Mobile hard disk read-write load prediction and optimization method based on AI

Abstract

The invention relates to the technical field of mobile hard disk read-write load storage, in particular to a mobile hard disk read-write load prediction and optimization method based on AI, and the method comprises the steps: collecting read-write request load instantaneous change data, read-write operation period intensity data, file size distribution entropy data and LBA thermodynamic diagram data in real time; acquiring a dynamic weight factor of the load instantaneous change data by adopting an exponential weighted moving average model, and acquiring a main frequency band energy proportion of the periodic intensity data to calculate an I / O response coefficient; obtaining a variable coefficient of the I / O response coefficient, and calculating an output probability value between the I / O response coefficient and the entropy data by adopting kernel density estimation mode matching so as to calculate a load threat coefficient; and executing a dynamic resource optimization strategy based on the load threat coefficient in combination with the LBA thermodynamic diagram data. The invention aims to reduce the number of physical writing times and track seeking time in the peak period, reduce the actual access addressing delay and improve the data access fluency.

Description

Technical Field

[0001] This application relates to the field of mobile hard drive read / write load storage technology, specifically to an AI-based method for predicting and optimizing mobile hard drive read / write load. Background Technology

[0002] With the surge in performance demands of mobile storage devices and the normalization of massive data processing, traditional portable hard drives face severe challenges such as performance degradation, shortened lifespan, and energy waste due to their inability to predict sudden read / write loads. Against this backdrop, AI-based portable hard drive read / write load analysis has emerged and developed rapidly. It utilizes machine learning and deep learning algorithms to accurately predict the intensity and type of future short-term or periodic read / write requests by analyzing massive data streams such as historical and current I / O operation patterns, file sizes, and access frequencies in real time, providing users with a more stable and reliable high-performance storage experience.

[0003] When handling AI-based mobile hard drive read / write loads, the highly random nature of user behavior and the instantaneous switching between multi-task scenarios can easily cause load characteristic drift. This leads traditional time-series models, relying on historical patterns, to misjudge peak loads or underestimate latency-sensitive operations. Current mainstream solutions employ ensemble learning frameworks, which enhance the model's adaptability to non-linear abrupt changes by fusing multiple weak learners and assigning dynamic weights, thereby improving the robustness of predictions for sudden loads with limited computing power. However, incremental training of this technology requires accumulating sufficient new anomalous samples to update model parameters. When encountering completely unknown behavioral patterns, the ensemble model, lacking relevant feature encoding, still exhibits severe prediction bias, potentially leading to buffer overflows or a surge in response latency. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides an AI-based method for predicting and optimizing read / write loads of mobile hard drives, thereby resolving existing problems.

[0005] The AI-based mobile hard drive read / write load prediction and optimization method in this application adopts the following technical solution:

[0006] One embodiment of this application provides an AI-based method for predicting and optimizing read / write loads of a mobile hard drive, the method comprising the following steps:

[0007] Real-time collection of instantaneous changes in read / write request load, read / write operation cycle intensity data, file size distribution entropy data, and LBA heatmap data;

[0008] For the data collected at each acquisition time and all previous acquisition times, an exponentially weighted moving average model is used to obtain the dynamic weighting factor of the instantaneous load change data at each acquisition time, and the main frequency band energy ratio of the periodic intensity data is obtained to calculate the I / O response coefficient at each acquisition time.

[0009] The coefficient of variation of the I / O response coefficient at all acquisition times is obtained within each sliding window. The difference between the I / O response coefficient and the entropy data at all acquisition times is calculated using kernel density estimation and pattern matching, and converted into an output probability value to calculate the load threat coefficient at each acquisition time.

[0010] Based on the load threat coefficient and combined with LBA heatmap data, a dynamic resource optimization strategy is executed.

[0011] Preferably, in step 1, the instantaneous load change data is obtained by embedding a counting unit in front of the I / O scheduler. In each data acquisition cycle, the counting unit records the number of I / O requests in that cycle and calculates the absolute difference between the number of requests in the previous data acquisition cycle and outputs the original instantaneous load change data.

[0012] Preferably, in step 1, the periodic intensity data is obtained by calculating the coefficient of variation of all adjacent I / O request time intervals within each data acquisition cycle, and the mapping result with the natural constant as the base and the negative number of the coefficient of variation as the exponent is used as the periodic intensity value.

