Solid state disk wear leveling global optimization method and system based on multi-granularity data features

Abstract

The invention relates to the technical field of data storage optimization, and discloses a solid state disk wear leveling global optimization method and system based on multi-granularity data features. The solid state disk wear leveling global optimization method specifically comprises the following steps that S101, lightweight data feature monitoring points with blocks, units and chips as basic granularity are deployed in a solid state disk controller, and original multi-dimensional data features related to wear are continuously collected and quantified in real time, all collected multi-dimensional data features are converted into dimensionless feature scores which can be transversely compared through normalization processing. According to the method, multi-dimensional feature fusion and normalization processing of three-level granularity of blocks, units and chips of a fixed hard disk are utilized, a prediction model based on a time sequence is combined, a dynamic risk thermodynamic diagram covering the whole disk is generated, a system can accurately position current and future potential abrasion hot spots, and the system reliability is improved. And passive equalization operation is converted into active risk prevention.

Description

Technical Field

[0001] This invention relates to the technical field of data storage optimization, and more specifically, to a global optimization method and system for wear leveling of solid-state drives based on multi-granularity data characteristics. Background Technology

[0002] Solid-state drives (SSDs) have become mainstream storage devices due to their high performance, low latency, and shock resistance. Their core storage medium is NAND flash memory. The physical characteristics of NAND flash memory limit the number of program / erase (P / E) cycles per cell. Excessive wear can lead to a decrease in data reliability and even complete failure. To overcome this defect, wear leveling technology has emerged and become one of the core algorithms of the flash translation layer (FTL) in SSD controllers. Existing wear leveling technologies are mainly divided into two categories: dynamic wear leveling prioritizes the operation on physical blocks with lower current wear when writing new data, thereby achieving initial wear distribution; static wear leveling goes a step further, actively exchanging static data stored in low-wear blocks with dynamic data in high-wear blocks, thus allowing data blocks that are not frequently updated to also participate in the wear process.

[0003] However, although existing technologies have alleviated the problem of uneven wear to some extent, their design concepts have fundamental limitations, resulting in optimization effects that fall far short of the ideal. First, most existing solutions make decisions based on a single granularity (usually a physical block) and a single dimension (P / E cycle count), failing to detect wear hotspots or severe wear dispersion that may form over a larger area. Second, existing technologies are inherently passive and reactive, making decisions solely based on historically accumulated P / E counts and lacking the ability to predict future wear trends. When the system identifies that a block has excessive wear, the risk has often already accumulated to a certain extent. At this point, remedial data migration not only has limited effectiveness but may also cause significant write amplification and performance jitter due to concentrated operations, thereby exacerbating overall wear and performance instability.

[0004] Based on the above, the existing technical problems severely restrict the full realization of the potential durability of SSDs, and a more intelligent, holistic and forward-looking solution is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a global optimization method and system for wear leveling of solid-state drives based on multi-granularity data features. This enables the system to accurately locate current and future potential wear hotspots, transforming passive leveling operations into proactive risk prevention. It provides unprecedented data support for global optimization decisions, thereby fundamentally improving the accuracy and effectiveness of wear leveling and aiming to solve the problems in the prior art.

[0006] This invention is implemented as follows: a global optimization method for wear leveling of solid-state drives based on multi-granularity data features, specifically including the following steps:

[0007] S101: Deploys lightweight data feature monitoring points at the block, cell, and chip level in the solid-state drive controller to collect and quantify raw multidimensional data features related to wear in real time and continuously. All collected multidimensional data features are normalized and converted into dimensionless feature scores that can be compared laterally.

[0008] S102: Input the feature scores into the state assessment engine based on the weighted scoring algorithm. The state assessment engine first calculates the instantaneous wear pressure score of each block to form a comprehensive assessment value, and generates a lightweight prediction model based on sliding window time series analysis. The lightweight prediction model predicts the wear rate and potential risk level of each block, unit and even chip in a specific future time window based on the trend of historical feature data, and outputs a risk heat map covering all physical blocks of the entire disk, including the current state and future predictions.

[0009] S103: Based on the generated risk heat map, a global mathematical model is constructed with the ultimate goal of maximizing the overall lifespan. The decision variables of the global optimization mathematical model are the mapping relationship between logical addresses and physical addresses during the next garbage collection or host write operation, thereby completing the optimization of the global mathematical model.

[0010] S104: Decision-making schemes are obtained through optimized global mathematical models, which are then transformed into operation instructions that can be executed by the flash conversion layer. Based on these instructions, specific task queues, including hot and cold data identification, data block merging, and physical address remapping, are dynamically generated. When the SSD is idle or under low load, a background deep optimization task is initiated to proactively move data from high-risk hotspot areas to low-wear coldspot areas and finely adjust the mapping table. When the SSD is under high load, the mode is switched to prioritize host I / O performance, performing only necessary lightweight balancing operations with minimal write amplification.

