Enterprise class hard disk fault tolerant redundant array of independent disks control

Abstract

The application discloses an enterprise-level hard disk fault-tolerant redundant array control device and relates to the technical field of storage control. The device comprises a hard disk state sensing and predicting unit, an active fault-tolerant control logic and a non-volatile cache. The unit outputs a hard disk health index by using a prediction model based on multidimensional data including physical sensing signals. The logic divides the hard disk into multiple grades such as sub-health and critical according to the index. For the sub-health hard disk, selective pre-migration of hot data and dynamic redundancy promotion of cold data are performed. For the critical hard disk, new write requests are intercepted to the non-volatile cache, and full data migration is started. The application can accurately predict faults and respond in stages, realize active defense, reduce the risk of data loss, and has little impact on business.

Description

Technical Field

[0001] This invention relates to the field of storage control technology, and in particular to an enterprise-level hard disk fault-tolerant redundancy array control device. Background Technology

[0002] Enterprise-level storage systems widely employ Redundant Array of Independent Disks (RAID) technology to improve data reliability and read / write performance. Traditional RAID controllers mostly use a passive fault-tolerance mode, meaning that the controller only initiates an emergency response, using parity data to rebuild data on a hot spare or replacement disk, after one of the hard drives in the array reports an uncorrectable read error or fails completely. However, this passive mechanism has inherent drawbacks: First, with the dramatic increase in single-disk capacity, the data reconstruction process takes an extremely long time, sometimes several hours or even days. During this period, the entire array is in a degraded state with reduced reliability. If another hard drive in the array fails at this time, it may lead to the permanent loss of all data in the entire array. This is known as the "reconstruction window risk".

[0003] Secondly, the data reconstruction process generates a large number of background read and write operations, which severely consume I / O bandwidth resources, causing a sharp decline in the access performance of the front-end business system and seriously affecting business continuity.

[0004] To provide early warnings, the industry has introduced Smart Monitoring, Analysis and Reporting Technology (SMART) for hard drives. Some existing technologies attempt to predict failures based on SMART data and proactively trigger data migration when an impending failure is predicted. However, these methods still have shortcomings: firstly, the accuracy of predictions relying solely on SMART attributes is limited, making it difficult to capture sudden failures caused by factors such as physical disturbances; secondly, their fault-tolerant response strategies are relatively simplistic, typically initiating global data migration only after the hard drive is deemed "on the verge of failure," lacking refined, low-cost response strategies for different risk levels, and failing to address the potential impact of the migration process itself on business performance and data consistency issues in extreme situations.

[0005] Therefore, existing technologies urgently need a more accurate, efficient, and business-transparent proactive fault-tolerance mechanism to overcome the risks and performance problems of passive reconstruction and to surpass the limitations of existing proactive early warning solutions. Summary of the Invention

[0006] The technical problem this invention aims to solve is that, in the case of existing hard disk array controllers in passive fault-tolerant mode, there is a high risk of data loss during the reconstruction window and the reconstruction process seriously affects business performance. In addition, existing active fault-tolerant solutions have problems such as inaccurate prediction and single and crude response strategies. This invention aims to provide an enterprise-level hard disk fault-tolerant redundant array control device that can achieve accurate fault prediction, multi-level risk response, and minimize the impact on business.

[0007] To address the aforementioned technical problems, this invention provides an enterprise-level hard disk fault-tolerant redundancy array control device, applicable to hard disk arrays comprising multiple member hard disks. The control device includes: Hard disk status sensing and prediction unit; Active fault-tolerant control logic; and Non-volatile cache; The hard disk status perception and prediction unit is configured to: based on multi-dimensional status data, including physical sensing signals, obtained from the member hard disks, and using a preset fault prediction model, output health status assessment indicators for each member hard disk. The active fault-tolerant control logic is connected to the hard disk status perception and prediction unit and is configured to: classify the health status of the member hard disk into at least a sub-healthy state and a critical state according to the health status assessment index. When any member hard drive is detected to enter the sub-healthy state, selective data pre-migration is performed on the hot data stored on the hard drive that has entered the sub-healthy state, which has an access popularity higher than a preset threshold, and dynamic redundancy enhancement is performed on the cold data stored on the hard drive that has an access popularity lower than the preset threshold. When any member hard drive is detected to enter the critical state, new write requests destined for the critical hard drive are intercepted and stored in the non-volatile cache, and a full data migration of the hard drive is initiated.

[0008] Furthermore, the physical sensing signals include signals from the vibration sensor built into the member hard drive and / or signals from the acoustic sensor built into the member hard drive.

[0009] Furthermore, the multi-dimensional status data also includes the SMART attribute of the member hard drive's self-monitoring, analysis, and reporting technology.

[0010] Furthermore, the fault prediction model is a temporal convolutional network (TCN) model or a long short-term memory (LSTM) model.

[0011] Furthermore, the health status assessment index is determined by at least one of the following methods: based on the Remaining Healthy Life Index (RULI); or based on a preset feature value extracted from the sudden failure probability curve.

[0012] Furthermore, the selective data pre-migration is performed during a service idle window period when the service load of the hard disk array is detected to be below a preset threshold.

[0013] Furthermore, the selective data pre-migration is implemented using striped write-time redirection technology.

[0014] Furthermore, the dynamic redundancy enhancement includes: Upgrade the data stripes that were originally protected by RAID 5 to RAID 6 protection; or The original (k, m) erasure code protected data stripe group is upgraded to (k, m+n) erasure code protected data stripe group, where n is an integer greater than or equal to 1.

[0015] Furthermore, the target location for the selective data pre-migration is a free space distributed across one or more member hard drives that are not in the sub-healthy or endangered state.

[0016] Furthermore, the hard disk status awareness and prediction unit includes a hardware accelerator for performing the fault prediction model inference.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Proactive Defense, Preventing Problems Before They Occur: By integrating physical sensor signals and advanced predictive models, this invention achieves early and accurate prediction of hard drive failures. Combining multi-level state division and proactive intervention, this invention transforms traditional reactive reconstruction into proactive defense, mitigating risks before failures occur and fundamentally avoiding "reconstruction window risks," significantly reducing the probability of permanent data loss due to sudden hard drive failures.

