A method for monitoring the health of a server storage system

By constructing load resistance characteristics and temperature fluctuation characteristics to calculate the hysteresis decay index, the problem of misjudgment of disk aging in existing technologies is solved, and high reliability and automated operation and maintenance of storage systems are achieved.

CN122086736AActive Publication Date: 2026-05-26ZHONGNAN INFORMATION TECH (SHENZHEN) CO LTD +1
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202610560231.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-27
Publication Date
2026-05-26
Estimated Expiration
2046-04-27

AI Technical Summary

Technical Problem

Existing technologies cannot accurately separate sudden business operations from disk aging caused by temperature fluctuations, resulting in storage systems being unable to provide accurate early warnings before hardware failure, thus affecting the operational reliability of large-scale storage clusters.

Method used

By collecting disk request stream data, load resistance characteristics and temperature fluctuation characteristics are constructed, a hysteresis decay index is calculated, and a health score is obtained by combining the benchmark decay index, enabling in-depth logical deduction and early warning of disk health status.

Benefits of technology

It accurately captures the underlying physical response capability and media fatigue of the disk, and eliminates interference from environmental thermodynamic fluctuations, thereby improving the operational reliability and automated operation and maintenance efficiency of the storage system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122086736A_ABST
    Figure CN122086736A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of storage system operation and maintenance technology, specifically relating to a health monitoring method for server storage systems. The method includes: collecting request stream data from a target disk to obtain input / output request rate sequences and input / output latency sequences; dividing time windows and obtaining load resilience characteristics based on the transient increase in input / output latency and the transient decrease in input / output request rate; synchronously collecting the surface temperature of the physical media to obtain temperature fluctuation characteristics, thereby determining the hysteresis degradation index of the target disk; calculating a health score based on the hysteresis degradation index and a benchmark degradation index, and performing early warning and storage scheduling operations. This invention can accurately capture underlying hidden performance degradation, effectively eliminate short-term performance jitter interference, and improve the operational reliability and data security of the storage architecture.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of storage system operation and maintenance technology. More specifically, this invention relates to a method for monitoring the health of a server storage system. Background Technology

[0002] With the expansion of cloud computing and big data businesses, the stability of server storage systems directly determines the continuity of upper-layer businesses. Currently, in the field of server storage system health monitoring, the main approach is to directly read the internal status attributes of storage devices or monitor static thresholds for average input / output latency. Alarms are triggered when relevant values ​​exceed the baseline.

[0003] However, in the actual operating environment of modern data centers, server load exhibits extremely strong transient bursts. When the physical media inside the storage device is performing low-level physical operations such as data erasure and write redirection, if it is intertwined with high-concurrency business, it is very easy to cause severe fluctuations in tail latency. Traditional static threshold and average statistical methods cannot accurately separate the reasonable latency surges caused by high-concurrency business flows from the endogenous processing lags caused by actual aging. When the business flood recedes, the storage disks with declining health will expose the problem of low resource scheduling efficiency, resulting in a significant physical lag phenomenon as their response time falls back.

[0004] Fixed threshold strategies often homogenize asynchronous fluctuations that include physical degradation information. In addition, the conductivity of the internal media of the disk and the thermodynamic throttling mechanism of the control chip are significantly affected by the ambient temperature. Abnormally high temperatures can easily trigger delays caused by frequency reduction protection logic. Existing technologies cannot isolate short-term performance jitter interference caused by environmental thermodynamic fluctuations. Relying solely on static indicators can easily lead to misjudgments of physical degradation, making it difficult for the system to provide accurate early warnings before hardware failure, which seriously restricts the operational reliability of large-scale storage clusters. Summary of the Invention

[0005] To address the aforementioned technical problem of misjudging disk aging due to sudden business interruptions and temperature fluctuations, this invention provides a method for monitoring the health of a server storage system, comprising: The system collects request stream data from the target disk and performs statistical analysis according to a set sampling period. It obtains the input / output request rate and average completion time of input / output requests for each sampling period, forming an input / output request rate sequence and an input / output latency sequence, respectively. The input / output request rate sequence and the input / output latency sequence are divided into several time windows of equal length. Based on the transient increase in input / output latency and the transient decrease in input / output request rate at each sampling point within each time window, the system obtains the load resistance characteristics of the target disk within each time window. Simultaneously, the system collects the physical media surface temperature of the target disk for each sampling period to obtain the temperature fluctuation characteristics within each time window. Based on the load resistance characteristics and the temperature fluctuation characteristics, it determines the hysteresis degradation index of the target disk. Based on the hysteresis degradation index of the target disk and a preset baseline degradation index, it calculates the health score of the target disk and executes corresponding early warning and storage scheduling operations based on the health score.