[0011] S105: After the decision-making plan is implemented, the solid-state drive enters a continuous verification and adjustment cycle. By comparing the multi-granularity feature data collected before and after the strategy execution, the actual effect of this wear leveling operation is quantitatively evaluated. Based on the comparison of actual effects, if it is higher than expected, the initiative of similar strategies is appropriately enhanced; if the cost-effectiveness is low or negative effects are produced, the feature weights, penalty coefficients or strategy trigger thresholds are automatically adjusted. Through a continuous data-driven feedback mechanism, it is ensured that the entire wear leveling optimization can adapt to different usage scenarios, workloads and various stages of the hard drive's life cycle.

[0012] Furthermore, in S101, lightweight data feature monitoring points at the block, cell, and chip level are deployed in the solid-state drive controller to continuously collect and quantify raw, multi-dimensional data features related to wear in real time, including:

[0013] At the block level, monitor and record the programming / erase cycle count, number of valid pages, last access timestamp, and data retention time for each physical block;

[0014] At the unit granularity, the standard deviation and range of the P / E cycle counts of all blocks within the unit are statistically analyzed and aggregated to quantify the wear dispersion within the unit, while also recording the total data throughput of the unit.

[0015] At the chip level, monitor the overall average wear level, the difference between the largest and smallest wear blocks, and the rate of increase in potential error bit rate due to read interference and programming interference.

[0016] Furthermore, all collected multidimensional data features are normalized and transformed into dimensionless feature scores that can be compared laterally, including:

[0017] By using the maximum-minimum normalization or Z-score normalization mathematical methods, raw data from different monitoring sources are uniformly mapped to a preset dimensionless standard interval, eliminating the comparison barriers between different physical quantities, so that the wear dispersion within the same unit can be directly and meaningfully compared with the temperature reading of a chip.

[0018] The normalized basic feature values ​​are assigned dynamically changing weight coefficients based on their contribution to the wear impact under different life cycles and workloads, and a comprehensive feature score is synthesized through a weighted fusion algorithm.

[0019] The feature score represents the current instantaneous state of the monitored object. When arranged in a time sequence, it forms a dynamically evolving data stream. By analyzing the trend and pattern of the data stream, a multi-level health and load feature profile is constructed, from block to unit to chip, which reflects both the real-time state and predicts future risks.

[0020] Furthermore, in S102, the lightweight prediction model predicts the wear rate and potential risk level of each block, cell, and even chip within a specific future time window based on historical characteristic data trends, including:

[0021] The lightweight prediction model establishes an independent time series database for each monitoring point, stores its historical feature scores at fixed periods, and uses a sliding window mechanism to select data from the most recent N periods as samples. It then uses a lightweight exponential smoothing method or ARIMA model for analysis to capture the short-term trend and periodic pattern of wear changes, where N is greater than 1.

[0022] The trained prediction model extrapolates the time series of each monitoring point and outputs its predicted feature score for the next M periods. By comparing it with a preset dynamic threshold, the expected wear rate for that period is calculated, and finally a risk prediction report covering all areas is generated. The risk prediction report is mapped to a specific risk level, where M is less than N.

[0023] Furthermore, by comparing the data with a preset dynamic threshold, the expected wear rate within that time period is calculated, ultimately generating a risk prediction report covering all areas of the system. This risk prediction report is mapped to specific risk levels, including:

[0024] The preset dynamic threshold is not a fixed value, but is determined based on its own historical performance. The threshold is dynamically calculated based on historical data, so that the risk judgment standard can adapt to the characteristics of different solid-state drives and the overall health status of the entire drive, thereby calculating the statistically significant expected wear rate within the future time window.

[0025] The predicted feature score and wear rate of each solid-state drive are compared with its corresponding dynamic threshold. Based on the magnitude and duration of exceeding the threshold, a predefined risk mapping matrix is ​​used to classify them into low, medium, and high risk levels. Finally, a structured and visualized global risk prediction report is automatically generated, which accurately identifies all potential failure points that need to be given priority and dealt with across the entire disk.

[0026] Further, in S103, the decision variables of the global optimization mathematical model are the mapping relationship between logical addresses and physical addresses during the next garbage collection or host write operation, thereby completing the optimization of the global mathematical model. The mapping relationship constraints include:

[0027] The wear of each block must not exceed the safety threshold in a short period of time, the write amplification caused by a single operation must not exceed the set upper limit, and real-time performance response requirements must be met.

[0028] The global mathematical model seeks a Pareto optimal solution by balancing the immediate wear leveling effect with long-term performance overhead, ensuring the most effective global lifetime extension at an acceptable performance cost.

[0029] Furthermore, in S104, a decision scheme is obtained through an optimized global mathematical model, which is then transformed into operation instructions that can be executed by the flash memory conversion layer. Based on these operation instructions, a specific task queue is dynamically generated, including tasks such as hot and cold data identification, data block merging, and physical address remapping.

[0030] The global optimization model obtains the optimal decision variables through the solver, namely the best mapping relationship from logical page to physical page and the target location of the data to be migrated. According to the FTL protocol specification, the optimal decision variables are transformed into a series of atomic operation instructions and written to the physical block and the mapping table entries from logical address LBA to physical address PBA are updated.

