Solid state disk read-write instruction intelligent scheduling method and system based on strategy optimization

Abstract

The present application relates to the technical field of computer control, and particularly provides a solid state disk read-write instruction intelligent scheduling method and system based on policy optimization, comprising obtaining the global state features of the solid state disk at the current time, the global state features including the depth distribution of the read-write queue, the occupation state of the flash channel, the erase-write number mapping diagram of each physical block and the trigger intensity of the current garbage collection; inputting the global state features into a preset policy optimization scheduling model, outputting the scheduling action at the current time through the policy network of the policy optimization scheduling model, the scheduling action including scheduling instructions from the read queue, scheduling instructions from the write queue, executing the garbage collection operation or keeping the current operation unchanged; controlling the controller of the solid state disk to execute the scheduling action, and obtaining the state changes and performance index data of the solid state disk caused after executing the scheduling action; the present application effectively reduces the negative impact of read interference on the service life, and improves the reliability of long-term operation of the solid state disk.

Description

Technical Field

[0001] This invention relates to the field of computer control technology, and in particular to a method and system for intelligent scheduling of solid-state drive read / write instructions based on strategy optimization. Background Technology

[0002] With the rapid development of big data, cloud computing, and artificial intelligence technologies, data centers are placing increasingly stringent performance demands on storage systems. Solid-state drives (SSDs), with their advantages of high read / write speeds, low power consumption, and high reliability, have gradually replaced traditional hard disk drives (HDDs) as the mainstream storage device. However, the performance of SSDs is highly dependent on the internal controller's scheduling strategy for read / write commands. A good scheduling strategy can effectively reduce read / write latency, increase throughput, and extend the lifespan of SSDs.

[0003] Prior art 1, Chinese Patent Application No.: CN201611202274.4, discloses a read / write scheduling method and apparatus in a solid-state drive (SSD). This invention discloses a read / write scheduling method in an SSD, comprising: storing read messages and write messages into a read message queue and a write message queue, respectively; adjusting the priorities of the read message queue and the write message queue; and processing the read messages and write messages. This invention also discloses a read / write scheduling apparatus in an SSD, comprising a storage module, an adjustment module, and a message processing module. This invention avoids the problem that in a message queue, due to a fixed read / write order, the write operation time is significantly longer than the read operation time, and the channel flash memory controller cannot process subsequently written messages for a long period. It achieves real-time, dynamic management of message queue priorities and the number of read and write messages processed, shortens the read latency, and improves the overall performance of the SSD. Prior art two, Chinese patent application number: CN202510924575.0, discloses an artificial intelligence-based intelligent read / write optimization method for solid-state drives (SSDs), including an intelligent read / write scheduler, adaptive wear leveling, read / write performance optimization, fault prediction and health management, and security protection and efficiency improvement. By analyzing usage patterns, optimizing caching strategies, and predicting faults through machine learning algorithms, it improves read / write performance and data security, significantly enhancing the read / write speed, stability, and security of SSDs to meet diverse application scenarios. It also predicts faults to prevent SSD failures. Through AI-driven intelligent read / write optimization, SSDs can achieve synergistic optimization of performance, lifespan, and energy efficiency under complex loads, providing highly reliable storage solutions for data centers, edge computing, and other scenarios. Combining machine learning, strategy optimization, and hardware collaboration technologies, it can improve SSD performance, lifespan, and energy efficiency. Prior art three, Chinese patent application number: CN201711068250.9, discloses a method for improving the hybrid read / write performance of solid-state drives. Its characteristic is that write requests are directly added to the request queue in the order they are received. The entry of a write request into the queue and the completion of the write request operation trigger an update of the block page information table at the current write point. Read requests undergo two levels of arbitration to determine whether a high-priority response can be given and the optimal insertion position of that high-priority response in the request queue: the first level of arbitration determines whether a high-priority response to the read request can be given based on the block page information table, and the second level of arbitration indexes the insertion position information to obtain the insertion position of the request in the request queue; the second level is based on the effectiveness of the first level of arbitration. By introducing two levels of arbitration, a suitable insertion position for the read request can be quickly found, improving read response efficiency. The overall scheduling process consumes few resources, is simple to implement but highly efficient, and can effectively improve the overall performance in hybrid read / write scenarios. Current technologies 1, 2, and 3 face challenges in efficiently applying strategy optimization to solid-state drive (SSD) read / write instruction scheduling. These challenges include ensuring real-time scheduling decisions while handling a high-dimensional, continuous internal state space and balancing short-term latency with long-term performance metrics (such as wear leveling). Therefore, this invention provides a strategy-optimized intelligent scheduling method and system for SSD read / write instructions. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: In one aspect, the present invention provides a strategy-optimized intelligent scheduling method for solid-state drive read / write instructions, including obtaining the global state characteristics of the solid-state drive at the current moment, wherein the global state characteristics include the depth distribution of the read / write queue, the occupancy status of the flash memory channel, the erase / write count mapping of each physical block, and the current garbage collection trigger strength. The global state features are input into a preset policy optimization scheduling model, and the policy network of the policy optimization scheduling model outputs the scheduling action at the current moment. The scheduling action includes scheduling instructions from the read queue, scheduling instructions from the write queue, performing garbage collection operations, or keeping the current operation unchanged. The controller that controls the solid-state drive executes the scheduling action and obtains the state changes and performance index data of the solid-state drive caused by the execution of the scheduling action; The reward value at the current moment is calculated based on the performance index data, and sample data is constructed using the reward value, the global state features, the scheduling action, and the global state features at the next moment. The policy optimization scheduling model is then iteratively trained using the sample data.

[0005] In one optional implementation, calculating the reward value at the current moment based on the performance index data includes: Determine the performance parameters at the current moment, including average read / write response latency, read / write throughput, and additional overhead caused by read / write switching; The performance parameters are weighted according to preset weighting coefficients to generate a comprehensive reward value; The formula for calculating the reward value is as follows: Where, in the formula, This represents the calculated reward value. This indicates the average response time within the current time window. This represents the throughput within the current time window. This indicates the performance loss caused by frequent switching between read and write operations. , and These represent the maximum normalized thresholds for the corresponding parameters, and α, β, and γ are the preset weighting coefficients for balancing delay, throughput, and switching loss, respectively, with α+β+γ=1.