[0013] Preferably, the file size distribution entropy data in step 1 is obtained by statistically analyzing the Shannon entropy values ​​of all files with a size less than or equal to 4KB.

[0014] Preferably, the method for calculating the I / O response coefficient in step 2 is as follows:

[0015] Calculate the value of an exponential function with the natural constant as the base and the negative of the energy percentage of the main frequency band as the exponent;

[0016] The product of the mean of the dynamic weighting factors of all acquisition times prior to each acquisition time and the value of the exponential function is used as the I / O response coefficient for each acquisition time.

[0017] Preferably, the load threat coefficient at each acquisition moment is determined by the normalized value of the product of the mean of the coefficients of variation within all sliding windows and the output probability value.

[0018] Preferably, the execution method of the dynamic resource optimization strategy is as follows: when B is greater than or equal to the preset danger threshold, small files are preferentially written to the SSD cache pool and scheduled using the LRU algorithm; when B is less than the preset danger threshold, the LBA heatmaps at all collection times are clustered, and the clustering results are used to predict and identify densely accessed areas in real time, and the data blocks in the area are preloaded with hot data into the cache.

[0019] Preferably, the small file is a file with a size of less than or equal to 4KB.

[0020] Preferably, in step 4, when clustering the LBA heatmaps at all acquisition times, DBSCAN clustering algorithm is used, and the neighborhood radius of the algorithm is adaptively improved by the following formula: ε=ε0*(1+μ); where ε is the improved neighborhood radius, ε0 is the preset initial neighborhood radius, and μ is the instantaneous load change data at the current acquisition time.

[0021] Preferably, in step 4, the method of using clustering results to predict and identify densely accessed areas in real time and preloading hot data blocks in the area to the cache is as follows: the cluster with the most data points is recorded as the hot cluster, the centroid of the LBA hot cluster is calculated at all collection times, and the historical centroid movement trajectory of the hot cluster at all data collection times is used as input. The weighted moving average direction prediction algorithm is used to output the location of the hot cluster centroid in the LBA heatmap within the next 5 seconds, so as to predict the data area most likely to be accessed in the next 5 seconds, and the data blocks in the area are preloaded to the cache.

[0022] This application has at least the following beneficial effects:

[0023] 1. This application addresses the problem of load characteristic drift caused by the randomness of user behavior and multi-task switching in traditional time series models, which leads to misjudgment of sudden peaks and failure of latency-sensitive operations. It adopts an exponentially weighted moving average model and Fourier spectrum energy analysis to dynamically integrate the intensity of load bursts and periodic regularity, thereby solving the problem of prediction inaccuracy caused by load characteristic drift in information processing and storage, and accurately quantifying the overall load instability.

[0024] 2. This application addresses the problem that traditional models lack adaptability to unknown behavioral patterns, leading to cache overflow and uncontrolled delays in high-priority requests. It adopts sliding coefficient of variation and kernel density estimation (KDE) pattern matching to fuse the intensity of load mutations and the probability of pattern anomalies, eliminating the limitations of single indicators and comprehensively quantifying the complex threats caused by the strong randomness of user behavior, thereby improving robustness to unknown patterns.

[0025] 3. This application sets a threshold based on the load threat coefficient to determine the risk status of load fluctuations. For high-risk warning status, small files are written to the SSD cache pool first, and LRU scheduling + batch sequential disk writing to mechanical hard drives is adopted. For stable and predictable status, predicted data blocks are preloaded to the high-speed cache based on the ARIMA model, thereby reducing the number of physical writes and seek time during peak periods, reducing actual access addressing latency, and improving data access smoothness. Attached Figure Description

[0026] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart of the AI-based mobile hard drive read / write load prediction and optimization method provided in this application. Detailed Implementation

[0028] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive objective, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the AI-based mobile hard drive read / write load prediction and optimization method proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0030] The following section, in conjunction with the accompanying drawings, details the specific scheme of the AI-based mobile hard drive read / write load prediction and optimization method provided in this application.

[0031] One embodiment of this application provides an AI-based method for predicting and optimizing read / write loads of mobile hard drives.