[0031] Upon receiving these atomic operation instructions, the system instantly identifies hot and cold data based on data access frequency statistics, dynamically groups and sorts the instructions, prioritizes the migration of hot data in high-risk areas to ensure real-time performance, arranges cold data merging and remapping operations to be executed in a low-priority queue in the background, and dynamically adjusts the execution rate and concurrency of the queue according to the real-time load to balance wear leveling efficiency and I / O performance response.

[0032] Furthermore, in S105, after the decision-making plan is executed, the solid-state drive enters a continuous verification and adjustment cycle. The actual effect of this wear leveling operation is quantitatively evaluated by comparing the multi-granularity feature data re-collected before and after the strategy execution, including:

[0033] After the decision-making plan is completed, a new round of comprehensive multi-granular data feature collection is immediately launched to obtain key indicators including the latest wear variance, maximum wear value, average wear degree and actual write amplification factor of the target area and surrounding related areas, and to accurately compare these real-time data after execution with the historical benchmark data recorded before execution.

[0034] By calculating the changes in key performance indicators, the gain assessment results of this operation are quantitatively evaluated. The gain assessment results are directly input into the parameter self-adjustment engine as a feedback signal. The parameter self-adjustment engine drives the weight coefficients or strategy trigger thresholds in the global mathematical model to be dynamically fine-tuned.

[0035] Compared with existing technologies, the global optimization method and system for solid-state drive wear leveling based on multi-granularity data features provided by this invention has the following beneficial effects:

[0036] 1. By introducing multi-granularity data feature acquisition and analysis, a refined and forward-looking perception of the health status of solid-state drives (SSDs) is achieved. By using the fusion and normalization of multi-dimensional features at the block, unit, and chip levels of the SSD, and combining them with a time-series-based prediction model, a dynamic risk heat map covering the entire drive is generated. This enables the system to accurately locate current and future potential wear hotspots, transforming passive wear leveling operations into proactive risk prevention. This provides unprecedented data support for global optimization decisions, thereby fundamentally improving the accuracy and effectiveness of wear leveling.

[0037] 2. By constructing a global optimization model with multi-objective constraints and forming an adaptive adjustment mechanism with closed-loop feedback, the optimal balance between lifespan extension and performance overhead is achieved. Dynamic weight coefficients and fuzzy logic control are adopted to enable the optimization objective to intelligently respond to different lifespan stages and workloads. Statistical tests are used to quantitatively evaluate the operational effect and drive parameter self-optimization. This not only significantly improves the overall lifespan and reliability of the solid-state drive, but also effectively suppresses write amplification and ensures the performance stability of the system under various application scenarios, forming a complete intelligent, robust, and adaptive solution.

[0038] A global optimization system for solid-state drive wear leveling based on multi-granularity data features is used to execute the aforementioned global optimization method for solid-state drive wear leveling. The system includes:

[0039] The data feature monitoring and processing module is used to collect raw multidimensional data features related to wear in real time in the solid-state drive controller at the basic granularity of blocks, cells and chips, and to normalize and weight and fuse all collected multidimensional data features to generate feature scores that can be compared horizontally.

[0040] The status assessment and prediction module is used to receive feature scores, calculate the instantaneous wear pressure score of each block through a weighted scoring algorithm, and establish a prediction model based on sliding window time series analysis to predict the wear rate and risk level of each granular unit within a specific future time window, and generate a risk heat map covering the entire system.

[0041] The global optimization decision module is used to construct a global optimization mathematical model based on the risk heat map with the goal of maximizing the overall lifespan. The solver then obtains the optimal decision scheme for the mapping relationship between logical addresses and physical addresses.

[0042] The strategy execution and scheduling module is used to convert decision schemes into executable operation instructions for the flash conversion layer, dynamically generate data block merging and physical address remapping task queues, and intelligently schedule foreground and background balanced operations according to system load status.

[0043] The effect verification and parameter adjustment module is used to quantitatively evaluate the operation effect by comparing multi-granular feature data before and after the strategy is executed, and adaptively adjust the feature weights, penalty term coefficients and strategy trigger thresholds based on the evaluation results.

[0044] Furthermore, the data feature monitoring and processing module includes:

[0045] The data acquisition unit is used to continuously and in real time acquire raw, multi-dimensional data features related to wear at the block, cell, and chip levels. Specifically, this includes: monitoring programming / erase cycle counts, number of valid pages, last access timestamp, and data retention time at the block level; calculating the standard deviation and range of P / E cycle counts for all blocks within a cell at the cell level to quantify wear dispersion and recording the total data throughput of that cell; and monitoring the overall average wear level, the difference between the largest and smallest worn blocks, and the rate of change in error bit rate caused by read and programming interference at the chip level.

[0046] The data normalization unit is used to receive raw multidimensional data from the data acquisition unit. Through the maximum-minimum normalization or Z-score normalization mathematical method, it maps the raw data with different dimensions and ranges to a preset dimensionless standard interval, eliminates the comparison barrier between different physical quantities, and provides a standardized data foundation for subsequent fusion calculation.