[0018] 2. Minimize Business Impact: The fault-tolerant operation for "sub-healthy" disks is precise and intelligent. By selectively migrating frequently accessed hot data only during business downtime and employing write-time redirection technology to avoid I / O congestion and performance jitter, business continuity is maximized, avoiding the severe impact on business performance caused by traditional rebuild processes. 3. Optimize resource utilization: Through stripe-level dynamic redundancy upgrades and downgrades, protection levels are temporarily increased only for a portion of cold data on "unhealthy" disks, rather than upgrading the entire array to RAID level. This "on-demand protection" strategy avoids the huge storage space waste caused by global redundancy upgrades while ensuring data security, achieving a good balance between reliability and cost efficiency.

[0019] 4. High reliability under extreme conditions: The write interception and caching mechanism for "endangered" disks provides a final layer of data insurance for hard drives that are about to fail completely. Even if the hard drive suddenly goes offline during migration, the latest written data is protected in the non-volatile cache, aiming to ensure that critical business data is not lost in the most critical moments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the system architecture of an enterprise-level hard disk fault-tolerant redundant array control device according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating an enterprise-level hard disk fault-tolerant redundant array control method according to an embodiment of the present invention. Figure 3 This is a signaling interaction timing diagram of the response processing for a hard drive in a sub-healthy state according to an embodiment of the present invention. Figure 4 This is a schematic diagram of strip-level dynamic redundancy upgrade according to an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of the fault prediction model (TCN) network structure according to an embodiment of the present invention.

[0023] Explanation of reference numerals in the attached diagram: 100, Host; 110, Control device; 111, Host interface; 112, Hard disk status awareness and prediction unit; 113, Active fault-tolerant control logic; 114, Non-volatile cache; 115, Redundant computing engine; 116, Hard disk interface; 120, Hard disk array; 121, Member hard disk; 303, Sub-healthy disk; 304, Target healthy disk; 400, RAID5 stripe; 401, Data block D1; 402, Data block D2; 403, Data block D3; 404, P parity block; 405, Q parity block; 410, RAID6 stripe. Detailed Implementation

[0024] 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. The described embodiments should not be considered as limitations on the invention. All other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention. Unless otherwise defined, all technical and scientific terms used in the embodiments of this invention have the same meaning as commonly understood by those skilled in the art. The terminology used in the embodiments of this invention is for the purpose of describing the embodiments of this invention only and is not intended to limit the invention. Before further detailed description of the embodiments of this invention, some nouns and terms involved in the embodiments of this invention are explained, and the nouns and terms involved in the embodiments of this invention are subject to the following interpretations.

[0025] (1) Hard disk status perception and prediction unit: refers to a functional module or component used to evaluate the health status of each member hard disk in real time and in depth. It can collect multi-dimensional status data including physical sensing signals and analyze the data using a preset fault prediction model to output an evaluation index that can quantify the health status of the hard disk.

[0026] (2) Active fault-tolerant control logic: refers to a logic control unit or processing core that executes a preset, hierarchical, and active fault-tolerant operation strategy based on the health status assessment indicators output by the hard disk status perception and prediction unit, so as to avoid risks before the hard disk actually fails.

[0027] (3) Sub-healthy state: This refers to an intermediate risk state in which the hard drive has shown signs of early performance degradation or increased failure risk, but has not yet reached complete failure, as determined by health status assessment indicators (such as Remaining Life Index RULI, Probability of Sudden Failure, etc.). The hard drive in this state can still provide normal service, but its reliability is lower than the normal health level.

[0028] (4) Endangered state: This refers to a high-risk state in which the hard drive is about to or is very likely to fail completely in a short period of time, as determined by health status assessment indicators. This state indicates that hard drive failure is inevitable or highly probable, and the most urgent measures need to be taken immediately to preserve the data.

[0029] (5) Stripe-level Redirect-on-Write: This refers to a data migration or update technique. Its core principle is that when a data stripe needs to be modified or migrated, it is not overwritten at the original physical location. Instead, the new data is written to a new, unused physical location. After the data is completely written to the new location, an atomic address mapping table update redirects the logical address pointing to the data stripe back to the new physical location. This method ensures the interruptibility of the data migration or update process and data consistency.

[0030] Please see Figure 1 This invention aims to provide a proactive fault-tolerant hard disk array control technology solution to address the technical problems of high data loss risk and significant impact on business performance caused by the passive reconstruction method used in existing hard disk arrays when facing member hard disk failures. The core idea of ​​this solution is to accurately and proactively predict the health status of member hard disks and implement multi-level, low-impact proactive fault-tolerant interventions based on the prediction results, thereby preventing failures before they occur and achieving seamless data protection.

[0031] This invention provides an enterprise-level hard disk fault-tolerant redundant array control device 110, which is applied to a hard disk array 120 including multiple member hard disks 121. The core function of the control device 110 is jointly performed by a hard disk status awareness and prediction unit 112 and an active fault-tolerant control logic 113, supplemented by a non-volatile cache 114 to cope with extreme risk scenarios. The hard disk status awareness and prediction unit 112 is responsible for continuously "checking" each member hard disk 121. Specifically, it is configured to calculate a quantitative health status assessment index for each member hard disk 121 based on multi-dimensional status data obtained from each member hard disk 121 using a preset fault prediction model. The characteristic of the "multi-dimensional status data" here is that it breaks through the limitations of traditional reliance on only SMART attributes, and also incorporates physical sensor signals that can reflect subtle physical changes in the hard disks into the scope of consideration, thereby providing a richer data foundation for accurate prediction.

[0032] After obtaining the health status assessment indicators, the proactive fault-tolerant control logic 113 takes over the processing. This logic is designed to transform abstract assessment indicators into concrete, executable fault-tolerant actions. First, based on the health status assessment indicators, it performs a "health classification" of the member hard disk 121, dividing its health status into multiple levels, including at least "sub-healthy state" and "endangered state." This classification mechanism is a prerequisite for achieving refined, low-cost fault tolerance, avoiding a heavyweight "one-size-fits-all" approach to all risks.

[0033] When the proactive fault-tolerant control logic 113 detects that any member hard drive 121 has entered a "sub-healthy state," it initiates a combined response strategy designed to "mitigate risk." On one hand, it performs selective data pre-migration on the hard drive for hot data (i.e., first data) whose access frequency exceeds a preset threshold. This is because hot data is accessed and modified frequently, and its loss or inconsistency has the greatest impact on business operations should the hard drive fail. By migrating it in advance to a member hard drive not in the sub-healthy or critical state, the risk to core data can be effectively reduced. On the other hand, for cold data (i.e., second data) stored on the hard drive whose access frequency does not exceed the preset threshold, dynamic redundancy enhancement is performed. Although cold data is accessed less frequently, it still needs protection. Dynamic redundancy enhancement is a highly cost-effective protection method that temporarily enhances the resilience of this data by adding redundant information (such as additional check blocks) without requiring large-scale data migration, thus saving valuable I / O bandwidth.