[0032] Specifically, the following AI-based methods for predicting and optimizing external hard drive read / write load are provided. Please refer to [link / reference]. Figure 1 The method includes the following steps:

[0033] Step 1: Real-time collection of instantaneous changes in read / write request load, read / write operation cycle intensity data, file size distribution entropy data, and LBA heatmap data.

[0034] For all the data collected below, the data collection period is set to 1 second in this embodiment.

[0035] A lightweight monitoring module is deployed at the firmware layer of the mobile hard drive controller to collect instantaneous changes in read / write request load, which is used to quantify the intensity of these instantaneous load changes. Specifically, a counting unit is embedded in the front end of the I / O scheduler. In each data acquisition cycle, the counting unit records the number of I / O requests in that cycle and calculates the absolute difference between the number of requests and the number of requests in the previous data acquisition cycle, outputting the raw instantaneous load change data.

[0036] A periodic detection probe is embedded in the host-side file system driver layer to collect read / write operation cycle intensity data, which is used to identify the regularity of load fluctuations. Specifically, within each data acquisition cycle, the coefficient of variation of the time interval of all adjacent I / O requests within each data acquisition cycle is calculated, and the mapping result with the natural constant as the base and the negative number of the coefficient of variation as the exponent is used as the cycle intensity value.

[0037] An entropy calculation unit is integrated into the USB protocol parsing layer to collect file size distribution entropy data, which is used to quantify the randomness of access patterns. Specifically, within each collection cycle, the entropy calculation unit counts all files smaller than or equal to 4KB and calculates the distribution entropy value of all files smaller than or equal to 4KB within that collection cycle based on Shannon entropy, thus obtaining the entropy data.

[0038] A high-precision timer is installed on the host to timestamp all acquisition modules with microsecond-level precision. All data points record the precise start time of acquisition, and all sequences generate data points at the exact second according to the data acquisition cycle, ensuring that the same second-level timestamp contains the three types of data mentioned above (instantaneous changes in read / write request load, read / write operation cycle intensity, and file size distribution entropy). The three types of data acquired at all acquisition moments are normalized in real time to eliminate the influence of units. For each acquisition moment, the data of each type acquired at all previous acquisition moments are sorted to obtain the time-series data of each type of data at each acquisition moment.

[0039] Thus, the instantaneous load change sequence, periodic intensity sequence, and entropy sequence at each acquisition moment are obtained.

[0040] Furthermore, the starting LBA and length of each request are recorded at the I / O scheduler layer. Histogram statistics are used to aggregate the starting logical block address and length of each I / O request, and the distribution of physical storage areas covered by all I / O requests per second is statistically analyzed to obtain an LBA heatmap at each sampling time. This allows for real-time capture of the spatial distribution characteristics of the most frequently accessed physical storage areas per second. The acquisition of the LBA heatmap is a well-known technique and will not be elaborated upon here.

[0041] Step 2: For the data collected at each acquisition time and all previous acquisition times, use the exponentially weighted moving average model to obtain the dynamic weighting factor of the instantaneous load change data at each acquisition time, and obtain the main frequency band energy ratio of the periodic intensity data to calculate the I / O response coefficient at each acquisition time.

[0042] In real-time interactive scenarios involving massive amounts of data in information processing and storage, the strong randomness of user behavior and the instantaneous switching of multiple tasks make it difficult for traditional time-series models to adapt to load characteristic drift, leading to problems such as misjudgment of sudden read / write peaks and failure of response to latency-sensitive operations.

[0043] Therefore, using the instantaneous load change sequence as input, and employing the exponentially weighted moving average (EWMA) model, the dynamic weighting factor at the t-th acquisition time is set to... Where X t X t-1 The instantaneous load change data at the t-th and t-1-th acquisition times are used to quantify the severity of load fluctuations. This allows for a larger weighting when read / write request volumes change significantly within the time window of adjacent acquisition periods, making the model more attentive to sudden load characteristics. The mean of the dynamic weighting factors for all acquisition times prior to each acquisition time is calculated. This quantifies the intensity of load fluctuations, providing nonlinear weighting for bursty traffic in information processing and storage support. The exponentially weighted moving average model is a well-known technique and will not be elaborated further.