[0047] The feature fusion unit receives normalized data, assigns dynamically changing weight coefficients to the data features based on their contribution to wear impact under different lifecycles and workloads, and synthesizes a comprehensive feature score through a weighted fusion algorithm. This score is arranged in a time series to form a dynamically evolving data stream, which is used to construct a multi-level health and load feature profile that reflects real-time status and predicts future risks. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the global optimization method for solid-state drive wear leveling based on multi-granularity data features proposed in this invention.

[0049] Figure 2 This is a flowchart illustrating the process of real-time and continuous collection and quantification of original multi-dimensional data features related to wear in the global optimization method for wear leveling of solid-state drives based on multi-granularity data features proposed in this invention.

[0050] Figure 3 This is a schematic diagram of the global optimization system for solid-state drive wear leveling based on multi-granularity data features proposed in this invention.

[0051] Figure 4This is a schematic diagram of the data feature monitoring and processing module in the global optimization method for solid-state drive wear leveling based on multi-granularity data features proposed in this invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] The implementation of the present invention will be described in detail below with reference to specific embodiments.

[0054] In the accompanying drawings of this embodiment, the same or similar reference numerals correspond to the same or similar components. In the description of this invention, it should be understood that if terms such as "upper," "lower," "left," and "right" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting this invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0055] Reference Figure 1-2 As shown, the global optimization method for solid-state drive wear leveling based on multi-granularity data features specifically includes the following steps:

[0056] S101: Deploys lightweight data feature monitoring points at the block, cell, and chip level in the solid-state drive controller to collect and quantify raw multidimensional data features related to wear in real time and continuously. All collected multidimensional data features are normalized and converted into dimensionless feature scores that can be compared laterally.

[0057] Specifically, lightweight data feature monitoring points at the block, cell, and chip level are deployed in the solid-state drive controller to continuously collect and quantify raw, multi-dimensional data features related to wear in real time, including:

[0058] At the block level, monitor and record the programming / erase cycle count, number of valid pages, last access timestamp, and data retention time for each physical block;

[0059] At the unit granularity, the standard deviation and range of the P / E cycle counts of all blocks within the unit are statistically analyzed and aggregated to quantify the wear dispersion within the unit, while also recording the total data throughput of the unit.

[0060] At the chip level, monitor the overall average wear level, the difference between the largest and smallest wear blocks, and the rate of increase in potential error bit rate due to read interference and programming interference;

[0061] S102: Input the feature scores into the state assessment engine based on the weighted scoring algorithm. The state assessment engine first calculates the instantaneous wear pressure score of each block to form a comprehensive assessment value, and generates a lightweight prediction model based on sliding window time series analysis. The lightweight prediction model predicts the wear rate and potential risk level of each block, unit and even chip in a specific future time window based on the trend of historical feature data, and outputs a risk heat map covering all physical blocks of the entire disk, including the current state and future predictions.

[0062] The lightweight prediction model, based on historical characteristic data trends, predicts the wear rate and potential risk level of each block, cell, and even chip within a specific future time window, including:

[0063] The lightweight prediction model establishes an independent time series database for each monitoring point, stores its historical feature scores at fixed periods, and uses a sliding window mechanism to select data from the most recent N periods as samples. It then uses a lightweight exponential smoothing method or ARIMA model for analysis to capture the short-term trend and periodic pattern of wear changes, where N is greater than 1.

[0064] The trained prediction model extrapolates the time series of each monitoring point and outputs its predicted feature score for the next M periods. By comparing it with the preset dynamic threshold, the expected wear rate for that period is calculated, and finally a risk prediction report covering all areas is generated. The risk prediction report is mapped to a specific risk level, where M is less than N.

[0065] S103: Based on the generated risk heat map, construct a global mathematical model with the ultimate goal of maximizing the overall lifespan. The decision variables of the global optimization mathematical model are the mapping relationship between logical addresses and physical addresses during the next garbage collection or host write operation, thereby completing the optimization of the global mathematical model.

[0066] S104: Decision-making schemes are obtained through optimized global mathematical models, which are then transformed into operation instructions that can be executed by the flash conversion layer. Based on these instructions, specific task queues, including hot and cold data identification, data block merging, and physical address remapping, are dynamically generated. When the SSD is idle or under low load, a background deep optimization task is initiated to proactively move data from high-risk hotspot areas to low-wear coldspot areas and finely adjust the mapping table. When the SSD is under high load, the mode is switched to prioritize host I / O performance, performing only necessary lightweight balancing operations with minimal write amplification.

[0067] Specifically, a decision scheme is obtained through an optimized global mathematical model, which is then transformed into operation instructions that can be executed by the flash memory conversion layer. Based on these operation instructions, a specific task queue is dynamically generated, including tasks such as hot and cold data identification, data block merging, and physical address remapping.

[0068] The global optimization model uses a solver to obtain the optimal decision variables, namely the best mapping relationship between logical pages and physical pages and the target location of the data to be migrated. According to the FTL protocol specification, the optimal decision variables are transformed into a series of atomic operation instructions and written to the physical block and the mapping table entries from logical address LBA to physical address PBA are updated.