[0034] As time progresses, if the health of a member hard drive 121 continues to deteriorate and is determined by the proactive fault-tolerant control logic 113 to be in a "critical state," it indicates that the hard drive may "suddenly fail" at any time, and the most decisive measures must be taken. The proactive fault-tolerant control logic 113 will immediately activate the "emergency plan": First, all new write requests sent to the hard drive in the critical state will be intercepted and transferred to the non-volatile cache 114 in the control device 110. Since the non-volatile cache 114 has the characteristic of not losing data when power is lost, this provides power loss protection for all new write data. At the same time, a full data migration of the hard drive will be immediately initiated, transferring all the data stored on it to member hard drives that are not in the sub-healthy or critical state as soon as possible. Through this mechanism of 'combining write interception and full migration,' new data is guaranteed not to be lost, and a time window is provided for the existing data to complete the migration, thereby maximizing the protection of data integrity and business continuity in extreme cases.

[0035] Furthermore, in a preferred embodiment, to improve the accuracy of fault prediction, the physical sensing signals acquired by the hard disk status sensing and prediction unit 112 may specifically include signals from the vibration sensor built into the member hard disk 121 and / or signals from the acoustic sensor built into the member hard disk 121. The vibration and acoustic characteristics of the hard disk during operation, with their minute abnormal changes, are often early signs of problems with the internal mechanical structure or read / write head assembly. For example, bearing wear may lead to an increase in specific vibration frequency components, and abnormal movement of the read / write head arm may produce unique acoustic patterns. By introducing these physical sensing signals, the present invention can capture more subtle early fault characteristics that traditional SMART systems cannot detect, thereby significantly improving the input information dimension and prediction accuracy of the fault prediction model and solving the problem of inaccurate prediction in the prior art.

[0036] In another preferred embodiment, the multi-dimensional status data may further include the Self-Monitoring, Analysis, and Reporting Technology (SMART) attributes of the member hard drive 121. SMART attributes are a long-established industry standard for hard drive health monitoring, encompassing dozens of parameters such as remapped sector count, seek error rate, and power-on time. While using SMART alone for prediction has limitations, combining it with the aforementioned physical sensing signals can create complementary advantages. Physical sensing signals excel at capturing sudden, physical anomalies, while SMART attributes focus more on recording long-term, cumulative wear and errors. By integrating both as common inputs to the fault prediction model, a more comprehensive and three-dimensional hard drive health profile can be constructed, enabling the prediction results to reflect long-term trends while remaining sensitive to short-term fluctuations, thereby further improving the robustness and reliability of the prediction.

[0037] In one alternative implementation, the fault prediction model can be specifically a Temporal Convolutional Network (TCN) model or a Long Short-Term Memory (LSTM) model. Both are advanced deep learning models, particularly adept at processing time-series data. The vibration, acoustic, and SMART attribute data of a hard drive are essentially time-varying sequential signals. The TCN model, through its causal and dilated convolutional structures, can efficiently capture local patterns and long-term dependencies in time series data, and boasts high computational parallelism, making it suitable for real-time inference on resource-constrained hardware (such as the accelerator within the control unit 110). The LSTM model, through its unique gating unit (input gate, forget gate, output gate) design, effectively solves the gradient vanishing and exploding problems in traditional recurrent neural networks, demonstrating excellent modeling capabilities for ultra-long-term dependencies in time series data. Employing these advanced prediction models, it is possible to automatically learn and extract deep features highly correlated with hard drive faults from complex multidimensional time-series data, achieving higher prediction accuracy than traditional threshold-based or simple statistical methods.

[0038] Furthermore, to provide a clear quantitative basis for the assessment and grading of health status, the health status assessment indicators can be determined in at least one of the following ways: based on the Remaining Useful Life Index (RULI); or based on preset feature values ​​extracted from the burst failure probability curve. RULI (Remaining Useful Life Index) is an intuitive indicator that predicts the remaining effective working time or percentage of the hard drive from the current moment until the expected failure. For example, an RULI of 80% indicates that the hard drive is expected to have 80% of its remaining lifespan remaining. The burst failure probability curve describes the change in the probability of a sudden, catastrophic failure of the hard drive over a future period (e.g., the next 24 hours). Various feature values ​​can be extracted from this curve, such as peak probability, probability growth slope, and duration of high probability. By setting thresholds for these indicators (e.g., RULI below 70% is defined as sub-healthy, below 20% as critical; or a failure probability exceeding 50% in the next 24 hours is defined as critical), the active fault-tolerant control logic 113 can achieve automated and precise grading of the hard drive's health status.

[0039] In a preferred embodiment, the selective data pre-migration is performed during a business idle window period when the service load of the hard disk array 120 is detected to be below a preset threshold. This is to achieve the core objective of minimizing the impact of the present invention on business operations. Data migration, especially cross-disk data copying or reconstruction, inevitably consumes I / O bandwidth and CPU resources. If migration is performed during peak business periods, it will inevitably compete for resources with normal business read and write requests, leading to decreased business performance and increased latency. Therefore, the proactive fault-tolerant control logic 113 can have a built-in load monitoring module to continuously monitor the IOPS, throughput, or CPU utilization of the entire hard disk array 120. When these indicators remain below a preset low-load threshold for a period of time (e.g., IOPS is below 30% of total capacity), the system is considered to have entered a "business idle window period." Initiating data pre-migration at this time can maximize the use of idle system resources, avoid perceptible impact on critical business performance, and achieve harmonious coexistence between fault-tolerant operations and business operations.