[0006] In an optional implementation, before obtaining the global state characteristics of the solid-state drive at the current moment, the method further includes: The original flash memory physical state data is matrix-encoded to transform the block state of the physical plane into a two-dimensional state matrix. The policy optimization scheduling model is constructed, which includes a feature extraction network and a decision network. The feature extraction network uses a convolutional neural network to extract the spatial features of the two-dimensional state matrix, and the decision network uses a fully connected layer combined with an attention mechanism to process the queue state features. The training process of the policy optimization scheduling model includes a policy network and a value network; The policy network is used to output the probability distribution of each scheduling action based on the input global state features, and to select a scheduling action based on the probability distribution. The value network is used to evaluate the state value function estimate corresponding to the current state and the scheduling action taken; The value network is updated by minimizing the error between the estimated state value function and the actual cumulative return, and the gradient of the value network is used to guide the parameter update of the policy network.

[0007] In one optional implementation, updating the value network and guiding the parameter updates of the policy network specifically includes: Calculate the time difference error, which is the sum of the reward value at the current moment and the estimated state value at the next moment after processing with the discount factor, minus the estimated state value at the current moment; The parameters of the value network are updated by backpropagation using the time difference error. The parameters of the policy network are updated by backpropagation using the product of the log probability of the time difference error and the gradient. It also includes the step of building an experience replay pool: The sample data consisting of the old state, scheduling action, reward value and new state is stored in the experience replay pool; During each training iteration, a preset number of sample data are randomly selected from the experience replay pool as training batches. The parameters of the policy optimization scheduling model are updated using the sample data in the training batch to break the temporal correlation between sample data and improve the stability of training.

[0008] In one optional implementation, a preset number of sample data are randomly drawn from the experience replay pool, including: Calculate the absolute value of the time difference error for each sample data in the experience replay pool; The sampling priority of each sample data is determined based on the magnitude of the absolute value of the time difference error. The sample data with the larger absolute value of the time difference error has a higher probability of being sampled. The sample data in the experience replay pool is extracted with priority according to the sampling priority, so as to improve the model's utilization efficiency of rare or high-value states. It also includes action exploration and the use of balance strategies: In the early stages of model training, a large exploration rate is used to control the solid-state drive controller to randomly execute scheduling actions in order to traverse as much of the state space as possible. As the number of training steps increases, the exploration rate is gradually reduced according to a preset decay function. In the later stages of model training, the learned optimal strategy is mainly used to select scheduling actions. During the model deployment phase, the parameters of the strategy-optimized scheduling model are frozen, and deterministic scheduling actions are output directly based on the input state.

[0009] In an optional implementation, obtaining the global state characteristics of the solid-state drive at the current moment further includes: Monitor the waiting time of currently pending read / write commands; If there is an instruction whose waiting time exceeds the preset time threshold, the priority flag of the instruction will be set when generating global state features; When outputting scheduling actions, the strategy optimization scheduling model adjusts the weight of instructions with the priority identifier to increase their probability of being selected for scheduling. The global state features also include the time-series features of the read / write instruction stream sent by the host. Extract the time series features of the read and write instruction stream within a preset time window, and then input the time series features and flash memory internal physical state features into the strategy optimization scheduling model. The strategy optimization scheduling model combines the time series characteristics to predict the load pressure within the future time window, and adjusts the current garbage collection trigger frequency in advance based on the prediction results.

[0010] In one optional implementation, adjusting the current garbage collection trigger frequency in advance based on the prediction results includes: When the write load pressure is predicted to be greater than the preset first pressure threshold within a future time window, the scheduling action is controlled to select to perform an active garbage collection operation in order to free up free space. When the read load pressure is predicted to be greater than the preset second pressure threshold within a future time window, the scheduling action is suppressed to select the garbage collection operation to ensure that the flash channel resources are used to respond to read and write requests.

[0011] In one optional implementation, during the operation of the solid-state drive, an exploration mechanism is activated within a preset safe time period to collect online sample data under the current actual operating conditions; The deployed strategy optimization scheduling model is updated in small batches using the online sample data. Calculate the loss value of the model on the validation set before and after the update. If the loss value decreases, replace the current model parameters with the updated model parameters; otherwise, keep the current model parameters unchanged. The policy optimization agent model employs a dual-deep Q-network architecture, comprising an evaluation network and a target network. The evaluation network selects actions, and the target network calculates the target Q-value to stabilize the training process. When updating network parameters, the mean squared error loss between the evaluation network's predicted Q-value and the target Q-value is minimized. The convergence process of the loss function follows the following update rules: in, To evaluate the network parameters, For the parameters of the target network, Let D be the discount factor, and D be the batch data set sampled from the playback buffer. These represent the state, action, reward, and next state, respectively.

[0012] In one optional implementation, when calculating the reward value, the impact of read interference on the solid-state drive is further considered, and the specific reward correction formula is as follows: in, This represents the corrected system reward value. Indicates the base reward value. Indicates the number of physical blocks affected. This represents the raw read error rate of the i-th physical block. This indicates the preset bit error rate safety threshold. This is the penalty coefficient; the reward value is deducted when the read error rate exceeds the threshold.

[0013] Another aspect of the present invention provides a strategy-optimized intelligent scheduling system for solid-state drive read / write instructions, comprising: The status acquisition module is used to monitor and acquire the global status characteristics of the solid-state drive in real time. The intelligent decision-making module has a pre-trained strategy optimization scheduling model built in, which is used to receive the global state features sent by the state acquisition module and output scheduling actions. An execution control module, connected to the intelligent decision-making module and the solid-state drive controller, is used to convert the scheduling action into control commands and send them to the solid-state drive controller for execution; The feedback update module is used to obtain the performance indicators after the scheduling action is executed, calculate the reward value, and periodically update the model in the intelligent decision-making module using the generated sample data.