[0069] Upon receiving these atomic operation instructions, hot and cold data are identified in real time based on data access frequency statistics. The instructions are dynamically grouped and sorted, and hot data migration in high-risk areas is prioritized to ensure real-time performance. Cold data merging and remapping operations are scheduled to be executed in a low-priority queue in the background. The execution rate and concurrency of the queue are dynamically adjusted according to the real-time load to balance wear leveling efficiency and I / O performance response.

[0070] S105: After the decision-making plan is implemented, the solid-state drive (SSD) enters a continuous verification and adjustment cycle. By comparing the multi-granularity feature data collected before and after the strategy execution, the actual effect of this wear leveling operation is quantitatively evaluated. Based on the comparison of actual effects, if it is higher than expected, the initiative of similar strategies is appropriately enhanced; if the cost-effectiveness is low or negative effects are produced, the feature weights, penalty coefficients, or strategy trigger thresholds are automatically adjusted. Through a continuous data-driven feedback mechanism, it is ensured that the entire wear leveling optimization can adapt to different usage scenarios, workloads, and various stages of the hard drive's life cycle. By introducing multi-granularity data feature collection and analysis, a refined and forward-looking perception of the SSD's health status is achieved. By using the fusion and normalization of multi-dimensional features at the block, unit, and chip levels of the fixed hard drive, combined with a time-series-based prediction model, a dynamic risk heat map covering the entire drive is generated. This enables the system to accurately locate current and future potential wear hotspots, transforming passive leveling operations into proactive risk prevention. This provides unprecedented data support for global optimization decisions, thereby fundamentally improving the accuracy and effectiveness of wear leveling.

[0071] In S101 of this embodiment, all collected multidimensional data features are normalized and converted into dimensionless feature scores that can be compared laterally, including:

[0072] By using the maximum-minimum normalization or Z-score normalization mathematical methods, raw data from different monitoring sources are uniformly mapped to a preset dimensionless standard interval, eliminating the comparison barriers between different physical quantities, so that the wear dispersion within the same unit can be directly and meaningfully compared with the temperature reading of a chip.

[0073] The normalized basic feature values ​​are assigned dynamically changing weight coefficients based on their contribution to the wear impact under different life cycles and workloads, and a comprehensive feature score is synthesized through a weighted fusion algorithm.

[0074] The feature score represents the current instantaneous state of the monitored object. When arranged in a time sequence, it forms a dynamically evolving data stream. By analyzing the trend and pattern of the data stream, a multi-level health and load feature profile is constructed, from block to unit to chip, which reflects both real-time status and predicts future risks.

[0075] In S102 of this embodiment, the expected wear rate within the time period is calculated by comparing it with a preset dynamic threshold, and a risk prediction report covering all areas of the entire system is finally generated. The risk prediction report is mapped with specific risk levels, including:

[0076] The preset dynamic threshold is not a fixed value, but is determined based on its own historical performance. The threshold is dynamically calculated based on historical data, so that the risk assessment standard can adapt to the characteristics of different solid-state drives and the overall health of the entire drive, thereby calculating the statistically significant expected wear rate within the future time window.

[0077] The predicted feature score and wear rate of each solid-state drive are compared with its corresponding dynamic threshold. Based on the magnitude and duration of exceeding the threshold, a predefined risk mapping matrix is ​​used to classify them into low, medium, and high risk levels. Finally, a structured and visualized global risk prediction report is automatically generated, which accurately identifies all potential failure points that need to be given priority and dealt with across the entire disk.

[0078] In S103 of this embodiment, when the decision variable of the global optimization mathematical model is the next garbage collection or host write operation, the mapping relationship from logical address to physical address is constrained to complete the optimization of the global mathematical model. The mapping relationship constraint includes:

[0079] The wear of each block must not exceed the safety threshold in a short period of time, the write amplification caused by a single operation must not exceed the set upper limit, and real-time performance response requirements must be met.

[0080] The global mathematical model seeks a Pareto optimal solution by balancing the immediate wear leveling effect with long-term performance overhead, ensuring the most effective global lifetime extension at an acceptable performance cost.

[0081] In S105 of this embodiment, after the decision-making scheme is executed, the solid-state drive enters a continuous verification and adjustment cycle. The actual effect of this wear leveling operation is quantitatively evaluated by comparing the multi-granularity feature data re-collected before and after the strategy execution, including:

[0082] After the decision-making plan is completed, a new round of comprehensive multi-granular data feature collection is immediately launched to obtain key indicators including the latest wear variance, maximum wear value, average wear degree and actual write amplification factor of the target area and surrounding related areas, and to accurately compare these real-time data after execution with the historical benchmark data recorded before execution.

[0083] By calculating the changes in key performance indicators, the gain assessment results of this operation are quantitatively evaluated. The gain assessment results will be directly input into the parameter self-adjustment engine as a feedback signal. The parameter self-adjustment engine drives the weight coefficients or strategy trigger thresholds in the global mathematical model to be dynamically fine-tuned.