[0040] Furthermore, to ensure the security, efficiency, and transparency of the data migration process itself, the selective data pre-migration preferably employs striped write-time redirection technology. For example... Figure 3 As shown, when the proactive fault tolerance control logic 113 decides to migrate a hot data stripe on the sub-healthy disk 303 to the target healthy disk 304, it does not lock the stripe or block access to it. If, during the migration process, the host 100 initiates a write request for the stripe, the proactive fault tolerance control logic 113 will use a write-on-redirect mechanism to directly redirect the new write operation to a new location on the target healthy disk 304 and update a temporary address pointer. The old data on the original sub-healthy disk 303 remains unchanged and continues to serve as the data source for the migration. The entire migration process can be viewed as a series of background operations of "reading old data - writing to the new location". Only after all data blocks have been successfully copied to the target healthy disk 304 will the proactive fault tolerance control logic 113 perform an atomic operation to permanently and indivisibly point the logical address of the data stripe to the new set of physical locations. The advantage of this technology is that the migration process can be safely interrupted and resumed at any time, and it is completely transparent to upper-layer applications. The read and write operations of the host 100 will not be suspended, thus ensuring business continuity.

[0041] In another preferred embodiment, the specific implementation of the dynamic redundancy enhancement may include: upgrading a data stripe originally protected by RAID 5 to RAID 6 protection; or upgrading a data stripe group originally protected by (k, m) erasure coding to (k, m+n) erasure coding protection, where n is an integer greater than or equal to 1. Figure 4As shown, a typical RAID 5 stripe 400 consists of multiple data blocks (such as data blocks D1 401, D2 402, and D3 403) and a P parity block 404, which can only tolerate the failure of a single member disk. When the hard drive carrying a data block in this stripe enters a suboptimal state, the redundancy calculation engine 115 within the active fault tolerance control logic 113 reads all the data blocks of that stripe, calculates and generates an additional Q parity block 405, and writes it to the free space of the array, thus forming an upgraded RAID 6 stripe 410. This new stripe can tolerate the failure of any two member disks, greatly enhancing its reliability. Similarly, for a system using erasure coding, a (k, m) stripe group represents a group of k data blocks and m parity blocks, which can tolerate the failure of m blocks. The dynamic redundancy enhancement operation is to calculate and add n additional parity blocks, making it a (k, m+n) protection level. This "on-demand, temporary, and local" redundancy enhancement method saves storage space significantly compared to globally adjusting the RAID level, achieving a better balance between reliability and cost.

[0042] In one optional implementation, the target location for the selective data pre-migration is a distributed free space located on one or more member hard drives 121 that are not in the sub-healthy or critical state. This distributed migration strategy aims to avoid performance bottlenecks and single points of failure. If all the hot data to be migrated is concentrated on a dedicated hot spare disk, the write bandwidth of that hot spare disk is likely to become a bottleneck, limiting the migration speed; at the same time, this also creates a new single point of failure. By adopting a distributed strategy, the active fault-tolerant control logic 113 will break down the data stripes to be migrated and migrate them to free blocks on multiple member hard drives in the array that are not in the sub-healthy or critical state, according to a load balancing algorithm. For example, stripe A is migrated to hard drive X, and stripe B is migrated to hard drive Y. The advantage of doing this is that the write pressure of the migration is distributed across multiple hard drives, which can be executed in parallel, greatly improving the migration efficiency. At the same time, the migrated data is also stored in a distributed manner, further improving the overall reliability of the data. Even if a target healthy disk subsequently fails, only a small portion of the migrated data will be affected, making recovery faster and easier.

[0043] Please refer to it again. Figure 1 This invention also provides an active fault-tolerant hard disk array control method, which corresponds to the function of the aforementioned control device 110. The method aims to achieve active and intelligent fault-tolerant management of the hard disk array 120 through a series of ordered steps. This method can be executed by the control device 110, or it can run as a pure software control module on a general-purpose server to manage its hard disk resources.

[0044] like Figure 2 As shown, the core process of this method begins with a continuous "acquisition and prediction" step (201). In this step, the control system periodically or in real time acquires multi-dimensional status data from each member hard disk 121 in the hard disk array 120. This data includes, but is not limited to, the aforementioned physical sensing signals (such as vibration and acoustics) and SMART attributes. Subsequently, using a preset fault prediction model (202), the acquired data is analyzed and calculated to output an assessment index for each member hard disk 121 that accurately reflects its current and future health trends.

[0045] After obtaining the health assessment indicators for each hard drive, the method proceeds to the "status grading" step (203). The system compares these indicator values ​​with preset thresholds to classify the health status of each member hard drive 121 into different levels such as "healthy," "sub-healthy," or "endangered." This is a crucial decision-making step, providing a basis for determining the intensity of subsequent intervention measures.

[0046] If a member hard drive 121 is determined to be in a "sub-healthy state," the process will proceed to the "sub-healthy state response" step (204). In this step, the system will execute a combination of strategies: on the one hand, selective data pre-migration will be performed on the hot data with high access frequency on the hard drive, "moving" it to a safer location; on the other hand, dynamic redundancy enhancement will be performed on the remaining cold data with low access frequency on the hard drive, "hardening" its protection. This series of operations aims to mitigate potential, medium-level risks with minimal performance cost.

[0047] If the condition of a member hard drive 121 deteriorates further and is determined to enter a "critical state," the process will proceed to the "critical state response" step (205). This is an emergency procedure where the system immediately intercepts all new write requests destined for the critical hard drive and writes them to a secure non-volatile cache to ensure the safety of new data to the greatest extent possible. Simultaneously, the system will initiate a full data migration of the critical hard drive with the highest priority, quickly "evacuating" all existing data from it to member hard drives that are not in the sub-healthy or critical state. Through this series of emergency measures, even if the hard drive ultimately fails completely, data integrity can be guaranteed to the greatest extent possible.

[0048] In a preferred embodiment, the fault prediction model can also be a Temporal Convolutional Network (TCN) model or a Long Short-Term Memory (LSTM) model. In the "Acquire and Predict" steps (201, 202), the collected multidimensional time-series data is input into the pre-trained TCN or LSTM model for inference computation. The RULI or failure probability curve output by the model serves as the health status assessment metric. This ensures the prediction accuracy and reliability at the method level.

[0049] In another preferred embodiment, the selective data pre-migration step is specifically implemented using stripe write-time redirection technology. When performing the "sub-health state response" step (204), if a data stripe needs to be migrated, the system first allocates space at the target location and then copies the data from the source location to the target location. During this period, any write operations on that stripe will be redirected to the new version at the target location. Once all data copying is complete and verification is successful, the system completes the switch from the logical address to the new physical address through an atomic metadata update. This ensures business continuity and strong data consistency during method execution.