[0014] Effects of the invention: This invention introduces a strategy optimization algorithm into the core controller of a solid-state drive (SSD), constructing an intelligent scheduling agent. This agent continuously learns and optimizes the scheduling strategy for read and write instructions through constant interaction with the storage system environment. Unlike existing technologies that rely on fixed-rule arbitration, the scheduling strategy of this invention can automatically adjust as the SSD media ages and data writing patterns change, possessing the adaptive capability to maintain optimal performance in complex and dynamic environments.

[0015] By introducing a carefully designed reward function, this invention enables the intelligent scheduling agent to make decisions that go beyond simply focusing on immediate read / write requests. Instead, it considers the long-term impact of operations on the internal physical state of the SSD (such as garbage collection pressure and read interference risks). This decision-making mechanism, based on global long-term benefits, effectively avoids write amplification and GC blocking issues caused by over-optimizing read requests, achieving an optimal balance between I / O performance and SSD lifespan.

[0016] For mixed read / write loads common in scenarios such as data centers, the intelligent scheduling agent of this invention can accurately identify the current load characteristics and dynamically adjust the "queueing" strategy and concurrency of read commands in the queue. Compared with the two-level arbitration of the existing technology, this invention can predict the consequences of command execution at a finer granularity, thereby maximizing the utilization of flash channel bandwidth while ensuring that the read response time meets the requirements, significantly improving the overall throughput and service quality of SSDs in mixed scenarios.

[0017] This invention employs strategy optimization techniques such as experience replay pooling and target networks. While ensuring model training stability, it effectively controls computational resource consumption, enabling deployment on resource-constrained SSD controller chips. Furthermore, through an online learning mechanism, the system can quickly adapt to new application scenarios and complete strategy upgrades without manual intervention, significantly improving the intelligence level and operational efficiency of the storage system. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart for obtaining the global state characteristics of a solid-state drive at the current moment, as provided in Embodiment 2 of the present invention. Figure 3 This is a system architecture diagram provided in Embodiment 3 of the present invention; Figure 4 This is a block diagram of the electronic device provided in Embodiment 4 of the present invention; Figure 5 This is a block diagram of a computer-readable storage medium provided in Embodiment 4 of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Hereinafter, the terms "first," "second," etc., are used for descriptive convenience only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0021] In this invention, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed mechanical connection, a detachable mechanical connection, or an integral part; or, "connection" can be a direct connection or an indirect connection through an intermediate medium. Furthermore, unless otherwise explicitly specified and limited, the term "coupling" should be interpreted broadly. For example, "coupling" can be a direct electrical connection, such as physical contact and electrical conduction between two components; it can also be understood as an electrical connection between different components in a circuit structure through physical lines capable of transmitting electrical signals, such as copper foil or wires on a printed circuit board (PCB), to transmit electrical signals; or, "coupling" can be an indirect electrical connection between two components through an intermediate medium; or, "coupling" can be an electrical connection between two components in a non-contact manner, such as an electrical connection between two components using capacitive coupling to transmit electrical signals.

[0022] In this embodiment of the invention, directional terms such as "up," "down," "left," and "right" may be defined relative to the orientation of the components shown in the accompanying drawings. It should be understood that these directional terms can be relative concepts, used for relative description and clarification, and can change accordingly depending on the orientation of the components in the accompanying drawings.

[0023] Example 1: like Figure 1 As shown, this embodiment of the invention provides a strategy-optimized intelligent scheduling method for solid-state drive read / write instructions, comprising the following steps: Step S100: Obtain the global state characteristics of the solid-state drive at the current moment. The global state characteristics include the depth distribution of the read and write queues, the occupancy status of the flash channels, the erase and write count mapping of each physical block, and the current garbage collection trigger strength. Step S200: Input the global state features into the preset policy optimization scheduling model, and output the scheduling action at the current time through the policy network of the policy optimization scheduling model. The scheduling action includes scheduling instructions from the read queue, scheduling instructions from the write queue, performing garbage collection operations, or keeping the current operation unchanged. Step S300: Control the solid-state drive controller to execute the scheduling action, and obtain the solid-state drive status changes and performance index data caused by the execution of the scheduling action; Step S400: Calculate the reward value at the current moment based on the performance index data, and construct sample data using the reward value, the global state features, the scheduling action, and the global state features at the next moment. Use the sample data to iteratively train the policy optimization scheduling model.

[0024] In the above embodiments, the step of obtaining the global state characteristics of the solid-state drive at the current moment includes: Queue depth distribution acquisition: Real-time scanning of the read and write command queues inside the SSD controller, counting the number of commands currently waiting to be processed, and classifying them according to command type and the flash channel to which the logical address belongs, forming a multi-dimensional queue depth vector. Simultaneously, the waiting time of each command is monitored; if the waiting time of any command exceeds a preset threshold (e.g., 10ms), it is marked as high priority in the status.

[0025] Channel occupancy status monitoring: Poll all flash memory transfer channels to check if there is any ongoing data transfer or programming / erasing operation, and generate a binary or normalized channel occupancy status vector (e.g., [occupied, idle, occupied, occupied]).

[0026] Physical block erase / write count mapping generation: Read the wear leveling table in the Flash Translation Layer (FTL) to obtain the current erase / write count (P / ECycle) of all physical blocks. Encode this raw data into a matrix, transforming the block state of the physical plane into a two-dimensional matrix (e.g., a 32x32 matrix), where the value of each element represents the normalized wear level of the corresponding physical block.

[0027] Garbage Collection (GC) Trigger Intensity Assessment: Detect the current state of the garbage collection mechanism, calculate the remaining proportion of free blocks and the number of dirty blocks, assess the urgency of the current system performing garbage collection, and convert it into a value between 0 and 1 as the trigger intensity feature.

[0028] Time series feature extraction (optional): Extract the density and type ratio of read and write command streams sent by the host within a preset time window (e.g., the past 10ms) to form time series features, which can be used to help predict future load.