[0006] This invention obtains load resilience characteristics by comprehensively considering the transient increase in input / output latency and the transient drop in input / output request rate at sampling points within a time window. It can accurately capture asynchronous physical hysteresis phenomena accompanied by high response latency during periods of low business load, truly reflecting the physical response capability and media fatigue of the disk's underlying layer. This effectively overcomes the technical limitations of traditional static fixed thresholds, which are prone to misjudging normal business congestion and real physical hysteresis when facing complex burst business flows. This invention combines load resilience characteristics and temperature fluctuation characteristics to jointly determine the hysteresis degradation index, fully considering the latency caused by the overheating and frequency reduction protection logic of the control chip under abnormal high temperature fluctuations. It successfully isolates the short-term performance jitter interference caused by environmental thermodynamic fluctuations, thereby extracting the real degradation characteristics caused purely by the irreversible aging of the underlying hardware media. Based on the hysteresis degradation index of the target disk and the preset benchmark degradation index, this invention calculates a continuously converging health score, realizing a deep logical deduction from the surface request flow phenomenon to the health evolution of the underlying physical devices of the server storage. It proactively isolates the risk of data collapse caused by hardware degradation and improves the operational reliability and automated operation and maintenance efficiency of the underlying storage architecture of large-scale data centers.

[0007] Preferably, obtaining the input / output request rate corresponding to each sampling period includes: statistically analyzing the request stream data to obtain the number of input / output requests in each sampling period, dividing the number of input / output requests in each sampling period by the sampling period duration to obtain the input / output request rate corresponding to each sampling period.

[0008] Preferably, the load resistance characteristic satisfies the expression: In the formula, Indicates the first Load resistance characteristics within a time window; Indicates the number of sampling points included within the time window; Indicates the first Within the first time window Input / output delay at each sampling point; Indicates the first The mean of the input and output delays of all sampling points within a time window; Indicates the first Within the first time window Input / output request rate per sampling point; Indicates the first The average input / output request rate of all sampling points within a time window; This indicates the operation of retrieving the maximum value; Represents an exponential function with the natural constant as its base; Indicates the sequence number of the sampling point within the time window; The smoothing penalty constant represents the input / output delay; This represents the smoothing penalty constant for the input / output request rate.

[0009] This invention amplifies the nonlinear difference between input and output delays exceeding the mean by using an exponential function, and extracts the positive proportion of the drop in input and output request rates by taking the maximum value operation. Multiplying and accumulating the two can capture abnormal physical conditions of high latency accompanied by low business load periods, effectively filtering out the interference of normal response fluctuations. This allows the calculated load resistance characteristics to intuitively and sensitively map the physical media wear and background management pressure of the target disk.

[0010] Preferably, obtaining the temperature fluctuation characteristics within each time window includes: calculating the standard deviation of the physical medium surface temperature of the target disk within the time window, as the temperature fluctuation characteristics within that time window.

[0011] Preferably, the hysteresis recession index satisfies the following expression: In the formula, Indicates the first The lag recession index corresponding to each time window; This indicates the number of historical time windows being traced. Indicates the tracing sequence number of the historical time window; Indicates the first Load resistance characteristics within a time window; Indicates the first Temperature fluctuation characteristics within a time window; Represents an exponential function with the natural constant as its base; The smoothing penalty constant represents the temperature fluctuation characteristics.

[0012] This invention utilizes temperature fluctuation characteristics to adaptively reduce the weight of load resistance features in historical time windows, and combines a time decay mechanism to give higher evaluation weight to recent data. When the target disk experiences severe thermodynamic state oscillations, it can dynamically reduce the feature contribution ratio of abnormal periods, weaken the weight of system performance jitter caused by temperature fluctuations, and eliminate the interference caused by short-term environmental thermodynamic fluctuations. This allows the final determined hysteresis decay index to reflect the real response hysteresis caused by the irreversible aging of the underlying hardware media.

[0013] Preferably, the hysteresis recession index satisfies the following expression: In the formula, This indicates the lag recession index for the current time window; This represents the lagging recession index for the previous time window. Indicates the recursive smoothing coefficient; Indicates the first Load resistance characteristics within a time window; Indicates the first Temperature fluctuation characteristics within a time window; The smoothing penalty constant represents the temperature fluctuation characteristics.