[0044] Subsequently, using the periodic intensity sequence as input, Fourier spectrum energy analysis is performed, with the frequency band set from 0.1Hz to 1Hz to cover second-level cycles, conforming to the typical load fluctuation characteristics of a portable hard drive. The output is the main frequency band energy percentage E, which characterizes the concentration of periodic loads to meet the low-latency requirements of information processing and storage. The Fourier spectrum energy analysis process is a well-known technique and will not be described in detail here.

[0045] Based on the above analysis, the I / O response coefficient A at each acquisition time is calculated using the following formula:

[0046]

[0047] in, It is the average of the dynamic weighting factors of all acquisition times before each acquisition time. It is used to sensitively capture the instantaneous changes in read and write request load between adjacent acquisition cycles, reflecting the instantaneous change intensity of the mobile hard drive load. The larger the value, the more drastic the load change. E is the main frequency band energy ratio of the periodic intensity sequence, which characterizes the concentration of periodic load. Its value represents the proportion of energy in the frequency band to the total energy. It is used to identify the regularity of the load. The larger the E value, the stronger the periodicity of the load, the more predictable the load, and the lower the volatility.

[0048] The I / O response coefficient A describes the degree of overall load instability of a portable hard drive in real-time data interaction scenarios involving massive amounts of data processing and storage, at each data acquisition moment. Through a dynamic weighted average of burst intensity and periodic concentration, it reflects the risk of load characteristic drift caused by the randomness of user behavior and instantaneous switching between multiple tasks. A larger A value indicates a higher risk of sudden read / write peaks and latency-sensitive operation failures.

[0049] Step 3: Obtain the coefficient of variation of the I / O response coefficients at all acquisition times within each sliding window. Use kernel density estimation mode matching to calculate the difference between the I / O response coefficients and entropy data at all acquisition times and convert it into an output probability value to calculate the load threat coefficient at each acquisition time.

[0050] In information processing and storage support services, portable hard drives need to cope with the real-time processing requirements of massive heterogeneous data streams. However, due to the strong spatiotemporal randomness of user behavior in cloud-edge-device collaborative scenarios, traditional prediction models cannot capture sudden shifts in load characteristics, leading to a surge in misjudgment rates of unknown access patterns, which in turn causes serious problems such as cache overflow and latency-sensitive operation blocking.

[0051] Therefore, the I / O response coefficients A calculated at all acquisition times prior to each acquisition time are sorted in ascending order of time to obtain the I / O response coefficient sequence for each acquisition time. Further, using the I / O response coefficient sequence as input, a sliding window analysis algorithm is used, with a window length of 30 seconds to balance real-time performance and statistical significance. The coefficient of variation within each sliding window is output, and the mean of the coefficients of variation within all sliding windows is calculated to quantify the abrupt change in load fluctuation. A larger mean indicates a more severe fluctuation. The sliding window analysis algorithm is a well-known technique and will not be described in detail here.

[0052] Subsequently, using the I / O response coefficient sequence and entropy sequence as input, pattern matching based on kernel density estimation (KDE) is employed with a Gaussian kernel bandwidth of 0.1, outputting the anomaly probability P. Specifically, the difference between the I / O response coefficient sequence and the entropy sequence is calculated using Kullback-Leibler divergence, and the difference is converted into an output probability value P using a Gaussian kernel, thus characterizing the degree to which the current load pattern deviates from the historical normal baseline. Both kernel density estimation (KDE)-based pattern matching and Kullback-Leibler divergence are well-known techniques and will not be elaborated further.

[0053] Based on the above analysis, the load threat coefficient B for each data collection moment is constructed, and the specific calculation formula is as follows:

[0054] B = sig(CV*P)

[0055] Wherein, sig() is the sigmoid function, CV is the mean coefficient of variation of the I / O response coefficient sequence within all sliding windows, which quantifies the relative dispersion of the sequence and characterizes the sudden intensity of load fluctuations in the external hard drive. It directly reflects the strong randomness of user behavior and the risk of instantaneous load jumps caused by instantaneous switching of multiple tasks. The larger the CV value, the more severe the load fluctuation and the higher the unpredictability. P is the output probability value between the I / O response coefficient sequence and the entropy value sequence, which measures the statistical distance of the current load pattern from the historical normal pattern and characterizes the threat level of the external hard drive encountering unknown or malicious behavior. The larger the P value, the more the current load pattern deviates from the normal baseline.