[0050] To illustrate the structure of the invention more specifically, a basic embodiment will be described below. Please refer to [link / reference]. Figure 1 This embodiment provides an enterprise-level hard disk fault-tolerant redundant array control device 110, which exists as an independent hardware entity (such as a PCIe card or a rack-mount controller), communicates with the upper-level host 100 (such as a server) through the host interface 111, and connects to and manages a hard disk array 120 composed of multiple member hard disks 121 through the hard disk interface 116.

[0051] The control device 110 integrates several key units that implement the core functions of this invention. The hard disk status awareness and prediction unit 112 is a highly specialized processing unit that connects to each member hard disk 121 via an out-of-band management channel (such as I2C or SGPIO, not shown in the figure) and can actively acquire various status information of the hard disks. The active fault-tolerant control logic 113 is the core control unit of the entire device, typically implemented by one or more high-performance processors (such as CPU or FPGA), responsible for executing all the decision-making and control processes described herein. The non-volatile cache 114 can be implemented by NVDIMM, MRAM, or DRAM with backup power, providing a reliable write data buffer in critical situations. In addition, the device includes a redundant computing engine 115 for performing RAID checksum or erasure coding calculations, and a host interface 111 and a hard disk interface 116, responsible for data and command interaction with the host 100 and the hard disk array 120, respectively.

[0052] In this basic embodiment, the hard disk status awareness and prediction unit 112 is configured to acquire multi-dimensional status data from the member hard disk 121. This data may include conventional parameters such as the hard disk's internal temperature, read / write error rate, and seek time. The unit uses a preset fault prediction model, such as a statistical model based on multiple linear regression or support vector machines, to comprehensively analyze this data and output a comprehensive health score as a health status assessment indicator.

[0053] Based on this health score, the proactive fault-tolerant control logic 113 classifies the health status of member hard drives 121 into three levels: healthy (score > 80), sub-healthy (score between 50 and 80), and critical (score < 50). These thresholds are exemplary settings and can be adjusted according to the hard drive model and business needs in actual applications.

[0054] When hard drive A is detected to be in a sub-healthy state, the proactive fault tolerance control logic 113 initiates corresponding processing. It first analyzes the access logs of the data blocks stored on hard drive A, identifying the 20% of data with the highest access frequency as hot data (first data). Subsequently, it starts a background task to copy these hot data blocks one by one to a designated hot spare disk in the hard drive array 120 or to the reserved space of other member hard drives not in the sub-healthy or critical state; this is selective data pre-migration. For the remaining 80% of cold data on hard drive A (second data), the proactive fault tolerance control logic 113 calls the redundancy calculation engine 115 to add an extra checksum to the stripe containing this data, thereby temporarily improving its data protection level; this is dynamic redundancy enhancement.

[0055] When hard disk B is detected to be in a critical state, the proactive fault tolerance control logic 113 immediately changes the write path to hard disk B. All new data sent by host 100 that was originally intended to be written to hard disk B is redirected to the non-volatile cache 114, and a write success confirmation is immediately returned to host 100. Simultaneously, the proactive fault tolerance control logic 113 initiates a full disk replication task with the highest priority, migrating all data on hard disk B to member hard disks that are not in the aforementioned sub-healthy or critical state as quickly as possible. After the full data migration is complete, the new data previously cached in the non-volatile cache 114 will also be written to the new location after migration, thus completing the entire fault tolerance process.

[0056] Through the above structure and operating method, this basic embodiment realizes a complete proactive fault-tolerant closed loop of "perception-prediction-tiered response". It can mitigate hard drive failure risks to a certain extent in advance, without relying on specific advanced prediction models or specific sensors, simply through the analysis of routine state data and multi-level response strategies, representing a significant improvement over traditional passive reconstruction mechanisms.

[0057] The specific implementation of the fault prediction model based on multi-sensor fusion temporal convolutional network (TCN) is as follows: To further improve the accuracy and robustness of fault prediction for member hard drive 121, this embodiment provides a specific implementation of the fault prediction model described in the above embodiments. The model is preferably a deep learning model based on a Temporal Convolutional Network (TCN), and its structural design, data processing, training process, and output calculation method are described below.

[0058] I. Model Network Structure See Figure 5 The TCN model used in this embodiment is an end-to-end deep learning network, and its network structure specifically includes: Input Layer: This layer receives preprocessed, multi-dimensional time-series data with a fixed time window size. Let the time window size be... (For example, (representing data collected at the most recent 256 time points), with each time point having a data dimension of [missing information]. ( The value is determined by the following "Input Data Processing", for example... Then the input layer receives a shape of Tensor.

[0059] Causal Dilated Convolutional Layers: This model contains stacked layers. Each residual block has a core causal dilated convolutional layer. Preferably, The structure of each residual block is as follows: First sub-layer: One-dimensional causal dilated convolution, kernel size. =3, Dilation Rate = (in The indexes for the residual blocks start from 0, so the inflation rates for the four blocks are 1, 2, 4, and 8, respectively. This convolutional layer ensures that the output sequence length matches the input sequence length through zero padding.

[0060] The second sub-layer consists of weight normalization and the ReLU activation function.

[0061] The third sub-layer is a regular convolutional layer (1×1 Conv) with a kernel size of 1, used to adjust the number of channels to match the number of input channels of the residual block for residual connections.

[0062] The fourth sub-layer is the Dropout layer, with a dropout rate set to 0.2 to prevent overfitting.

[0063] Skip Connection: The output of the fourth sub-layer is added to the original input to obtain the final output of the residual block.

[0064] Global Average Pooling (GAP): A global average pooling layer is connected after the last residual block. This layer performs pooling along the time dimension, pooling data of length... The time-series features are compressed into a fixed-length feature vector. This operation aims to reduce the number of model parameters and extract global features from the time series.

[0065] Fully Connected Output Layer: This layer takes the feature vectors output from the GAP layer and feeds them into one or more fully connected layers, ultimately mapping them to the output space. Specifically, the output layer of this model contains two parallel output heads: The first output head is a neuron that uses a linear activation function to output a predicted value of the Remaining Healthy Life Index (RULI), which is a continuous value between 0 and 1.

[0066] Second output head: is a header containing... The output layer of 1 neuron (e.g.) The softmax activation function is used to output a discretized probability distribution of the sudden failure probability curve for the next 24 hours. This curve divides the next 24 hours into... The model outputs the probability of a sudden failure occurring within each time interval.

[0067] II. Input Data Processing Flow To ensure the model can extract effective features from the raw physical sensor signals and SMART data, this embodiment employs the following process for preprocessing and feature engineering of the input data, ultimately generating the shape described above. The tensor. This processing flow can be executed by the preprocessing module inside the hard disk state awareness and prediction unit 112.