[0029] By combining multi-dimensional hardware states (channels, block wear) with software states (queues, instruction flow), a comprehensive and high-precision decision-making basis is provided for the intelligent scheduling model, solving the problem of narrow vision caused by traditional scheduling relying solely on a single queue depth. The addition of time-series features and GC trigger intensity enables the model to "anticipate" future pressures (such as an impending write flood), thus reacting in advance and avoiding reactive responses. Prioritizing long-waiting instructions ensures the system's immediate response capability to latency-sensitive tasks.

[0030] Example 2: like Figure 1 As shown, based on Example 1, the steps provided in this embodiment of the invention include calculating the reward value at the current moment based on the performance index data, including: Determine the performance parameters at the current moment, including average read / write response latency, read / write throughput, and additional overhead caused by read / write switching; The performance parameters are weighted according to preset weighting coefficients to generate a comprehensive reward value; The formula for calculating the reward value is as follows: Where, in the formula, This represents the calculated reward value. This indicates the average response time within the current time window. This represents the throughput within the current time window. This indicates the performance loss caused by frequent switching between read and write operations. , and These represent the maximum normalized thresholds for the corresponding parameters, and α, β, and γ are the preset weighting coefficients for balancing delay, throughput, and switching loss, respectively, with α+β+γ=1.

[0031] In an optional implementation, before obtaining the global state characteristics of the solid-state drive at the current moment, the method further includes: The original flash memory physical state data is matrix-encoded to transform the block state of the physical plane into a two-dimensional state matrix. The policy optimization scheduling model is constructed, which includes a feature extraction network and a decision network. The feature extraction network uses a convolutional neural network to extract the spatial features of the two-dimensional state matrix, and the decision network uses a fully connected layer combined with an attention mechanism to process the queue state features. The training process of the policy optimization scheduling model includes a policy network and a value network; The policy network is used to output the probability distribution of each scheduling action based on the input global state features, and to select a scheduling action based on the probability distribution. The value network is used to evaluate the state value function estimate corresponding to the current state and the scheduling action taken; The value network is updated by minimizing the error between the estimated state value function and the actual cumulative return, and the gradient of the value network is used to guide the parameter update of the policy network.

[0032] In one optional implementation, updating the value network and guiding the parameter updates of the policy network specifically includes: Calculate the time difference error, which is the sum of the reward value at the current moment and the estimated state value at the next moment after processing with the discount factor, minus the estimated state value at the current moment; The parameters of the value network are updated by backpropagation using the time difference error. The parameters of the policy network are updated by backpropagation using the product of the log probability of the time difference error and the gradient. It also includes the step of building an experience replay pool: The sample data consisting of the old state, scheduling action, reward value and new state is stored in the experience replay pool; During each training iteration, a preset number of sample data are randomly selected from the experience replay pool as training batches. The parameters of the policy optimization scheduling model are updated using the sample data in the training batch to break the temporal correlation between sample data and improve the stability of training.

[0033] In one optional implementation, a preset number of sample data are randomly drawn from the experience replay pool, including: Calculate the absolute value of the time difference error for each sample data in the experience replay pool; The sampling priority of each sample data is determined based on the magnitude of the absolute value of the time difference error. The sample data with the larger absolute value of the time difference error has a higher probability of being sampled. The sample data in the experience replay pool is extracted with priority according to the sampling priority, so as to improve the model's utilization efficiency of rare or high-value states. It also includes action exploration and the use of balance strategies: In the early stages of model training, a large exploration rate is used to control the solid-state drive controller to randomly execute scheduling actions in order to traverse as much of the state space as possible. As the number of training steps increases, the exploration rate is gradually reduced according to a preset decay function. In the later stages of model training, the learned optimal strategy is mainly used to select scheduling actions. During the model deployment phase, the parameters of the strategy-optimized scheduling model are frozen, and deterministic scheduling actions are output directly based on the input state.

[0034] In an optional implementation, obtaining the global state characteristics of the solid-state drive at the current moment further includes: Monitor the waiting time of currently pending read / write commands; If there is an instruction whose waiting time exceeds the preset time threshold, the priority flag of the instruction will be set when generating global state features; When outputting scheduling actions, the strategy optimization scheduling model adjusts the weight of instructions with the priority identifier to increase their probability of being selected for scheduling. The global state features also include the time-series features of the read / write instruction stream sent by the host. Extract the time series features of the read and write instruction stream within a preset time window, and then input the time series features and flash memory internal physical state features into the strategy optimization scheduling model. The strategy optimization scheduling model combines the time series characteristics to predict the load pressure within the future time window, and adjusts the current garbage collection trigger frequency in advance based on the prediction results.

[0035] In one optional implementation, adjusting the current garbage collection trigger frequency in advance based on the prediction results includes: When the write load pressure is predicted to be greater than the preset first pressure threshold within a future time window, the scheduling action is controlled to select to perform an active garbage collection operation in order to free up free space. When the read load pressure is predicted to be greater than the preset second pressure threshold within a future time window, the scheduling action is suppressed to select the garbage collection operation to ensure that the flash channel resources are used to respond to read and write requests.

[0036] In one optional implementation, during the operation of the solid-state drive, an exploration mechanism is activated within a preset safe time period to collect online sample data under the current actual operating conditions; The deployed strategy optimization scheduling model is updated in small batches using the online sample data. Calculate the loss value of the model on the validation set before and after the update. If the loss value decreases, replace the current model parameters with the updated model parameters; otherwise, keep the current model parameters unchanged. The policy optimization agent model employs a dual-deep Q-network architecture, comprising an evaluation network and a target network. The evaluation network selects actions, and the target network calculates the target Q-value to stabilize the training process. When updating network parameters, the mean squared error loss between the evaluation network's predicted Q-value and the target Q-value is minimized. The convergence process of the loss function follows the following update rules: in, To evaluate the network parameters, For the parameters of the target network, Let D be the discount factor, and D be the batch data set sampled from the playback buffer. These represent the state, action, reward, and next state, respectively.