[0084] This technical solution achieves an optimal balance between lifespan extension and performance overhead by constructing a global optimization model with multi-objective constraints and forming an adaptive adjustment mechanism with closed-loop feedback. It adopts dynamic weight coefficients and fuzzy logic control to enable the optimization objective to intelligently respond to different lifespan stages and workloads, and uses statistical tests to quantitatively evaluate the operational effect and drive parameter self-optimization. This not only significantly improves the overall lifespan and reliability of solid-state drives, but also effectively suppresses write amplification and ensures the performance stability of the system under various application scenarios, forming a complete intelligent, robust, and adaptive solution.

[0085] Reference Figure 3-4 As shown, a global optimization system for solid-state drive wear leveling based on multi-granularity data features is used to execute the aforementioned global optimization method for solid-state drive wear leveling. The system includes:

[0086] The data feature monitoring and processing module is used to collect raw multidimensional data features related to wear in real time in the solid-state drive controller at the basic granularity of blocks, cells and chips, and to normalize and weight and fuse all collected multidimensional data features to generate feature scores that can be compared horizontally.

[0087] The status assessment and prediction module is used to receive feature scores, calculate the instantaneous wear pressure score of each block through a weighted scoring algorithm, and establish a prediction model based on sliding window time series analysis to predict the wear rate and risk level of each granular unit within a specific future time window, and generate a risk heat map covering the entire system.

[0088] The global optimization decision module is used to construct a global optimization mathematical model based on the risk heat map with the goal of maximizing the overall lifespan. The solver then obtains the optimal decision scheme for the mapping relationship between logical addresses and physical addresses.

[0089] The strategy execution and scheduling module is used to convert decision schemes into executable operation instructions for the flash conversion layer, dynamically generate data block merging and physical address remapping task queues, and intelligently schedule foreground and background balanced operations according to system load status.

[0090] The effect verification and parameter adjustment module is used to quantitatively evaluate the operation effect by comparing multi-granular feature data before and after strategy execution, and adaptively adjust feature weights, penalty term coefficients and strategy trigger thresholds based on the evaluation results. By constructing a global optimization model with multi-objective constraints and forming an adaptive adjustment mechanism with closed-loop feedback, the optimal balance between lifespan extension and performance overhead is achieved. Dynamic weight coefficients and fuzzy logic control are used to enable the optimization objective to intelligently respond to different life cycle stages and workloads, and statistical tests are used to quantitatively evaluate the operation effect and drive parameter self-optimization. This not only significantly improves the overall lifespan and reliability of solid-state drives.

[0091] In this embodiment, the data feature monitoring and processing module includes:

[0092] The data acquisition unit is used to continuously and in real time acquire raw, multi-dimensional data features related to wear at the block, cell, and chip levels. Specifically, this includes: monitoring programming / erase cycle counts, number of valid pages, last access timestamp, and data retention time at the block level; calculating the standard deviation and range of P / E cycle counts for all blocks within a cell at the cell level to quantify wear dispersion and recording the total data throughput of that cell; and monitoring the overall average wear level, the difference between the largest and smallest worn blocks, and the rate of change in error bit rate caused by read and programming interference at the chip level.

[0093] The data normalization unit is used to receive raw multidimensional data from the data acquisition unit. Through the maximum-minimum normalization or Z-score normalization mathematical method, it maps the raw data with different dimensions and ranges to a preset dimensionless standard interval, eliminates the comparison barrier between different physical quantities, and provides a standardized data foundation for subsequent fusion calculation.

[0094] The feature fusion unit receives normalized data, assigns dynamically changing weight coefficients to the data features based on their contribution to wear and tear under different lifecycles and workloads, and synthesizes a comprehensive feature score through a weighted fusion algorithm. This score is arranged in a time series to form a dynamically evolving data stream, which is used to construct a multi-layered health and load feature profile that reflects real-time status and predicts future risks. By introducing multi-granularity data feature acquisition and analysis, a refined and forward-looking perception of the health status of solid-state drives is achieved.

[0095] This technical solution utilizes the fusion and normalization of multi-dimensional features at the block, unit, and chip levels of a fixed hard drive, combined with a time-series-based prediction model, to generate a dynamic risk heatmap covering the entire drive. This enables the system to accurately locate current and future potential wear hotspots, transforming passive balancing operations into proactive risk prevention. It provides unprecedented data support for global optimization decisions, effectively suppresses write amplification, and ensures the system's performance stability in various application scenarios, forming a complete intelligent, robust, and adaptive solution.

