% and increases resource utilization in the optimal and stable zone by 28%.

[0247] The accuracy rate of identifying sub-health conditions reached 94.7%, and the median early warning time was 18 minutes.

[0248] The dynamic data replacement process is executed automatically approximately 50 times per day, with 92% of hot data task transfers completed within 50 milliseconds.

[0249] SLA defaults due to hardware failures decreased by 76%;

[0250] Maintenance personnel have shifted from passively responding to faults to proactively preventing maintenance, resulting in a 40% improvement in maintenance efficiency.

[0251] The sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0252] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A dual-state storage collaborative resource allocation pre-detection and dynamic compensation method, applied to a resource allocation system, characterized in that, The resource allocation system includes a resource management server and probes deployed at each allocable resource endpoint, and includes the following steps: The S100 resource management server constructs a dual-state feature model containing steady-state and transient features for each allocable resource, and collects and maintains the latest dual-state feature data of each resource in real time through probes deployed at each resource end; S200. When a resource allocation request containing service quality requirements is received, the resource management server reads the current dual-state characteristics of each candidate resource from the candidate resource pool based on the latest dual-state feature data; then it calls the pre-stored collaborative decision matrix to map the dual-state characteristics of each candidate resource to the collaborative decision matrix to determine its quadrant, and selects the target resource that meets the service quality requirements from the candidate resource pool according to the preset quadrant priority order. S300. After resource allocation, the resource management server uses a trend prediction model to predict the transient characteristic data of the allocated resources. If the prediction result exceeds the safety threshold, the resource is marked as sub-healthy and a dynamic replacement instruction is generated. The S400 resource management server divides tasks running on sub-healthy resources into hot data tasks and warm data tasks. It selects resources in the optimal and stable zone from the candidate resource pool as shadow resources. Hot data tasks are transparently transferred through underlying network technology, while warm data tasks are migrated gradually until the task migration is complete.

2. The resource allocation pre-detection and dynamic replacement method for dual-state storage collaboration according to claim 1, characterized in that, S100 includes: The resource management server loads the metadata definition file of the dual-state feature model. The metadata definition file specifies the steady-state feature field group and the transient feature field group. The steady-state feature field group includes at least one of the following: hardware model, total capacity, and nominal performance parameters. The transient feature field group includes at least one of the following: number of real-time input / output operations per second, current load rate, and read / write response latency. The resource management server identifies allocable resources through a resource discovery mechanism and creates an independent model instance for each discovered resource. The resource management server sends query commands to each resource through the standard hardware management interface protocol to obtain its steady-state characteristic data, and fills the obtained steady-state characteristic data into the steady-state characteristic field of the corresponding model instance to realize the initialization assignment of the model instance; The resource management server selects a compatible probe version based on the hardware model and operating system type of each resource and distributes it to each allocable resource. It also sends a collection strategy configuration file to the probe, specifying that transient feature data is collected at a first frequency of hundreds of milliseconds and steady-state feature update data is collected at a second frequency of minutes. The resource management server receives transient feature data reported by the probe, performs smoothing and denoising processing using an exponentially weighted moving average algorithm, and then updates the corresponding model instance in the high-speed memory database using an overwrite write method. At the same time, it receives steady-state feature update data reported by the probe, updates the corresponding steady-state feature fields in the model instance, and persists them to the backend relational database.

3. The resource allocation pre-detection and dynamic replacement method for dual-state storage collaboration according to claim 2, characterized in that, In the exponentially weighted moving average algorithm, the smoothed value at the current time is calculated by weighting the measured value at the current time and the smoothed value at the previous time using a smoothing coefficient. The smoothing coefficient ranges from 0 to 1. The closer the smoothing coefficient is to 1, the greater the weight of the current measured value. The closer the smoothing coefficient is to 0, the greater the weight of the historical value.

4. The resource allocation pre-detection and dynamic replacement method for dual-state storage collaboration according to claim 1, characterized in that, Specifically, S200 includes: After receiving a resource allocation request that includes quality of service requirements, the resource management server queries the dual-state feature database to form a candidate resource pool based on the resource type and basic capacity conditions specified in the request. The current dual-state feature data of each resource in the candidate resource pool is read in batches from the high-speed memory database. The current dual-state feature data is the latest data maintained in real time by S100. The read bi-state feature data is normalized to map feature values ​​of different dimensions to a standard interval; The pre-stored collaborative decision matrix is ​​invoked. The collaborative decision matrix is ​​a two-dimensional coordinate system constructed with the normalized steady-state feature score as the horizontal axis and the transient feature score as the vertical axis. The two-dimensional coordinate system is divided into four quadrants by the preset steady-state feature threshold and transient feature threshold. The normalized bi-state features of each candidate resource are used as coordinate points and mapped to the collaborative decision matrix to determine its quadrant. Candidate resources are sorted according to a preset quadrant priority order, and the resource with the highest priority is selected as the target resource for this allocation.