[0037] In one optional implementation, when calculating the reward value, the impact of read interference on the solid-state drive is further considered, and the specific reward correction formula is as follows: in, This represents the corrected system reward value. Indicates the base reward value. Indicates the number of physical blocks affected. This represents the raw read error rate of the i-th physical block. This indicates the preset bit error rate safety threshold. This is the penalty coefficient; the reward value is deducted when the read error rate exceeds the threshold.

[0038] Among them, step S101: input the global state features into the model and output the scheduling action; Step S102: Control the execution of scheduling actions and obtain state changes and performance indicators; Step S103: Calculate the reward value, construct sample data, and iteratively train the model; In the above embodiments, feature extraction and fusion utilize a convolutional neural network to process a two-dimensional physical block erase / write count mapping map to extract spatial features (such as wear hotspot areas) inside the flash memory.

[0039] A fully connected layer combined with an attention mechanism is used to process queue states and time-series features. The attention mechanism automatically increases the weight of high-priority instructions.

[0040] The extracted spatial features are concatenated with the queue state features to form the final global state vector.

[0041] Policy network decision: Input the global state vector into the policy network of the policy optimization model (such as the Actor network or the evaluation network of DDQN).

[0042] Action probability output: The policy network outputs the probability distribution or Q-value of each candidate action. Candidate actions typically include: Action A: Dispatch instructions from the read queue (process read requests).

[0043] Action B: Dispatch instructions from the write queue (process write requests).

[0044] Action C: Perform garbage collection operation (actively move data).

[0045] Action D: Keep the current operation unchanged (suspend or wait).

[0046] Action selection: Based on the probability distribution and combined with an exploration-exploitation strategy (such as ε-greedy), a specific scheduling action is selected. In the early stages of training, the model may randomly try actions; in the later stages of training or deployment, it tends to select the optimal action with the highest probability.

[0047] Intelligently weighing multiple objectives, unlike simple "read-first" or "write-first" rules, deep neural networks can comprehensively consider multiple factors such as wear and tear, latency, and GC pressure, and make globally optimal dynamic decisions in complex nonlinear scenarios.

[0048] Automated feature learning, utilizing CNNs and attention mechanisms, enables the model to automatically discover complex features that are difficult for humans to design from raw state data (such as the implicit correlation between wear distribution and read / write performance), thus improving the accuracy of scheduling.

[0049] Adaptive to load changes: Regardless of whether the load is sequential or random read / write, the model can automatically adjust its output actions based on the input features without requiring manual rewriting of scheduling rules.

[0050] Instruction execution: The solid-state drive controller firmware receives the scheduling action output in step two, parses it into low-level hardware control signals (such as sending flash read page instructions, write page instructions, or block erase instructions), and drives the flash media to perform the corresponding operations.

[0051] State transition capture: After the action is completed or a fixed time step (TimeSlot) has elapsed, the system enters the next time step t+1. The operation in step one is executed again to collect the new global state characteristics St+1.

[0052] Performance metrics statistics: Within the time window of action execution, key performance indicators (KPIs) are statistically analyzed. Latency: Calculates the average response time of read and write commands.

[0053] Throughput: The amount of data processed per unit of time.

[0054] Overhead: Records the additional time lost due to switching between read and write operations.

[0055] Reliability (optional): Changes in read error rate after the execution of the action.

[0056] Closed-loop feedback mechanism: This step establishes a complete closed loop from "decision-making" to "results", which is the basis for the strategy optimization to be able to self-optimize.

[0057] Quantifying environmental impact: By accurately measuring the performance changes after an action is executed, the physical response of the hardware is transformed into data metrics that the model can understand, enabling the algorithm to objectively evaluate the merits of each decision.

[0058] Reward value calculation: Calculate the reward value R based on the performance indicators obtained in step three and the preset weighting coefficients.

[0059] The base reward is a weighted sum of latency (lower is better), throughput (higher is better), and switching overhead (lower is better). Samples are stored in an experience replay pool. During training, samples are prioritized based on their Temporal Difference Error (TDError), with "unexpected" or "high-value" samples being used to update the model first. The TDError is calculated as the current reward plus the discounted value of the next state estimate, minus the current state's value estimate. Backpropagation using the TDError updates the parameters of the policy network (determining action) and the value network (evaluating performance). If a Dual Deep Q-Network (DDQN) is used, the evaluation network selects actions, and the target network calculates the target Q-value to stabilize the training process.

[0060] Through continuous trial-and-error, evaluation-and-correction, the model can gradually learn the optimal strategy to adopt under various complex conditions. As training data accumulates, the scheduling strategy becomes increasingly sophisticated. By employing priority experience replay and DDQN architecture, the temporal correlation of sample data is effectively broken, solving the problem of divergence in policy optimization training and accelerating model convergence. By introducing a discount factor, the model not only focuses on current performance metrics but also considers the long-term impact of current actions on the future (e.g., sacrificing some performance for GC now is to prevent future performance avalanche), thereby achieving a balance and optimization of the system's long-term performance.

[0061] In one embodiment, the global state characteristics of the solid-state drive at the current moment are collected in real time: The solid-state drive controller collects its global status every 1ms, and the specific data collected includes: Read / Write Queue Status: The current depth of the read instruction queue is 12, and the depth of the write instruction queue is 26. The queue depth distribution of each channel is obtained by splitting by channel. At the same time, the waiting time of each instruction is monitored: There is one read instruction in this batch whose waiting time exceeds the preset threshold of 10ms. Therefore, the priority flag corresponding to this instruction is set to 1 in the global features.

[0062] Flash Channel Status: Of the four flash channels, two are currently busy and two are idle, encoded as an occupancy vector of [1,1,0,0].

[0063] Physical block erase / write characteristics: This involves pre-constructing four 32×32 erase / write count normalization matrices, simultaneously reading the read error rate of all current physical blocks, and counting the number of physical blocks with error rates exceeding the security threshold along with their corresponding error rates.

[0064] Running status: Passive garbage collection is currently triggered with a trigger strength of 0.6 (normalized to a range of 0-1, where a larger value indicates a higher amount of garbage collection tasks being processed); the read and write instruction density of the past 10 1ms time windows is extracted to obtain time series features: the density of upcoming write requests continues to increase, and the corresponding time series feature vector is encoded and input into the model.