[0068] Raw data acquisition and alignment: The following raw data from the member hard disk 121 is acquired synchronously with a sampling period of 1 second: Vibration signal: The original acceleration waveforms in the X, Y, and Z directions are read from the built-in triaxial vibration sensor, with a sampling rate of 2560 Hz for each direction.

[0069] Acoustic signal: The raw waveform of sound pressure is read from the built-in acoustic sensor at a sampling rate of 2560 Hz.

[0070] SMART properties: Read a predefined set of key SMART property values, including but not limited to: Raw Read Error Rate, Spin Retry Count, Reallocated Sectors Count, Current Pending Sector Count, Uncorrectable Sector Count, Write Error Rate, etc.

[0071] Feature extraction: For the raw waveform data collected above, extract its time-domain and frequency-domain statistical features within the current 1-second window.

[0072] Temporal characteristics: Calculate the mean, standard deviation, maximum, minimum, kurtosis, and skewness of the vibration and acoustic signals for each axis (X, Y, Z).

[0073] Frequency domain characteristics: Perform a Fast Fourier Transform (FFT) on the vibration and acoustic signals for each axis to calculate its power spectral density (PSD). Extract the following frequency domain characteristics from the PSD: frequency centroid, frequency bandwidth, and energy percentage within a preset frequency band (e.g., 0-100Hz, 100-500Hz, 500-1000Hz, and above 1000Hz).

[0074] SMART features: The collected raw SMART values ​​are normalized and used as part of the input features.

[0075] All the extracted features are concatenated to form a feature with dimension [missing information]. Feature vectors (e.g.) ).

[0076] Time window construction and standardization: The past Each sampling period ( The feature vectors of each sampling point within approximately 4.27 minutes are arranged in chronological order, forming a shape... The original input tensor.

[0077] The tensor is Z-score standardized (i.e., the mean is subtracted and the standard deviation is divided) on each feature dimension to eliminate the influence of different feature units, thus obtaining the standard tensor input to the TCN model. .in .

[0078] III. Model Training Process The TCN model needs to be pre-trained on a large amount of hard drive data before it can be used for actual inference. The training process is as follows.

[0079] Source and annotation of the training dataset: Data Source: Collect complete lifecycle data from hard drives of the same model or series in a real data center environment, from brand new operation to complete failure. The data content must include vibration signals, acoustic signals, and SMART attributes that are completely consistent with the aforementioned "input data processing" process, and the acquisition cycle is also 1 second.

[0080] Data labeling: Generating labels for each time window (i.e., each input sample) in the training dataset. A label consists of two parts: RULI label (Percentage Remaining Life): The actual lifespan of a hard drive is defined as the total runtime from brand new to the point of irreversible failure. For a running The RULI true value of samples collected at specific time points is calculated as follows: .

[0081] Burst Failure Tag (Probability of Failure in the Next 24 Hours): A one-hot encoded vector is generated based on whether the hard drive experienced a burst failure (i.e., a catastrophic failure without prior warning) within 24 hours after the sample time point. If in the future... If the vector fails within the specified time interval, then the first time interval of the vector is... One bit is 1, and the rest are 0.

[0082] Loss Function: Due to the model having two output heads and employing a multi-task learning strategy, the total loss function is... The weighted sum of the two loss functions: For RULI prediction (regression task), the mean absolute error (MAE) loss is used, i.e. .

[0083] For predicting the probability of sudden failures (classification task), the categorical cross-entropy loss is used, i.e. .

[0084] The total loss function is: ,in and For weighting coefficients, preferably, , .

[0085] Optimizer and training parameters: Optimizer: The Adam optimizer (adaptive moment estimator) is used.

[0086] Learning Rate: The initial learning rate is set to... .

[0087] Learning Rate Scheduler: Employs a cosine annealing strategy, gradually decaying the learning rate from its initial value to a predetermined level during training. .

[0088] Batch Size: Set to 256 or 512 depending on the GPU memory size.

[0089] Training epochs: The total number of training epochs is set to 200. Early stopping is also used, which terminates training early if the loss on the validation set does not decrease within 20 consecutive epochs to prevent overfitting.

[0090] Training / Validation Set Splitting: All labeled data are randomly divided into training and validation sets at a ratio of 80% / 20%.

[0091] IV. Calculation of Health Status Assessment Indicators During the inference phase, the TCN model, for the current input time window, simultaneously outputs a RULI prediction value (a scalar value between 0 and 1) and a set of sudden failure probability values ​​for the next 24 hours (a value of length ). (vector of the data). Based on the above original output, the hard disk status awareness and prediction unit 112 further calculates the health status assessment index for use by the active fault-tolerant control logic 113.

[0092] RULI Index: The scalar value directly output by the model's first output header is the Remaining Healthy Lifetime Index (RULI). The closer this value is to 1, the healthier the hard drive; the closer it is to 0, the closer it is to the end of its lifespan. For example, RULI = 0.72 means that the estimated remaining lifespan of the hard drive is 72% of its total lifespan.

[0093] Sudden failure probability characteristic value: The output length of the second output head of the model is The vector, where the first... The element represents the future number... The time period (e.g., the first time period) One and a half hours, if The probability of a sudden failure occurring within a given timeframe. Based on this probability curve, the hard disk status sensing and prediction unit 112 can extract the following preset feature values ​​as the basis for status classification: Peak Probability: The maximum value in the vector represents the maximum risk of sudden failure of the hard drive in the next 24 hours.

[0094] Cumulative Risk: Calculates the average or sum of probabilities over all time periods, representing the overall risk level of sudden failure in the short term.

[0095] Through the above methods, the TCN model can accurately and quantitatively output evaluation indicators reflecting the health status of the hard drive based on multi-source sensor signals, providing a reliable data foundation for subsequent multi-level fault-tolerant responses. It should be noted that the above model structure, parameters, and training methods are illustrative examples. Those skilled in the art, after learning of this embodiment, can make adaptive adjustments and optimizations according to the actual hard drive type, data scale, and computing resources; such adjustments do not depart from the scope of protection of this invention.

[0096] Compared with existing technologies, the fault prediction model implementation scheme proposed in this embodiment has the following significant advantages: 1. High Prediction Accuracy: By fusing multi-physical sensor signals such as vibration and acoustics with SMART data, and leveraging the powerful temporal feature extraction capabilities of the TCN model, this approach can capture subtle signs of early performance degradation and sudden failures in hard drives. Experimental data shows that compared to traditional methods that only use SMART attributes, this solution improves the accuracy of fault prediction by approximately 30% and reduces the false alarm rate by approximately 40%.