[0014] This invention utilizes a recursive smoothing coefficient to perform a first-order recursive filtering update on the observations within the current time window and the hysteresis decay index of the previous time window. By using the recursive smoothing coefficient to approximate the time decay factor, the weighted accumulation process of long-period historical time windows is transformed into a real-time update logic. This not only preserves the weighted memory of historical hysteresis characteristics and the temperature penalty mechanism, but also avoids the repeated traversal calculation of a large number of historical sequences. Thus, while ensuring the accuracy of the evaluation, it significantly reduces the system's computational resource consumption and improves the real-time processing efficiency under high-frequency sampling.

[0015] Preferably, the method for obtaining the benchmark degradation index is as follows: in response to the target disk being a brand new disk before its first online operation, a full-load random read / write pressure is applied to the target disk; under the full-load pressure condition, the input / output request rate sequence, input / output delay sequence, and continuous temperature observation sequence corresponding to the target disk are collected synchronously, the hysteresis degradation index of the target disk under the full-load pressure condition is calculated, and the hysteresis degradation index is used as the benchmark degradation index.

[0016] Preferably, the method for obtaining the benchmark decay index is as follows: in response to the target disk being an existing disk already running in the data center, a healthy reference disk cluster of the same batch and model as the target disk is retrieved, and the average hysteresis decay index of the healthy reference disk cluster during peak business periods is used as the benchmark decay index.

[0017] Preferably, the health score satisfies the expression: In the formula, Indicates the target disk is in the Health score corresponding to each time window; Indicates the first The lag recession index corresponding to each time window; Represents the hyperbolic tangent function; This represents the benchmark recession index.

[0018] This invention uses a hyperbolic tangent function to smoothly map the ratio of the hysteresis decay index to the benchmark decay index within a continuously converging interval, ensuring the stability of the evaluation results. When the current hysteresis deviates significantly from the benchmark characteristics, it can rapidly amplify the severity of physical decay, causing the health score to drop more rapidly, thus providing a reliable quantitative benchmark for assessing the risk of large-scale failure of the internal media of the target disk.

[0019] Preferably, the step of performing corresponding early warning and storage scheduling operations based on the health score includes: in response to the target disk's health score being lower than the set health alarm baseline for a consecutive preset number of time windows, triggering a hardware replacement work order generation operation, and initiating a distributed replica balance migration task for the data within the target disk.

[0020] The beneficial effects of this invention are as follows: This invention constructs load resilience features by extracting the transient increase in input / output latency and the transient drop in input / output request rate. This accurately captures the underlying implicit performance degradation of the storage medium when the background physical mechanism is obstructed, effectively overcoming the technical limitations of traditional static fixed thresholds, which are prone to misjudgment and omission when facing complex and sudden business flows. This invention introduces temperature fluctuation characteristics to adaptively reduce the weight of load resilience features from historical periods and combines a time decay mechanism to determine the hysteresis degradation index, successfully eliminating short-term performance jitter interference caused by the thermodynamic throttling effect of the controller, ensuring the authenticity and stability of long-term physical degradation feature extraction. This invention constructs a continuously converging health score system through benchmark degradation index calibration and nonlinear mapping, realizing a deep logical deduction from surface request flow phenomena to the health evolution of the underlying physical devices of server storage. This proactively isolates the risk of data collapse caused by hardware degradation, improving the operational reliability and automated operation and maintenance efficiency of the underlying storage architecture of large-scale data centers. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a health monitoring method for a server storage system according to the present invention; Figure 2 A schematic diagram illustrating the changes in target disk I / O request rate and I / O latency; Figure 3 This is a schematic diagram illustrating the changes in load resistance characteristics; Figure 4 This is a schematic diagram illustrating the changes in temperature fluctuation characteristics. Figure 5 A schematic diagram illustrating the changes in the target disk lag and degradation index; Figure 6 A diagram illustrating the dynamic health score changes of the target disk. Figure 7 This diagram illustrates a comparison of the monitoring effects of existing technologies and the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] This invention discloses a method for monitoring the health of a server storage system, referring to... Figure 1 This includes steps S1 to S4: S1. Collect the request stream data of the target disk, perform statistics according to the set sampling period, and obtain the input / output request rate and the average completion time of the input / output request corresponding to each sampling period, which constitute the input / output request rate sequence and the input / output delay sequence, respectively.