[0065] Finally, all the above features are concatenated into a complete global state feature according to the model requirements.

[0066] The strategy optimization model outputs scheduling actions: Input global state features into the policy optimization scheduling model during training: The convolutional neural network first extracts the spatial features of the erase / write count mapping matrix to obtain the overall wear distribution features of the flash memory. The self-attention layer assigns higher weights to long-waiting read instructions with priority labels, and finally concatenates them to obtain the complete global features. The policy network outputs the probability of each action, combined with the current exploration rate: by the 2000th step of training, the exploration rate has decayed to 0.2, so the highest-scoring action output by the model is selected with a 98% probability and random exploration with a 2% probability. Finally, the model outputs the current scheduled action as executing proactive garbage collection: the model predicts that the write load pressure in the future time window will exceed the first pressure threshold (more than 30 write requests per ms) based on the time series features of the input instruction stream. Therefore, it chooses to trigger proactive garbage collection in advance to free up free blocks and avoid passive garbage collection from occupying channel resources and increasing latency when write requests surge in the future.

[0067] Meanwhile, for read instructions with priority indicators, the model has made a 1.5-fold weight correction to the output probability of the "schedule read queue instruction" action, ensuring that even if garbage collection is selected this time, the read instruction will be scheduled first in the future.

[0068] The controller executes scheduling actions and collects status changes and performance metrics: After the controller executes the active garbage collection action, it records the execution time of this action. At the same time, it waits for 1 time window (1ms) to collect the new global state S_{t+1} and counts the performance indicators within this time window: average read / write response latency Delay=0.8ms, throughput Throughput=3200MB / s, read / write switching overhead Switchcost=2 switches totaling 80μs, currently there are 3 physical blocks whose read error rate exceeds the safety threshold Rthreshold=1e-5, and the average error rate is 2.3e-5.

[0069] Calculate the reward value and build a training sample to update the model: Calculate the corrected reward value: First, substitute the basic reward formula to calculate the basic reward R: In this embodiment, the normalized maximum threshold Tmax = 5ms, Throughputmax = 4000MB / s, and Switchmax = 10 switches totaling 400μs. The parameters are: R=0.5*(-0.8 / 5)+0.35*(3200 / 4000)-0.15*(80 / 400)=-0.08+0.28-0.03=0.17 Then, a reading interference penalty correction is added, and the correction formula is substituted in: Rmodify = 0.17 - 0.02 * [(2.3e-5 - 1e-5) + (1.8e-5 - 1e-5) + (2.1e-5 - 1e-5)] ≈ 0.17 - 0.02 * 4.2e-5 ≈ 0.169, and the final reward value is taken as 0.17. Construct samples and store them in the experience replay pool The quintuple (S_t, a_t, Rmodify, S_{t+1}, done) is used as a training sample. The absolute value of the time difference error of the sample is calculated to be 0.82. The sampling priority is assigned according to the magnitude of the time difference error and stored in the experience replay pool.

[0070] Each training iteration samples 32 samples from the experience replay pool according to priority to form a training batch: samples with larger temporal difference errors have a higher sampling probability. This allows for prioritizing the learning of high-value rare state samples, improving training efficiency. This embodiment adopts the DDQN update rule: the evaluation network selects actions, the target network calculates the target Q-value, and the parameters are updated by minimizing the mean squared error between the evaluation Q-value and the target Q-value. Every 1000 iterations, the evaluation network parameters are synchronized to the target network to stabilize the training process. Simultaneously, during the update process, backpropagation of the temporal difference error is used to update the parameters of both the policy network and the value network, breaking the temporal correlation of samples and preventing model training divergence.

[0071] Explore - Utilize Balance and Continuous Online Updates: During the model training phase, the exploration rate decay rule is strictly followed: in the early stage of training (the first 1000 steps), the exploration rate is above 0.8, and the controller will randomly select actions with a high probability to traverse the state space under different load combinations as much as possible; as the number of training steps increases, the exploration rate is gradually reduced, and after 5000 steps, the exploration rate drops to 0.01. The model mainly relies on the learned policy to select actions and complete the training convergence.

[0072] After the model is deployed to the solid-state drive for official operation, it will also be continuously iterated online: low-frequency exploration will be enabled only during safe periods when the solid-state drive is idle (such as a continuous 10-second window when the system idle rate exceeds 90%) to collect online samples under actual operating load. Gradient updates will be performed using a small batch (8 samples) each time, and the verification loss before and after the update will be calculated. If the loss decreases, the new parameters will be retained; otherwise, the parameters will be rolled back to ensure that the model can adapt to the user's actual usage load and will not experience performance degradation.

[0073] Example 3: like Figure 3 As shown in Example 1, this embodiment of the invention provides a strategy-optimized intelligent scheduling system for solid-state drive read / write instructions, comprising: The status acquisition module is used to monitor and acquire the global status characteristics of the solid-state drive in real time. The intelligent decision-making module has a pre-trained strategy optimization scheduling model built in, which is used to receive the global state features sent by the state acquisition module and output scheduling actions. An execution control module, connected to the intelligent decision-making module and the solid-state drive controller, is used to convert the scheduling action into control commands and send them to the solid-state drive controller for execution; The feedback update module is used to obtain the performance indicators after the scheduling action is executed, calculate the reward value, and periodically update the model in the intelligent decision-making module using the generated sample data.

[0074] In the above embodiments, the corresponding intelligent scheduling system for solid-state drive read / write instructions is divided into four functional modules, specifically deployed as follows: Status acquisition module: The hardware monitoring unit embedded in the solid-state drive controller automatically collects all status data every 1ms, including read / write queue, channel occupancy, erase / write count, bit error rate, instruction wait time, and instruction stream sequence. After encoding, it outputs the data to the intelligent decision module. Intelligent decision-making module: The computing unit deployed in the main control firmware has a built-in trained policy optimization scheduling model, which can output the optimal scheduling action within 10μs after receiving state features. Execution control module: Converts the scheduling actions output by the model into hardware control instructions that the main controller can recognize, and directly controls the flash memory channel to perform corresponding read, write or garbage collection operations; Feedback and Update Module: Responsible for collecting performance metrics after action execution, calculating reward values, storing training samples, and triggering model parameter updates during idle periods to achieve continuous model optimization.