[0056] The load threat coefficient B comprehensively quantifies the overall risk level faced by portable hard drives in dealing with real-time interaction of massive heterogeneous data during information processing and storage. By integrating the intensity of sudden changes in load fluctuations and the degree of abnormal deviation in load patterns, it reflects the complex threats caused by the strong randomness of user behavior and instantaneous switching of multiple tasks.

[0057] Step 4: Based on the load threat coefficient and combined with LBA heatmap data, execute dynamic resource optimization strategies.

[0058] In information processing and storage support services, mobile storage devices face challenges due to the highly random nature of user behavior, load characteristic drift caused by instantaneous switching between multiple tasks, and prediction challenges posed by completely unknown behavioral patterns. This leads to severe biases in traditional or existing AI prediction models when encountering sudden and atypical loads. Such biases directly cause a surge in cache overflow risks and uncontrolled delays in high-priority request responses for mobile hard drives.

[0059] Therefore, a dynamic resource optimization strategy needs to be designed based on the load threat coefficient B, which accurately reflects the probability of load fluctuations and abnormal behavior, to intelligently adapt to the complex and ever-changing application scenarios in information processing and storage. The specific strategy design is as follows:

[0060] 1. The preset threat threshold T is 0.5;

[0061] 2. When B ≥ T, the system is considered to be in a high-risk warning state. To cope with the sudden load caused by a large number of small file writes, small files are prioritized for writing to a cache pool composed of high-performance SSDs. An LRU algorithm is used for intelligent scheduling, writing to the mechanical hard drive in batches and sequentially only when the load is stable. This significantly reduces the number of physical writes and seek time during peak periods, ensuring the timely write of critical transactional data in information processing and storage support services. Specifically, files with a size of 4KB or less are considered small files.

[0062] 2. When B < T, the current state is considered stable and predictable. Using the real-time LBA heatmap as input, the DBSCAN density clustering algorithm is used to cluster the real-time LBA heatmap. The initial minimum number of points and neighborhood radius are preset to 5 and 20, respectively. The neighborhood radius is adaptively improved based on load fluctuations. The specific improvement formula is ε = ε0 * (1 + μ), where ε is the improved neighborhood radius, ε0 is the preset initial neighborhood radius, and μ is the instantaneous load change data at the current acquisition time. The clustering granularity is dynamically adjusted according to the intensity of load fluctuations to ensure that the search range is expanded to capture scattered points during high burst access and that fine recognition capability is maintained during stable loads, thereby improving the accuracy and adaptability of dense access area identification. Finally, all clusters of the LBA heatmap are output to identify dense access areas in real time. The DBSCAN density clustering algorithm is a well-known technology and will not be described in detail here.

[0063] Furthermore, the cluster with the most data points is designated as the hotspot cluster. The centroid of the LBA hotspot cluster is calculated at all data collection times. Using the historical centroid movement trajectory of the hotspot cluster at all data collection times as input, a weighted moving average direction prediction algorithm is applied. Specifically, a time-exponential decay weight is used, with a decay factor of 0.8. The algorithm outputs the location of the hotspot cluster centroid in the LBA heatmap within the next 5 seconds to predict the most likely data region to be accessed within the next 5 seconds. Data blocks within the predicted region are proactively preloaded into a high-speed cache, such as RAM or SSD cache, significantly reducing the physical addressing wait time for subsequent actual accesses and improving the smoothness of data access for users in information processing and storage support services. The use of the weighted moving average direction prediction algorithm is a well-known technology and will not be elaborated upon here.

[0064] Thus, the invention of an AI-based method for predicting and optimizing read / write loads of portable hard drives has been completed.

[0065] The above technical features constitute the preferred embodiment of this application, which has strong adaptability and the best implementation effect. Unnecessary technical features can be added or removed according to actual needs to meet the needs of different situations.