5. The resource allocation pre-detection and dynamic replacement method for dual-state storage collaboration according to claim 4, characterized in that, The four quadrants include: The first quadrant has both steady-state and transient characteristics that are higher than the corresponding thresholds, i.e., resources in the optimal and stable region; The second quadrant, representing the potential resource area, consists of regions where steady-state characteristics are below a threshold but transient characteristics are above a threshold. Resources in the third quadrant, where steady-state characteristics are above the threshold and transient characteristics are below the threshold, are considered risk zone resources. Resources with both bi-state characteristics below the corresponding threshold in the fourth quadrant are considered to be in the elimination zone.

6. The resource allocation pre-detection and dynamic replacement method for dual-state storage collaboration according to claim 1, characterized in that, The S300 includes: After resource allocation, the resource management server will include the allocated resources in the continuous monitoring list, and its historical transient characteristic data will be organized into time series and stored in the time series database. A pre-set trend prediction model based on a long short-term memory neural network is invoked, and key transient feature data of the resource at the most recent preset number of time points are input into the model at a preset frequency to obtain a sequence of predicted values ​​at the next preset number of time points. The predicted value sequence is compared point by point with the preset safety threshold. If the prediction result shows that a certain key transient feature of the resource will reach or exceed the safety threshold within the preset time window in the future, the resource is judged to be in a sub-healthy state. Resources judged to be in a sub-healthy state will continue to be monitored, but will no longer be considered as candidate resources for new allocation. Generate dynamic fill-in instructions, which include a unique identifier for the sub-health resource, a list of tasks currently running on the resource, and the sensitivity pre-grading results of each task.

7. The resource allocation pre-detection and dynamic replacement method for dual-state storage collaboration according to claim 6, characterized in that, The key transient characteristics include device temperature, read / write response latency, and the number of input / output operations per second.

8. The resource allocation pre-detection and dynamic replacement method for dual-state storage collaboration according to claim 1, characterized in that, The S400 includes: The resource management server parses the dynamic fill instruction to obtain the unique identifier of sub-health resources, the task list, and the sensitivity pre-classification results of each task; Based on their sensitivity to changes in resource status, tasks are divided into hot data tasks and warm data tasks. Hot data tasks are those that are highly sensitive to latency and use long-lived connection communication; warm data tasks are those that have a higher tolerance for latency and use short-lived connection communication or belong to batch processing tasks. Based on the collaborative decision matrix, all resources in the optimal and stable zone are selected from the current candidate resource pool as candidate shadow resources; From the candidate shadow resource pool, taking into account factors such as current load rate, remaining capacity, and network distance from sub-healthy resources, the resource with the best overall conditions is selected as the target shadow resource. For hot data tasks, flow table rules are issued by the software-defined network controller to transparently transfer business traffic that originally pointed to sub-healthy resources to the target shadow resources; For warm data tasks, a gradual migration strategy is adopted. The resource configuration table is updated in the task scheduler, and newly initiated warm data tasks are allocated to the target shadow resource. Warm data tasks that are currently being executed on sub-healthy resources are allowed to continue running until they end naturally. Once all tasks on the sub-health resource platform have been migrated, the record replacement is complete.

9. The resource allocation pre-detection and dynamic replacement method for dual-state storage collaboration according to claim 8, characterized in that, The method for selecting the resource with the best comprehensive conditions as the target shadow resource includes: using a multi-factor weighted scoring method to calculate a comprehensive score based on the current load rate, remaining capacity ratio, and network distance to the sub-healthy resource of the candidate shadow resource, and selecting the resource with the highest comprehensive score as the target shadow resource; wherein, the current load rate has the highest weight in the scoring.

10. A resource allocation pre-detection and dynamic compensation system for implementing the dual-state storage collaboration method according to any one of claims 1-9, characterized in that, This includes a resource management server and probes deployed at each allocable resource endpoint; The resource management server includes: The dual-state feature modeling module is used to load the metadata definition file of the dual-state feature model, create an independent model instance for each allocable resource, and obtain steady-state feature data through the standard hardware management interface to initialize the model instance. The probe management module is used to deploy probes to each allocable resource end and issue acquisition strategies that include the acquisition frequency of dual-state features. The data receiving and preprocessing module is used to receive transient feature data and steady-state feature update data reported by the probe, and to perform smoothing and noise reduction processing on the transient feature data. The dual-state feature database includes a high-speed in-memory database for storing real-time transient feature data and a back-end relational database for persistently storing steady-state feature data. The resource allocation pre-detection module is used to receive resource allocation requests containing service quality requirements, and to determine the target resource based on the latest bi-state feature data maintained in the high-speed memory database by calling the collaborative decision matrix and quadrant priority. The trend prediction module is used to load a trend prediction model based on a long short-term memory neural network, identify the sub-health status of allocated resources, and trigger replacement commands. The dynamic fill module is used to respond to fill commands and perform differentiated smooth migration of hot data tasks and warm data tasks according to the task sensitivity level. The probes deployed at each allocable resource terminal are used to collect and report transient characteristic data and steady-state characteristic update data of the resources at different frequencies according to the collection strategy.

Images