[0075] This embodiment compares two schemes—traditional fixed-priority scheduling (read-first scheduling) and traditional dynamic scheduling—on the same consumer-grade solid-state drive hardware, and obtains the following results under different load scenarios: Performance improvement under mixed random read / write loads: Under a typical desktop mixed load of 70% reads and 30% writes, the average read / write response latency of this solution is 0.72ms, which is 40.5% lower than the 1.21ms of traditional read-first scheduling. Random 4K read / write throughput is increased by 18.2%, and the overhead of switching is reduced by 29% due to the reduction of meaningless read / write switching. Improved adaptability to sudden load surges: For scenarios involving sudden large file writes, this solution can predict the surge in write load in advance through the instruction stream sequence, triggering proactive garbage collection to free up space. The fluctuation range of sudden peak throughput is reduced from 35% in traditional solutions to 8%, avoiding the performance drop problem caused by "garbage collection blocking read and write after a write surge" in traditional solutions. Solid-state drive lifespan optimization: This solution incorporates a read interference bit error rate penalty into the reward function. The model will actively avoid frequent read operations on high-wear blocks. After running continuously for 1000 hours, the average bit error rate of the solid-state drive is 17% lower than that of the traditional solution, effectively reducing the negative impact of read interference on lifespan and improving the long-term reliability of the solid-state drive. Adapt to different load characteristics: Because the model supports continuous online learning, it can automatically adapt and adjust the scheduling strategy for different user habits (such as heavy downloads, heavy gaming, light office workloads, etc.), without the need for manual parameter configuration as with traditional fixed-rule scheduling, making it more versatile.

[0076] Example 4: Figure 4 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention is shown.

[0077] The electronic device may include a central processing unit / microprocessor / main control chip, etc. 4; and a storage medium 5, coupled to the central processing unit / microprocessor / main control chip, etc. 4, and storing computer-executable instructions therein for performing the steps of various methods of embodiments of the present invention when executed by the processor.

[0078] The central processing unit / microprocessor / main control chip, etc., can include, but are not limited to, one or more processors or microprocessors.

[0079] Storage medium 5 may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (e.g., hard disk, floppy disk, solid-state drive, removable disk, CD-ROM, DVD-ROM, Blu-ray disc, etc.).

[0080] In addition, the electronic device may also include (but is not limited to) a data bus 6, an input / output bus / external bus / device bus 7, a display 8, and input / output devices 9 (e.g., keyboard, mouse, speaker, etc.).

[0081] The central processing unit / microprocessor / main control chip, etc. 4 can communicate with external devices (8, 9, etc.) via I / O bus 7 through wired or wireless network (not shown).

[0082] The storage medium 5 may also store at least one computer-executable instruction for performing the steps of various functions and / or methods in the embodiments described herein when the central processing unit / microprocessor / main control chip, etc., 4 is running.

[0083] In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product, wherein one or more computer-executable instructions are executed by a processor to perform the steps of the various functions and / or methods in the embodiments described herein.

[0084] Figure 5 A schematic diagram of a computer-readable storage medium according to an embodiment of the present invention is shown.

[0085] like Figure 5 As shown, the non-transitory computer-readable storage medium 11 stores instructions, such as computer-readable instructions 10. When the computer-readable instructions 10 are executed by a processor, the various methods described above can be performed. The non-transitory computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-transitory non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, the non-transitory computer-readable storage medium 11 can be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions 10 stored on the computer-readable storage medium 11, the various methods described above can be performed.

[0086] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0087] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0088] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0089] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods of the various embodiments of this invention through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0090] The above 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 embodiments of the present invention.

Claims

1. A strategy-optimized intelligent scheduling method for solid-state drive read / write instructions, characterized in that, Includes the following steps: Obtain the global state characteristics of the solid-state drive at the current moment. The global state characteristics include the depth distribution of the read and write queues, the occupancy status of the flash channels, the erase and write count mapping of each physical block, and the current garbage collection trigger strength. The global state features are input into a preset policy optimization scheduling model, and the policy network of the policy optimization scheduling model outputs the scheduling action at the current moment. The scheduling action includes scheduling instructions from the read queue, scheduling instructions from the write queue, performing garbage collection operations, or keeping the current operation unchanged. The controller that controls the solid-state drive executes the scheduling action and obtains the state changes and performance index data of the solid-state drive caused by the execution of the scheduling action; The reward value at the current moment is calculated based on the performance index data, and sample data is constructed using the reward value, the global state features, the scheduling action, and the global state features at the next moment. The policy optimization scheduling model is then iteratively trained using the sample data.

2. The intelligent scheduling method for solid-state drive read / write instructions based on strategy optimization as described in claim 1, characterized in that, The reward value at the current moment is calculated based on the performance metric data, including: Determine the performance parameters at the current moment, including average read / write response latency, read / write throughput, and additional overhead caused by read / write switching; The performance parameters are weighted according to preset weighting coefficients to generate a comprehensive reward value; The formula for calculating the reward value is as follows: Where, in the formula, This represents the calculated reward value. This indicates the average response time within the current time window. This represents the throughput within the current time window. This indicates the performance loss caused by frequent switching between read and write operations. , and These represent the maximum normalized thresholds for the corresponding parameters, and α, β, and γ are the preset weighting coefficients for balancing delay, throughput, and switching loss, respectively, with α+β+γ=1.