[0025] Specifically, the system monitors and collects request stream data from the target disk in real time. It then statistically analyzes the request stream data according to a set sampling period to obtain the number of input / output requests and the average completion time of each input / output request within each sampling period. The number of input / output requests in each consecutive sampling period is divided by the sampling period duration to obtain the input / output request rate. The input / output request rates corresponding to each sampling period are then arranged sequentially according to the sampling period to form an input / output request rate sequence. Similarly, the average completion time of input / output requests in each consecutive sampling period is arranged sequentially according to the sampling period to form an input / output delay sequence.

[0026] In this embodiment, the sampling period is set to 10 milliseconds. Since the typical read / write completion time of a physical flash memory page is usually between a few milliseconds and tens of milliseconds, a sampling period of 10 milliseconds can accurately capture transient physical congestion. In other embodiments, implementers can set the sampling period according to the actual read / write characteristics of the physical flash memory page, such as 5 milliseconds or 20 milliseconds.

[0027] S2. Divide the input / output request rate sequence and the input / output delay sequence into several time windows of equal length. Based on the transient increase in input / output delay and the transient drop in input / output request rate at the sampling points within each time window, obtain the load resistance characteristics of the target disk within each time window.

[0028] It should be noted that in a healthy disk, input / output latency is usually positively correlated with input / output request rate. As the physical media of the disk ages, background tasks consume a large amount of main control chip resources, which in turn causes the actual response latency to fail to fall back in time with the decrease in external load, resulting in distorted performance evaluation. Therefore, in order to capture the physical lag phenomenon of high latency accompanied by low business load and to truly reflect the physical response capability and media fatigue of the disk, this invention obtains the load resistance characteristics of the target disk in each time window based on the input / output request rate sequence and the input / output latency sequence.

[0029] Specifically, the input / output request rate sequence and the input / output delay sequence are divided into several time windows of equal length. In this embodiment, the number of sampling points included in each time window is... The value is set to 50. Since the typical execution cycle of background maintenance of the physical media inside the target disk is usually between 100 milliseconds and 1000 milliseconds, setting 50 sampling points can correspond to a time span of 500 milliseconds, thereby fully covering at least one background physical operation cycle and filtering out extremely short-term transient data spikes. In other embodiments, the implementer can set the number of sampling points included in the time window according to the actual duration of the background maintenance cycle, such as 20 or 100.

[0030] For any given time window, calculate the load resistance characteristics within that time window: ; In the formula, Indicates the first Load resistance characteristics within a time window; Indicates the number of sampling points included within the time window; Indicates the first Within the first time window Input / output delay at each sampling point; Indicates the first The mean of the input and output delays of all sampling points within a time window; Indicates the first Within the first time window Input / output request rate per sampling point; Indicates the first The average input / output request rate of all sampling points within a time window; This indicates the operation of retrieving the maximum value; Represents an exponential function with the natural constant as its base; Indicates the sequence number of the sampling point within the time window; The smoothing penalty constant representing the input-output delay is used to prevent the denominator from being 0. In this embodiment, the limiting response time of the underlying physical medium based on the input-output delay is set to 0.1 milliseconds. The smoothing penalty constant representing the input / output request rate is used to prevent the denominator from being zero. In this embodiment, the minimum statistical cardinality for a single discrete request based on the input / output request rate is set to 0.1 times / millisecond. In other embodiments, implementers can adjust this according to the actual hardware performance indicators of the storage device and the system sampling accuracy. and .

[0031] Among them, the mean of input and output delay and the average input / output request rate This represents the baseline state of the target disk's macroscopic throughput and response within the current time window. This reflects the degree of sudden increase in delay, when the first Input / output delay at each sampling point Less than or equal to the mean of input and output delays hour, The output range is the minimum value between [0,1], which exponentially suppresses the normal response state; when the... Input / output delay at each sampling point Greater than the average input / output delay When, the ratio It is a positive value. The nonlinear differences are rapidly amplified by using the base of the natural constant, mapping the abnormal congestion level of the medium response, and the dimensionless nature of the variable is achieved by dividing by the mean. When the... Input / output request rate per sampling point Less than the average input / output request rate hour, Output a non-zero positive scaling value to activate feature capture, reflecting the degree of transient drop in the request rate. This is achieved by multiplying the degree of latency transient spike by the degree of request rate transient drop and iterating through the number of sampling points contained within the time window. Accumulation can amplify abnormal physical conditions such as high latency during periods of low business load. When the first... When the latency of each sampling point within a time window is greater and the request rate is lower, the load resistance characteristic obtained by multiplying and summing these characteristics will be more pronounced. The larger the value, the more severe the internal response latency of the target disk when the external request rate decreases, and the greater the wear and tear of the underlying physical media or the greater the background management pressure on the target disk.