[0096] In this embodiment, the entire operation process can be controlled by a computer to provide signal feedback and implement the steps sequentially. These are all conventional knowledge in current automation control, and will not be elaborated on in this embodiment.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A global optimization method for wear leveling of solid-state drives based on multi-granularity data features, characterized in that, Specifically, the following steps are included: S101: Deploys lightweight data feature monitoring points at the block, cell, and chip level in the solid-state drive controller to collect and quantify raw multidimensional data features related to wear in real time and continuously. All collected multidimensional data features are normalized and converted into dimensionless feature scores that can be compared laterally. S102: Input the feature scores into the state assessment engine based on the weighted scoring algorithm. The state assessment engine first calculates the instantaneous wear pressure score of each block to form a comprehensive assessment value, and generates a lightweight prediction model based on sliding window time series analysis. The lightweight prediction model predicts the wear rate and potential risk level of each block, unit and even chip in a specific future time window based on the trend of historical feature data, and outputs a risk heat map covering all physical blocks of the entire disk, including the current state and future predictions. S103: Based on the generated risk heat map, a global mathematical model is constructed with the ultimate goal of maximizing the overall lifespan. The decision variables of the global optimization mathematical model are the mapping relationship between logical addresses and physical addresses during the next garbage collection or host write operation, thereby completing the optimization of the global mathematical model. S104: Decision-making schemes are obtained through optimized global mathematical models, which are then transformed into operation instructions that can be executed by the flash conversion layer. Based on these instructions, specific task queues, including hot and cold data identification, data block merging, and physical address remapping, are dynamically generated. When the SSD is idle or under low load, a background deep optimization task is initiated to proactively move data from high-risk hotspot areas to low-wear coldspot areas and finely adjust the mapping table. When the SSD is under high load, the mode is switched to prioritize host I / O performance, performing only necessary lightweight balancing operations with minimal write amplification. S105: After the decision-making plan is implemented, the solid-state drive enters a continuous verification and adjustment cycle. By comparing the multi-granularity feature data collected before and after the strategy execution, the actual effect of this wear leveling operation is quantitatively evaluated. Based on the comparison of actual effects, if it is higher than expected, the initiative of similar strategies is appropriately enhanced; if the cost-effectiveness is low or negative effects are produced, the feature weights, penalty coefficients or strategy trigger thresholds are automatically adjusted. Through a continuous data-driven feedback mechanism, it is ensured that the entire wear leveling optimization can adapt to different usage scenarios, workloads and various stages of the hard drive's life cycle.

2. The global optimization method for solid-state drive wear leveling based on multi-granularity data features as described in claim 1, characterized in that, In the S101, lightweight data feature monitoring points at the block, cell, and chip level are deployed in the solid-state drive controller to continuously collect and quantify raw, multi-dimensional data features related to wear in real time, including: At the block level, monitor and record the programming / erase cycle count, number of valid pages, last access timestamp, and data retention time for each physical block; At the unit granularity, the standard deviation and range of the P / E cycle counts of all blocks within the unit are statistically analyzed and aggregated to quantify the wear dispersion within the unit, while also recording the total data throughput of the unit. At the chip level, monitor the overall average wear level, the difference between the largest and smallest wear blocks, and the rate of increase in potential error bit rate due to read interference and programming interference.

3. The global optimization method for solid-state drive wear leveling based on multi-granularity data features as described in claim 2, characterized in that, All collected multidimensional data features were normalized and transformed into dimensionless feature scores that could be compared laterally, including: By using the maximum-minimum normalization or Z-score normalization mathematical methods, raw data from different monitoring sources are uniformly mapped to a preset dimensionless standard interval, eliminating the comparison barriers between different physical quantities, so that the wear dispersion within the same unit can be directly and meaningfully compared with the temperature reading of a chip. The normalized basic feature values ​​are assigned dynamically changing weight coefficients based on their contribution to the wear impact under different life cycles and workloads, and a comprehensive feature score is synthesized through a weighted fusion algorithm. The feature score represents the current instantaneous state of the monitored object. When arranged in a time sequence, it forms a dynamically evolving data stream. By analyzing the trend and pattern of the data stream, a multi-level health and load feature profile is constructed, from block to unit to chip, which reflects both the real-time state and predicts future risks.

4. The global optimization method for solid-state drive wear leveling based on multi-granularity data features as described in claim 3, characterized in that, In S102, the lightweight prediction model predicts the wear rate and potential risk level of each block, cell, and even chip within a specific future time window based on historical characteristic data trends, including: The lightweight prediction model establishes an independent time series database for each monitoring point, stores its historical feature scores at fixed periods, and uses a sliding window mechanism to select data from the most recent N periods as samples. It then uses a lightweight exponential smoothing method or ARIMA model for analysis to capture the short-term trend and periodic pattern of wear changes, where N is greater than 1. The trained prediction model extrapolates the time series of each monitoring point and outputs its predicted feature score for the next M periods. By comparing it with a preset dynamic threshold, the expected wear rate for that period is calculated, and finally a risk prediction report covering all areas is generated. The risk prediction report is mapped to a specific risk level, where M is less than N.

5. The global optimization method for solid-state drive wear leveling based on multi-granularity data features as described in claim 4, characterized in that, By comparing the data with a preset dynamic threshold, the expected wear rate within that time period is calculated, ultimately generating a risk prediction report covering all areas of the entire system. This risk prediction report is mapped to specific risk levels, including: The preset dynamic threshold is not a fixed value, but is determined based on its own historical performance. The threshold is dynamically calculated based on historical data, so that the risk judgment standard can adapt to the characteristics of different solid-state drives and the overall health status of the entire drive, thereby calculating the statistically significant expected wear rate within the future time window. The predicted feature score and wear rate of each solid-state drive are compared with its corresponding dynamic threshold. Based on the magnitude and duration of exceeding the threshold, a predefined risk mapping matrix is ​​used to classify them into low, medium, and high risk levels. Finally, a structured and visualized global risk prediction report is automatically generated, which accurately identifies all potential failure points that need to be given priority and dealt with across the entire disk.