3. The intelligent scheduling method for solid-state drive read / write instructions based on strategy optimization as described in claim 1, characterized in that, Before obtaining the global state characteristics of the solid-state drive at the current moment, the method further includes: The original flash memory physical state data is matrix-encoded to transform the block state of the physical plane into a two-dimensional state matrix. The policy optimization scheduling model is constructed, which includes a feature extraction network and a decision network. The feature extraction network uses a convolutional neural network to extract the spatial features of the two-dimensional state matrix, and the decision network uses a fully connected layer combined with an attention mechanism to process the queue state features. The training process of the policy optimization scheduling model includes a policy network and a value network; The policy network is used to output the probability distribution of each scheduling action based on the input global state features, and to select a scheduling action based on the probability distribution. The value network is used to evaluate the state value function estimate corresponding to the current state and the scheduling action taken; The value network is updated by minimizing the error between the estimated state value function and the actual cumulative return, and the gradient of the value network is used to guide the parameter update of the policy network.

4. The intelligent scheduling method for solid-state drive read / write instructions based on strategy optimization as described in claim 1, characterized in that, The updating of the value network and the parameters guiding the policy network specifically include: Calculate the time difference error, which is the sum of the reward value at the current moment and the estimated state value at the next moment after processing with the discount factor, minus the estimated state value at the current moment; The parameters of the value network are updated by backpropagation using the time difference error. The parameters of the policy network are updated by backpropagation using the product of the log probability of the time difference error and the gradient. It also includes the step of building an experience replay pool: The sample data consisting of the old state, scheduling action, reward value and new state is stored in the experience replay pool; During each training iteration, a preset number of sample data are randomly selected from the experience replay pool as training batches. The parameters of the policy optimization scheduling model are updated using the sample data in the training batch to break the temporal correlation between sample data and improve the stability of training.

5. The intelligent scheduling method for solid-state drive read / write instructions based on strategy optimization as described in claim 1, characterized in that, Randomly extract a preset number of sample data from the experience replay pool, including: Calculate the absolute value of the time difference error for each sample data in the experience replay pool; The sampling priority of each sample data is determined based on the magnitude of the absolute value of the time difference error. The sample data with the larger absolute value of the time difference error has a higher probability of being sampled. The sample data in the experience replay pool is extracted with priority according to the sampling priority, so as to improve the model's utilization efficiency of rare or high-value states. It also includes action exploration and the use of balance strategies: In the early stages of model training, a large exploration rate is used to control the solid-state drive controller to randomly execute scheduling actions in order to traverse as much of the state space as possible. As the number of training steps increases, the exploration rate is gradually reduced according to a preset decay function. In the later stages of model training, the learned optimal strategy is mainly used to select scheduling actions. During the model deployment phase, the parameters of the strategy-optimized scheduling model are frozen, and deterministic scheduling actions are output directly based on the input state.

6. The intelligent scheduling method for solid-state drive read / write instructions based on strategy optimization as described in claim 1, characterized in that, The acquisition of the global state characteristics of the solid-state drive at the current moment also includes: Monitor the waiting time of currently pending read / write commands; If there is an instruction whose waiting time exceeds the preset time threshold, the priority flag of the instruction will be set when generating global state features; When outputting scheduling actions, the strategy optimization scheduling model adjusts the weight of instructions with the priority identifier to increase their probability of being selected for scheduling. The global state features also include the time-series features of the read / write instruction stream sent by the host. Extract the time series features of the read and write instruction stream within a preset time window, and then input the time series features and flash memory internal physical state features into the strategy optimization scheduling model. The strategy optimization scheduling model combines the time series characteristics to predict the load pressure within the future time window, and adjusts the current garbage collection trigger frequency in advance based on the prediction results.

7. The intelligent scheduling method for solid-state drive read / write instructions based on strategy optimization as described in claim 1, characterized in that, The aforementioned adjustment of the current garbage collection trigger frequency based on prediction results includes: When the write load pressure is predicted to be greater than the preset first pressure threshold within a future time window, the scheduling action is controlled to select to perform an active garbage collection operation in order to free up free space. When the read load pressure is predicted to be greater than the preset second pressure threshold within a future time window, the scheduling action is suppressed to select the garbage collection operation to ensure that the flash channel resources are used to respond to read and write requests.

8. The intelligent scheduling method for solid-state drive read / write instructions based on strategy optimization as described in claim 1, characterized in that, During the operation of the solid-state drive, an exploration mechanism is activated within a preset safe time period to collect online sample data under the current actual operating conditions. The deployed strategy optimization scheduling model is updated in small batches using the online sample data. Calculate the loss value of the model on the validation set before and after the update. If the loss value decreases, replace the current model parameters with the updated model parameters; otherwise, keep the current model parameters unchanged. The policy optimization agent model employs a dual-deep Q-network architecture, comprising an evaluation network and a target network. The evaluation network selects actions, and the target network calculates the target Q-value to stabilize the training process. When updating network parameters, the mean squared error loss between the evaluation network's predicted Q-value and the target Q-value is minimized. The convergence process of the loss function follows the following update rules: in, To evaluate the network parameters, For the parameters of the target network, Let D be the discount factor, and D be the batch data set sampled from the playback buffer. These represent the state, action, reward, and next state, respectively.

9. The intelligent scheduling method for solid-state drive read / write instructions based on strategy optimization as described in claim 2, characterized in that, When calculating the reward value, the impact of read interference on the solid-state drive is further considered. The specific reward correction formula is as follows: in, This represents the corrected system reward value. Indicates the base reward value. Indicates the number of physical blocks affected. This represents the raw read error rate of the i-th physical block. This indicates the preset bit error rate safety threshold. This is the penalty coefficient; the reward value is deducted when the read error rate exceeds the threshold.

10. A policy-optimized intelligent scheduling system for solid-state drive read / write instructions, as described in any one of claims 1 to 9, characterized in that, include: The status acquisition module is used to monitor and acquire the global status characteristics of the solid-state drive in real time. The intelligent decision-making module has a pre-trained strategy optimization scheduling model built in, which is used to receive the global state features sent by the state acquisition module and output scheduling actions. An execution control module, connected to the intelligent decision-making module and the solid-state drive controller, is used to convert the scheduling action into control commands and send them to the solid-state drive controller for execution; The feedback update module is used to obtain the performance indicators after the scheduling action is executed, calculate the reward value, and periodically update the model in the intelligent decision-making module using the generated sample data.

Images