[0032] For example, Figure 2 This is a diagram illustrating the changes in target disk I / O request rate and I / O latency. Figure 2 It can be seen that in the initial health phase of monitoring, the input / output delay and the input / output request rate exhibit a highly consistent positive correlation. In the thermodynamic interference phase of the middle phase of monitoring, the abnormal ambient temperature triggers the equipment's self-protection frequency reduction, resulting in an irrational and drastic increase in the input / output delay. In the physical media degradation phase of the later phase of monitoring, when the input / output request rate drops sharply, the input / output delay no longer falls back synchronously, but instead exhibits significant abnormal hovering and hysteresis.

[0033] Figure 3 This is a schematic diagram illustrating the changes in load resistance characteristics, from Figure 3 It can be seen that under normal healthy conditions, the load resistance characteristic is suppressed and remains stable in an extremely low range. When the system faces abnormal high temperature fluctuations or the underlying physical medium undergoes real aging, the load resistance characteristic increases in the corresponding range, proving that the load resistance characteristic constructed in this invention can amplify and capture the latent abnormal operating conditions of low load accompanied by high latency with extremely high sensitivity.

[0034] S3. Synchronously collect the physical media surface temperature of the target disk corresponding to each sampling period to obtain the temperature fluctuation characteristics within each time window, and determine the hysteresis degradation index of the target disk based on the load resistance characteristics and the temperature fluctuation characteristics.

[0035] It should be noted that the conductivity of the internal storage medium of the disk and the thermodynamic throttling mechanism of the control chip are significantly affected by the ambient temperature. Under abnormal high temperature fluctuations, delays caused by the overheating and frequency reduction protection logic of the control chip are prone to occur, which may lead to misjudgments of physical degradation based solely on the delay index. Therefore, in order to isolate the short-term performance jitter interference caused by environmental thermodynamic fluctuations and extract the true degradation characteristics caused purely by the irreversible aging of the underlying hardware medium, this invention determines the hysteresis degradation index of the target disk based on the load resistance characteristics and temperature fluctuation characteristics within each time window.

[0036] Specifically, the surface temperature of the physical medium corresponding to each sampling period is synchronously collected through the intelligent self-monitoring analysis and reporting technology interface built into the target disk, forming a continuous temperature observation sequence. The temperature fluctuation characteristics within each time window are calculated. ; In the formula, Indicates the first Temperature fluctuation characteristics within a time window; Indicates the number of sampling points included within the time window; Indicates the first Within the first time window Temperature at each sampling point; Indicates the first The average temperature of all sampling points within a time window; This indicates the sequence number of the sampling point within the time window. This invention utilizes the standard deviation of temperature within the current time window to reflect the instability of the medium's thermodynamic state within that window. The greater the temperature deviation of each sampling point from the mean, the stronger the temperature fluctuation characteristics. The larger.

[0037] Furthermore, the hysteresis index corresponding to each time window is calculated: ; In the formula, Indicates the first The hysteresis decay index corresponding to each time window reflects the disk response hysteresis rate under the influence of unit temperature fluctuation after time-weighted smoothing. This indicates the number of historical time windows being traced. Indicates the tracing sequence number of the historical time window; Indicates the first Load resistance characteristics within a time window; Indicates the first Temperature fluctuation characteristics within a time window; Represents an exponential function with the natural constant as its base; The smoothing penalty constant represents the temperature fluctuation characteristics.

[0038] When the Temperature fluctuation characteristics within a time window The larger the value, the more intense the thermodynamic fluctuations of the target disk during that period. In this case, the larger the denominator, the smaller the contribution ratio of that period to the calculation result, which in effect weakens the weight of system performance jitter caused by intense temperature fluctuations in the final health assessment. As a time decay factor, when tracing the sequence number The larger the value, i.e., the longer the time period, the smaller the decay weight, thus assigning a higher weight to time windows closer to the present moment; the hysteresis decay index is obtained by weighting and accumulating historical sequences. The larger the value, the more likely the target disk is to exhibit continuous and stable abnormal response lag after excluding short-term thermodynamic fluctuations, and the higher the probability of actual damage and degradation of its internal storage block media.