Claims

1. An AI-based method for predicting and optimizing read / write loads of portable hard drives, characterized in that: The method includes the following steps: Real-time collection of instantaneous changes in read / write request load, read / write operation cycle intensity data, file size distribution entropy data, and LBA heatmap data; For the data collected at each acquisition time and all previous acquisition times, an exponentially weighted moving average model is used to obtain the dynamic weighting factor of the instantaneous load change data at each acquisition time, and the main frequency band energy ratio of the periodic intensity data is obtained to calculate the I / O response coefficient at each acquisition time. The coefficient of variation of the I / O response coefficient at all acquisition times is obtained within each sliding window. The difference between the I / O response coefficient and the entropy data at all acquisition times is calculated using kernel density estimation and pattern matching, and converted into an output probability value to calculate the load threat coefficient at each acquisition time. Based on the load threat coefficient and combined with LBA heatmap data, a dynamic resource optimization strategy is executed.

2. The AI-based mobile hard drive read / write load prediction and optimization method as described in claim 1, characterized in that, In step 1, the instantaneous load change data is processed by a counting unit embedded in the front end of the I / O scheduler. In each data acquisition cycle, the counting unit records the number of I / O requests in that cycle and calculates the absolute difference between the number of requests in the previous data acquisition cycle, and outputs the original instantaneous load change data.

3. The AI-based mobile hard drive read / write load prediction and optimization method as described in claim 1, characterized in that, In step 1, the periodic intensity data is obtained by calculating the coefficient of variation of all adjacent I / O request time intervals within each data acquisition cycle, and using the mapping result with the natural constant as the base and the negative number of the coefficient of variation as the exponent, as the periodic intensity value.

4. The AI-based mobile hard drive read / write load prediction and optimization method as described in claim 1, characterized in that, In step 1, the file size distribution entropy data is obtained by statistically analyzing the Shannon entropy values ​​of all files with a size less than or equal to 4KB.

5. The AI-based mobile hard drive read / write load prediction and optimization method as described in claim 1, characterized in that, The method for calculating the I / O response coefficient in step 2 is as follows: Calculate the value of an exponential function with the natural constant as the base and the negative of the energy percentage of the main frequency band as the exponent; The product of the mean of the dynamic weighting factors of all acquisition times prior to each acquisition time and the value of the exponential function is used as the I / O response coefficient for each acquisition time.

6. The AI-based mobile hard drive read / write load prediction and optimization method as described in claim 1, characterized in that, The load threat coefficient at each acquisition moment is determined by the normalized value of the product of the mean of the coefficients of variation within all sliding windows and the output probability value.

7. The AI-based mobile hard drive read / write load prediction and optimization method as described in claim 1, characterized in that, The dynamic resource optimization strategy is implemented as follows: when B is greater than or equal to the preset danger threshold, small files are written to the SSD cache pool first and scheduled using the LRU algorithm; when B is less than the preset danger threshold, the LBA heatmaps at all collection times are clustered, and the clustering results are used to predict and identify densely accessed areas in real time, and the data blocks in the area are preloaded with hot data into the cache.

8. The AI-based mobile hard drive read / write load prediction and optimization method as described in claim 7, characterized in that, The small file is a file with a size of 4KB or less.

9. The AI-based mobile hard drive read / write load prediction and optimization method as described in claim 7, characterized in that, In step 4, when clustering the LBA heatmaps at all acquisition times, DBSCAN clustering algorithm is used, and the neighborhood radius of the algorithm is adaptively improved by formula: ε=ε0*(1+μ); where ε is the improved neighborhood radius, ε0 is the preset initial neighborhood radius, and μ is the instantaneous load change data at the current acquisition time.

10. The AI-based mobile hard drive read / write load prediction and optimization method as described in claim 7, characterized in that, In step 4, the clustering results are used to predict and identify densely accessed areas in real time. The method for preloading hot data in the data blocks within the area to the cache is as follows: the cluster with the most data points is recorded as the hot cluster. The centroid of the LBA hot cluster is calculated at all data collection times. The historical centroid movement trajectory of the hot cluster at all data collection times is used as input. The weighted moving average direction prediction algorithm is used to output the location of the LBA heatmap where the centroid of the hot cluster is located within the next 5 seconds. This is to predict the data area most likely to be accessed within the next 5 seconds and preload the data blocks within the area to the cache.

Images