[0097] 2. Strong predictive timeliness: The TCN model has the advantage of parallel computing when processing long-term series, with fast inference speed and the ability to achieve real-time prediction at the second level, which provides valuable lead time for subsequent hierarchical fault-tolerant response.

[0098] 3. Good generalization ability: By training on large-scale, multi-model hard disk datasets and introducing regularization techniques such as Dropout, the model can effectively adapt to hard disks under different workloads and environmental conditions, and has good versatility.

[0099] In a preferred embodiment, to further improve the accuracy of prediction and the fineness of response, the present invention introduces more advanced technical means. Specifically, the multi-dimensional state data acquired by the hard disk state perception and prediction unit 112, in addition to conventional parameters, also includes raw signals from the high-sampling-rate triaxial vibration sensor and acoustic sensor built into the member hard disk 121. These physical sensing signals can capture minute dynamic changes in precision mechanical components such as the spindle motor and read / write head arm during hard disk operation at millisecond or even finer time granularity. At the same time, traditional SMART attributes are also collected, forming a data input set containing multiple data types.

[0100] Building upon this foundation, the fault prediction model is no longer a simple statistical model, but has been upgraded to a Temporal Convolutional Network (TCN) model pre-trained on massive amounts of hard drive lifecycle data. The hard drive status awareness and prediction unit 112 integrates a dedicated TCN hardware accelerator, enabling real-time, low-latency inference computation on the collected multi-dimensional time-series data. The TCN model can automatically learn and extract deep temporal features strongly correlated with specific fault modes (such as bearing wear and head instability) from these seemingly noisy signals, thereby outputting a more accurate Remaining Health Life Index (RULI) and a probability curve of sudden failure in the next 24 hours.

[0101] The active fault-tolerant control logic 113 performs more dynamic and reliable state classification based on the RULI and failure probability curves output by the TCN model. For example, when the RULI first falls below 70% or the failure probability curve shows an abnormal rise, the hard drive is marked as "sub-healthy". When the RULI falls below 20% or the peak failure probability exceeds 50%, the hard drive is marked as "critical".

[0102] The response to sub-optimal operating conditions is also more intelligent. The timing of selective data pre-migration is strictly linked to the business load. The proactive fault-tolerant control logic 113 continuously monitors the overall I / O load of the hard disk array 120, and only initiates migration when it detects that the system has entered a business idle window (such as after a nighttime batch processing task has ended), thereby ensuring that the performance impact on daytime online services is close to zero.

[0103] More importantly, the data migration process employed striped write-time redirection technology. For example... Figure 3As shown, when migrating hot data on the sub-healthy disk 303, the proactive fault-tolerance control logic 113 does not lock the data. When the host 100 writes to this data during this period, the write operation is seamlessly redirected to a new location on the target healthy disk 304. After the migration is complete, the logical address is updated through an atomic address mapping table switch. This ensures that even if the migration process is time-consuming, it will not interrupt business operations, and the data will always maintain strong consistency.

[0104] For the cold data on the sub-health panel, the process of dynamic redundancy improvement is also more specific. For example... Figure 4 As shown, if a cold data stripe originally belonged to a RAID 5 stripe 400, consisting of data blocks D1 401, D2 402, D3 403 and a P parity block 404, the redundancy calculation engine 115 will dynamically calculate a new Q parity block 405 for it and logically combine it with the original stripe data to form a new, temporary RAID 6 stripe 410. In this way, the stripe's fault tolerance is improved from resisting single-disk failure to resisting dual-disk failure, and this only consumes the extra space of one parity block.

[0105] This invention can be specifically implemented as a hardware expansion card. For example, the expansion card can be designed to connect to the motherboard of a server or storage controller via a PCIe interface. The card can integrate one or more high-performance FPGA or ASIC chips to embed the active fault-tolerant control logic 113, redundant computing engine 115, and hardware accelerators for performing TCN model inference as described herein. The card can also have onboard non-volatile memory with backup power protection as a high-performance, highly reliable non-volatile cache 114. The card provides multiple hard drive interfaces 116, which can directly connect to multiple SAS or SATA hard drives to form a hard drive array 120. In this way, the active fault-tolerant capabilities described herein can be provided to the server or storage system.

[0106] Now, we describe a minimum protection scope embodiment that integrates several preferred features of the present invention and provides a high level of protection. The control device 110 of this embodiment is, in hardware, the aforementioned PCIe card. The hard disk array 120 it manages consists of 24 18TB SAS hard disks configured as a large storage pool.

[0107] The complete workflow of this embodiment is as follows: 1. Continuous Sensing and Prediction: The hard disk status sensing and prediction unit 112 collects raw waveform data from triaxial vibration and acoustic sensors from all 24 member hard disks 121 at a frequency of 1000 times per second, and collects all SMART attributes once per minute. These data streams are fed into the TCN hardware accelerator in real time. The TCN model continuously outputs the RULI value of each hard disk and the probability curve of sudden failure for the next 24 hours. 2. Sub-health event trigger: After running for a period of time, the RULI value of hard drive #5 slowly decreased from 95% to 68%, falling below the sub-health threshold of 70% for the first time. The active fault tolerance control logic 113 immediately marked hard drive #5 as "sub-healthy" and generated an internal alarm. 3. Sub-health Response Initiation: The proactive fault-tolerance control logic 113 analyzes the IO access logs and finds that the database index files stored on hard drive #5 and the root partition files of the virtual machine operating system are accessed most frequently. It marks the approximately 200GB data stripes corresponding to these files as hot data. Simultaneously, it begins monitoring the overall IOPS of the array. At 2:00 AM, the system IOPS drops below 5000, below the 30% load threshold, and the proactive fault-tolerance control logic 113 determines that the business idle window is open.

[0108] 4. Combined Fault-Tolerant Execution: The proactive fault-tolerance control logic 113 initiates a sub-health response for hard disk #5. On one hand, it performs selective data pre-migration: the 200GB hot data stripe is migrated in 64MB blocks using a distributed strategy to the free space of hard disks #10, #11, #12, and #13. The migration process uses stripe write-time redirection technology, which is transparent to upper-layer services. On the other hand, for the remaining approximately 15TB of cold data on hard disk #5 (mainly archived logs and backup files), which originally belonged to the (16,2) erasure coding stripe group, the proactive fault-tolerance control logic 113 dynamically calculates and adds a check block for each stripe group involved, temporarily upgrading it to the (16,3) protection level.