[0039] In this embodiment, the smoothing penalty constant of the temperature fluctuation characteristics is... Set to 0.1 degrees Celsius, smoothing penalty constant. It possesses physical dimensions consistent with temperature fluctuation characteristics to prevent the denominator from becoming zero and to suppress numerical explosion during extremely low temperature fluctuations. Since the temperature control accuracy of the constant temperature air conditioning system in data center server rooms is typically around 0.1 degrees Celsius, setting it to 0.1 degrees Celsius accurately reflects the temperature floor noise under constant temperature conditions. In other embodiments, implementers can adjust the settings according to the actual temperature control accuracy of the server room. The setting, for example, 0.05 degrees Celsius or 0.2 degrees Celsius.

[0040] In this embodiment, the number of historical time windows to be traced is... The value is set to 172,800. Since disk hardware aging is a long-term physical wear and tear accumulation process, in order to completely eliminate the interference of a single, occasional severe business load flow, it is necessary to cover the natural day-night cycle of typical business cycles. The 172,800 time windows cover the physical span of the past 24 hours. In other embodiments, implementers can set the number of historical time windows to be traced according to the cyclical changes of the business, such as 86,400 corresponding to 12 hours or 345,600 corresponding to 48 hours.

[0041] In practical engineering implementation, in order to significantly reduce the system's computational resource overhead under high-frequency sampling, the weighted accumulation process of historical time windows in the calculation of the hysteresis decay index can be equivalently transformed into a first-order recursive filtering formula for real-time updating. The equivalent calculation satisfies the expression: ; In the formula, This indicates the lag recession index for the current time window; This represents the lagging recession index for the previous time window. This represents the recursive smoothing coefficient, and its value is equal to... This is used to equivalently approximate the aforementioned time decay factor. ; Indicates the first Load resistance characteristics within a time window; Indicates the first Temperature fluctuation characteristics within a time window; The smoothing penalty constant represents the temperature fluctuation characteristics.

[0042] Preferably, in practical engineering applications, to avoid misjudgment or abnormal accumulation of features due to data loss or absolute zero values ​​caused by the target disk being in a long-term dormant state with no business requests, this embodiment also includes a filtering mechanism for extreme idle conditions: before acquiring the load resistance characteristics in each time window, the input / output request rate of all sampling points in the current time window is verified. If the absolute value of the average input / output request rate in a consecutive set number of time windows is 0, it is determined that the target disk has entered a deep dormant period. At this time, the calculation of the load resistance characteristics of the current time window will be paused, and the hysteresis decay index of the current time window will be set. The lagging recession index remains at the previous time window. This means pausing feature updates until the input / output request rate returns to a non-zero state, then resuming the normal calculation and recursive update logic of the hysteresis decay index to ensure evaluation stability under extreme business surges. In this embodiment, when the absolute value of the average input / output request rate within 10 consecutive time windows is 0, the target disk is determined to enter a deep sleep period. In other embodiments, implementers can set the number of consecutive time windows required to determine that the target disk has entered a deep sleep period, such as 5 or 20, based on the actual business access surge patterns or the sleep triggering strategy of the storage hardware.

[0043] For example, Figure 4 This is a schematic diagram illustrating the changes in temperature fluctuation characteristics. Figure 5 This is a diagram illustrating the changes in the target disk hysteresis decay index, from... Figure 4 and Figure 5 It can be seen that during the period of interference from abnormal high temperature fluctuations, although the load resistance characteristics increased, the hysteresis decay index was effectively flattened under the suppression effect of temperature fluctuation characteristics, thus eliminating the illusion of thermodynamic performance. However, when the real physical medium underwent irreversible decay in the later stage, the hysteresis decay index showed a step-like cliff-like increase, which verified that the present invention can filter out environmental noise interference while accurately extracting real long-cycle physical losses.

[0044] S4. Based on the hysteresis decay index of the target disk and the preset baseline decay index, calculate the health score of the target disk, and execute the corresponding early warning and storage scheduling operations according to the health score.