6. The global optimization method for solid-state drive wear leveling based on multi-granularity data features as described in claim 5, characterized in that, In S103, the decision variables of the global optimization mathematical model are the mapping relationship between logical addresses and physical addresses during the next garbage collection or host write operation, thereby optimizing the global mathematical model. The mapping relationship constraints include: The wear of each block must not exceed the safety threshold in a short period of time, the write amplification caused by a single operation must not exceed the set upper limit, and real-time performance response requirements must be met. The global mathematical model seeks a Pareto optimal solution by balancing the immediate wear leveling effect with long-term performance overhead, ensuring the most effective global lifetime extension at an acceptable performance cost.

7. The global optimization method for solid-state drive wear leveling based on multi-granularity data features as described in claim 6, characterized in that, In S104, a decision scheme is obtained through an optimized global mathematical model, which is then transformed into operation instructions that can be executed by the flash memory conversion layer. Based on these operation instructions, a specific task queue is dynamically generated, including hot and cold data identification, data block merging, and physical address remapping. The global optimization model obtains the optimal decision variables through the solver, namely the best mapping relationship from logical page to physical page and the target location of the data to be migrated. According to the FTL protocol specification, the optimal decision variables are transformed into a series of atomic operation instructions and written to the physical block and the mapping table entries from logical address LBA to physical address PBA are updated. Upon receiving these atomic operation instructions, the system instantly identifies hot and cold data based on data access frequency statistics, dynamically groups and sorts the instructions, prioritizes the migration of hot data in high-risk areas to ensure real-time performance, arranges cold data merging and remapping operations to be executed in a low-priority queue in the background, and dynamically adjusts the execution rate and concurrency of the queue according to the real-time load to balance wear leveling efficiency and I / O performance response.

8. The global optimization method for solid-state drive wear leveling based on multi-granularity data features as described in claim 7, characterized in that, In S105, after the decision-making plan is executed, the solid-state drive enters a continuous verification and adjustment cycle. The actual effect of this wear leveling operation is quantitatively evaluated by comparing the multi-granularity feature data collected before and after the strategy execution, including: After the decision-making plan is completed, a new round of comprehensive multi-granular data feature collection is immediately launched to obtain key indicators including the latest wear variance, maximum wear value, average wear degree and actual write amplification factor of the target area and surrounding related areas, and to accurately compare these real-time data after execution with the historical benchmark data recorded before execution. By calculating the changes in key performance indicators, the gain assessment results of this operation are quantitatively evaluated. The gain assessment results are directly input into the parameter self-adjustment engine as a feedback signal. The parameter self-adjustment engine drives the weight coefficients or strategy trigger thresholds in the global mathematical model to be dynamically fine-tuned.

9. A global optimization system for solid-state drive wear leveling based on multi-granularity data features, characterized in that, The system is used to execute the global optimization method for solid-state drive wear leveling according to any one of claims 1-8, the system comprising: The data feature monitoring and processing module is used to collect raw multidimensional data features related to wear in real time in the solid-state drive controller at the basic granularity of blocks, cells and chips, and to normalize and weight and fuse all collected multidimensional data features to generate feature scores that can be compared horizontally. The status assessment and prediction module is used to receive feature scores, calculate the instantaneous wear pressure score of each block through a weighted scoring algorithm, and establish a prediction model based on sliding window time series analysis to predict the wear rate and risk level of each granular unit within a specific future time window, and generate a risk heat map covering the entire system. The global optimization decision module is used to construct a global optimization mathematical model based on the risk heat map with the goal of maximizing the overall lifespan. The solver then obtains the optimal decision scheme for the mapping relationship between logical addresses and physical addresses. The strategy execution and scheduling module is used to convert decision schemes into executable operation instructions for the flash conversion layer, dynamically generate data block merging and physical address remapping task queues, and intelligently schedule foreground and background balanced operations according to system load status. The effect verification and parameter adjustment module is used to quantitatively evaluate the operation effect by comparing multi-granular feature data before and after the strategy is executed, and adaptively adjust the feature weights, penalty term coefficients and strategy trigger thresholds based on the evaluation results.

10. The global optimization system for solid-state drive wear leveling based on multi-granularity data features as described in claim 9, characterized in that, The data feature monitoring and processing module includes: The data acquisition unit is used to collect raw, multi-dimensional data features related to wear in real time and continuously, using blocks, units, and chips as the basic granularity. The data normalization unit is used to receive raw multidimensional data from the data acquisition unit and, through the maximum-minimum normalization or Z-score normalization mathematical method, uniformly map the raw data with different dimensions and ranges to a preset dimensionless standard interval, thereby eliminating the comparison barriers between different physical quantities. The feature fusion unit receives normalized data, assigns dynamically changing weight coefficients to the data features based on their contribution to wear impact under different lifecycles and workloads, and synthesizes a comprehensive feature score through a weighted fusion algorithm.

Images