[0109] 5. Endangered Event: During the aforementioned migration and redundancy enhancement process, the TCN model output of hard disk #21 suddenly changes drastically, with its sudden failure probability soaring from 1% to 60% within minutes. The active fault-tolerant control logic 113 immediately marks hard disk #21 as "endangered". 6. Endangered Response Takeover: Proactive fault-tolerant control logic 113 immediately executes the highest priority endangered response. All new write requests destined for hard disk #21 are intercepted and written to the non-volatile cache 114 on the PCIe card. Simultaneously, the system initiates a full data migration of hard disk #21 using all available background bandwidth, targeting a global hot spare disk #24 in the array.

[0110] 7. Results: Several hours later, the hot data migration and cold data redundancy enhancement of hard drive #5 were completed, and the risk was successfully mitigated. Several hours later, hard drive #21 completely stopped responding when the full migration was 80% complete. However, since the newly written data was already in the cache and most of the existing data had been migrated, the system only needed to use the verification information to reconstruct the remaining 20% ​​of the data to complete the recovery. Throughout the process, the risk of business interruption was significantly reduced and data loss was avoided.

[0111] This embodiment achieves the following comprehensive technical effect by combining all the above-mentioned technical features: 1. High prediction accuracy: By integrating physical sensor signals and the TCN model, the ability to predict slow degradation and sudden risks of hard drives is improved. 2. Low business impact: Through the combination of business idle window detection, selective migration and write-time redirection technologies, the impact of preventative fault-tolerant operations on business performance is almost negligible.

[0112] 3. Optimized resource efficiency: Distributed migration avoids hot spare disk bottlenecks, while dynamic redundancy enhancement achieves higher reliability at a lower space cost, thus optimizing resource utilization.

[0113] 4. Strong reliability guarantee: The write interception and full migration mechanism for critical states provide the ultimate guarantee for data security, ensuring data integrity even in extreme cases where the hard drive is about to fail completely.

[0114] The technical solution provided by this invention has a wide range of applications. In large-scale cloud computing data centers, the procurement volume of servers and hard drives is enormous. Even a failure rate of a few per thousand per year means that a large number of hard drives need to be replaced every day. By using this invention, proactive warnings can be issued and data migration can be completed before hard drive failures occur, transforming unplanned downtime maintenance into planned hard drive replacements, significantly improving operational efficiency and reducing the risk of business interruption due to hard drive failures.

[0115] In industries such as finance and securities, where data consistency and business continuity are extremely critical, any data loss or transaction interruption due to hardware failure can result in significant economic losses. This invention, through its accurate prediction and seamless fault-tolerance mechanism, can maximize the stable operation of the storage system, providing reliable data storage support for core transaction systems.

[0116] In the field of high-definition video surveillance, storage systems need to handle a large number of write requests around the clock, resulting in hard drives operating under high load for extended periods and a relatively high failure rate. Traditional RAID reconstruction processes are not only lengthy but also severely impact video playback performance. This invention can automatically handle hard drives that are about to fail without affecting real-time video recording and playback, ensuring the integrity and availability of surveillance footage.

[0117] Furthermore, in any field that relies on large-scale hard disk arrays for data storage and processing, such as scientific computing, media asset creation, and big data analysis, this invention can significantly improve the reliability, availability, and maintainability of data storage through its proactive, intelligent, and efficient fault-tolerance mechanism, thereby reducing system maintenance costs.

Claims

1. An enterprise-level hard disk fault-tolerant redundancy array control device, applied to a hard disk array comprising multiple member hard disks, characterized in that, The control device includes: Hard disk status sensing and prediction unit; Active fault-tolerant control logic; and Non-volatile cache; The hard disk status perception and prediction unit is configured to: based on multi-dimensional status data, including physical sensing signals, obtained from the member hard disks, and using a preset fault prediction model, output health status assessment indicators for each member hard disk. The active fault-tolerant control logic is connected to the hard disk status perception and prediction unit and is configured to: classify the health status of the member hard disk into at least a sub-healthy state and a critical state according to the health status assessment index. When any member hard drive is detected to enter the sub-healthy state, selective data pre-migration is performed on the hot data stored on the hard drive that has entered the sub-healthy state, which has an access popularity higher than a preset threshold, and dynamic redundancy enhancement is performed on the cold data stored on the hard drive that has an access popularity lower than the preset threshold. When any member hard drive is detected to enter the critical state, new write requests destined for the critical hard drive are intercepted and stored in the non-volatile cache, and a full data migration of the hard drive is initiated.

2. The apparatus according to claim 1, characterized in that, The physical sensing signals include signals from the vibration sensor built into the member hard drive and / or signals from the acoustic sensor built into the member hard drive.

3. The apparatus according to claim 1 or 2, characterized in that, The multi-dimensional status data also includes the SMART attribute, a self-monitoring, analysis, and reporting technology for the member hard drives.

4. The apparatus according to claim 1, characterized in that, The fault prediction model is either a temporal convolutional network (TCN) model or a long short-term memory (LSTM) model.

5. The apparatus according to claim 1, characterized in that, The health status assessment index is determined by at least one of the following methods: based on the Remaining Healthy Life Index (RULI); or based on a preset feature value extracted from the sudden failure probability curve.

6. The apparatus according to claim 1, characterized in that, The selective data pre-migration is performed during a service idle window period when the service load of the hard disk array is detected to be lower than a preset threshold.

7. The apparatus according to claim 1, characterized in that, The selective data pre-migration is implemented using striped write-time redirection technology.

8. The apparatus according to claim 1, characterized in that, The dynamic redundancy enhancement includes: Upgrade the data stripes that were originally protected by RAID 5 to RAID 6 protection; or The original (k, m) erasure code protected data stripe group is upgraded to (k, m+n) erasure code protected data stripe group, where n is an integer greater than or equal to 1.

9. The apparatus according to claim 1, characterized in that, The target location for the selective data pre-migration is the free space distributed across one or more member hard drives that are not in the sub-healthy or endangered state.

10. The apparatus according to claim 1, characterized in that, The hard disk status awareness and prediction unit includes a hardware accelerator for performing the inference of the fault prediction model.

Images