[0045] Specifically, before the target disk is first put into operation, a standard random read / write test tool is used to apply full-load random read / write pressure to it, saturating the underlying communication channel of the target disk. Under this full-load pressure condition, the corresponding input / output request rate sequence, input / output delay sequence, and continuous temperature observation sequence are collected synchronously. Then, using the methods in steps S2 and S3, the hysteresis degradation index of the target disk in brand new condition is obtained and used as the benchmark degradation index. Benchmark Recession Index It reflects the hysteresis baseline of a specific physical medium model under optimal conditions, providing an anchored physical basis for subsequent comparisons of degradation levels.

[0046] For existing disks already running in the data center, since it is impossible to perform stress tests before the initial deployment, a reference disk cluster of the same batch and model as the target disks, which is in a healthy operating state, is retrieved. The average hysteresis decay index of the reference disk cluster during peak business periods is used as the benchmark decay index. To prevent baseline drift in the benchmark degradation index due to overall physical aging of the reference disk cluster, the reference disk cluster needs to be periodically screened to dynamically remove disks with health scores below a preset health threshold, ensuring the timeliness and accuracy of the benchmark value.

[0047] In other embodiments, implementers may also derive and set a benchmark degradation index based on the media theoretical response model provided by the storage device manufacturer.

[0048] Calculate the health score of the target disk: ; In the formula, Indicates the target disk is in the Health score corresponding to each time window; Indicates the first The lag recession index corresponding to each time window; Represents the hyperbolic tangent function; This represents the benchmark recession index.

[0049] in, ratio Mapping to a continuous convergent interval of [0,1] makes the evaluation results exhibit smooth and stable characteristics, when the th The lag recession index corresponding to each time window Much larger than the benchmark recession index At this point, it indicates that the current media hysteresis has significantly deviated from the baseline characteristics. At this time, the hyperbolic tangent mapping result approaches 1, causing the target disk to... Health score corresponding to each time window A value close to 0 indicates an extremely high risk of large-scale internal media failure on the target disk.

[0050] Furthermore, when the target disk's health score is continuous When a consecutive time window falls below the set health alarm baseline, a hardware replacement work order is generated, and a distributed replica load balancing migration task for the data on the target disk is initiated. In this embodiment, the number of consecutive time windows is... The time window is set to 600. Since a single time window spans 500 milliseconds, setting 600 time windows corresponds exactly to a 5-minute physical span. This span effectively filters out instantaneous fluctuations in scores caused by software layer blockages such as extremely short-term system process freezes and occasional bus network retransmissions, ensuring the stability and reliability of the final alarm decision. In other embodiments, implementers can set the number of continuous time windows according to the actual server business's requirements for fault warning sensitivity, for example, shortening it to 300 or extending it to 1200. In this embodiment, the health alarm baseline is set to 20. When the health score is below 20, the risk of data loss due to physical disk damage increases sharply. Setting it to 20 can balance data security and hardware replacement costs. In other embodiments, implementers can set the health alarm baseline according to the strictness of the actual server business's tolerance for data loss, for example, 10 or 30.

[0051] For example, Figure 6 This is a diagram illustrating the dynamic health score changes of the target disk. Figure 7 This is a schematic diagram comparing the monitoring effects of existing technologies with those of the present invention. Figure 7 The paper demonstrates a comparison between existing technologies that rely solely on average input-output latency for monitoring and the present invention's monitoring method based on health scores. Figure 7 As can be seen, existing technologies cannot accurately isolate short-term performance fluctuations caused by factors such as ambient temperature. Under abnormal high-temperature fluctuations, their average input / output latency will exceed the set limit, resulting in false judgments of physical degradation. In contrast, the health score of this invention, by incorporating temperature fluctuation characteristics for noise reduction, remains stable and above the health alarm baseline under abnormal high-temperature fluctuations. It only falls below the health alarm baseline when the underlying physical medium of the target disk experiences real degradation. This demonstrates that this invention overcomes the technical defects of traditional static threshold monitoring based on input / output latency, which is prone to missed and false alarms, when facing complex external environmental interference.

Claims

1. A method for monitoring the health of a server storage system, characterized in that, include: Collect request stream data from the target disk, perform statistics according to the set sampling period, and obtain the input / output request rate and the average completion time of input / output requests for each sampling period, which constitute the input / output request rate sequence and the input / output delay sequence, respectively. The input / output request rate sequence and the input / output delay sequence are divided into several time windows of equal length. Based on the transient increase in input / output delay and the transient drop in input / output request rate at the sampling points within each time window, the load resistance characteristics of the target disk within each time window are obtained. The physical media surface temperature of the target disk is collected synchronously for each sampling period to obtain the temperature fluctuation characteristics within each time window. The hysteresis degradation index of the target disk is determined based on the load resistance characteristics and the temperature fluctuation characteristics. Based on the target disk's hysteresis decay index and the preset baseline decay index, calculate the target disk's health score, and execute corresponding early warning and storage scheduling operations according to the health score.

2. The method for monitoring the health of a server storage system according to claim 1, characterized in that, The step of obtaining the input / output request rate corresponding to each sampling period includes: The number of input and output requests in each sampling period is statistically analyzed. The number of input and output requests in each sampling period is divided by the sampling period duration to obtain the input and output request rate corresponding to each sampling period.

3. The method for monitoring the health of a server storage system according to claim 1, characterized in that, The load resistance characteristic satisfies the expression: ; In the formula, Indicates the first Load resistance characteristics within a time window; Indicates the number of sampling points included within the time window; Indicates the first Within the first time window Input / output delay at each sampling point; Indicates the first The mean of the input and output delays of all sampling points within a time window; Indicates the first Within the first time window Input / output request rate per sampling point; Indicates the first The average input / output request rate of all sampling points within a time window; This indicates the operation of retrieving the maximum value; Represents an exponential function with the natural constant as its base; Indicates the sequence number of the sampling point within the time window; The smoothing penalty constant represents the input / output delay; This represents the smoothing penalty constant for the input / output request rate.

4. The method for monitoring the health of a server storage system according to claim 1, characterized in that, The acquisition of temperature fluctuation characteristics within each time window includes: The standard deviation of the physical media surface temperature of the target disk within the calculated time window is used as the temperature fluctuation characteristic within that time window.

5. The method for monitoring the health of a server storage system according to claim 1, characterized in that, The lagging recession index satisfies the expression: ; In the formula, Indicates the first The lag recession index corresponding to each time window; This indicates the number of historical time windows being traced. Indicates the tracing sequence number of the historical time window; Indicates the first Load resistance characteristics within a time window; Indicates the first Temperature fluctuation characteristics within a time window; Represents an exponential function with the natural constant as its base; The smoothing penalty constant represents the temperature fluctuation characteristics.

6. The method for monitoring the health of a server storage system according to claim 1, characterized in that, The lagging recession index satisfies the expression: ; In the formula, This indicates the lag recession index for the current time window; This represents the lagging recession index for the previous time window. Indicates the recursive smoothing coefficient; Indicates the first Load resistance characteristics within a time window; Indicates the first Temperature fluctuation characteristics within a time window; The smoothing penalty constant represents the temperature fluctuation characteristics.

7. The method for monitoring the health of a server storage system according to claim 1, characterized in that, The method for obtaining the benchmark recession index is as follows: In response to the fact that the target disk is a brand new disk before its first online operation, a full-load random read and write pressure is applied to the target disk; under the full-load pressure condition, the input and output request rate sequence, input and output delay sequence and continuous temperature observation sequence corresponding to the target disk are collected synchronously, the hysteresis decay index of the target disk under the full-load pressure condition is calculated, and the hysteresis decay index is used as the benchmark decay index.

8. The method for monitoring the health of a server storage system according to claim 1, characterized in that, The method for obtaining the benchmark recession index is as follows: In response to the target disk being an existing disk already running in the data center, a healthy reference disk cluster of the same batch and model as the target disk is retrieved, and the average lag and decay index of the healthy reference disk cluster during peak business periods is used as the benchmark decay index.

9. A method for monitoring the health of a server storage system according to any one of claims 1-8, characterized in that, The health score satisfies the expression: ; In the formula, Indicates the target disk is in the Health score corresponding to each time window; Indicates the first The lag recession index corresponding to each time window; Represents the hyperbolic tangent function; This represents the benchmark recession index.

10. The method for monitoring the health of a server storage system according to claim 1, characterized in that, The step of performing corresponding early warning and storage scheduling operations based on health scores includes: In response to the target disk's health score falling below the set health alarm baseline for a preset number of consecutive time windows, a hardware replacement work order is generated, and a distributed replica balance migration task for the data within the target disk is initiated.

Citation Information

Patent Citations

  • Method for predicting service life of solid state disk

    CN119782714A

  • Data storage management system and method based on big data

    CN120596460A

  • Hard disk health evaluation method and storage device

    WO2022227373A1