Method for load balancing of a cluster of hosting servers based on prediction of traffic characteristics

Abstract

This invention relates to the field of server cluster load balancing technology, and particularly to a method for load balancing of managed server clusters based on business characteristic prediction. The method includes: acquiring time-series CPU utilization data of each node in the server cluster; calculating load burstability based on the CPU utilization time-series data and constructing a load burstability sequence; extracting autocorrelation characteristics from the load burstability sequence to construct a time-series clustering sequence; extracting the current time-series clustering measured sequence and the historical time-series clustering reference sequence, and evaluating the prediction reliability based on the deviation between the measured sequence and the reference sequence to construct a periodic consistency; dynamically adjusting the fusion weight parameters of historical periodic prediction values ​​and real-time short-term prediction values ​​according to the periodic consistency to obtain the fused predicted load, and scheduling business requests to corresponding nodes based on the fused predicted load. This invention achieves adaptive switching of strategies under peak resource pre-allocation and abnormal disturbances.

Description

Technical Field

[0001] This invention relates to the field of server cluster load balancing technology, and in particular to a method for load balancing of managed server clusters based on business characteristic prediction. Background Technology

[0002] With the rapid development of cloud computing and hosting services, managed server clusters have become the core infrastructure for large-scale internet businesses such as e-commerce transactions, financial settlements, and online education. In actual operation, user requests in these business scenarios are not evenly distributed, but are driven by business events such as promotional activities, transaction periods, and concentrated logins before classes, exhibiting significant bursty and periodic characteristics. This leads to server clusters facing enormous load pressure during peak business periods, and how to achieve efficient and balanced cluster load scheduling has become a key problem that the industry urgently needs to solve.

[0003] Chinese patent application CN121411984A discloses an adaptive load balancing method and system based on multi-dimensional intelligent perception. The method includes: periodically collecting and preprocessing multi-dimensional index data of the load status of each node through monitoring agents deployed on each node in a server cluster to obtain standardized status feature data, and then determining the dynamic weight of each dimension index using a preset weight calculation model; calculating the comprehensive load index of each node based on the dynamic weight of each dimension index, and then classifying each node into different load levels; dynamically generating a load balancing strategy based on the current load distribution and load level, combined with the time-series prediction results of historical load data; and calculating the load to be migrated according to the load balancing strategy, selecting the optimal target migration node, and executing the load migration operation.

[0004] Existing cluster load balancing technologies mainly include round-robin algorithms, weighted round-robin algorithms, least connections algorithms, and their variants. These algorithms share the characteristic of using the instantaneous load state of each node, such as current CPU utilization or active connections, as the scheduling basis, allocating tasks to the node with the lowest current load or the highest weight upon each request arrival. However, these methods are essentially passive response scheduling: scheduling decisions rely solely on the current single-point state information, lacking any predictive ability for future load evolution trends. When business peaks occur, these algorithms can only trigger task migration or traffic reallocation after node loads have exceeded normal levels, resulting in significant time lag in scheduling decisions. This makes it difficult to pre-allocate resources before peak periods arrive, leading to severe overload of local nodes in the cluster and a sharp increase in service response latency. Summary of the Invention

[0005] To address the problem that existing server cluster load balancing methods rely solely on instantaneous load status for passive, reactive scheduling, lacking the ability to predict future load trends, leading to scheduling delays during peak business periods and causing node overload and a sharp increase in service latency, this invention provides a managed server cluster load balancing method based on business characteristic prediction.

[0006] This invention provides a load balancing method for managed server clusters based on business characteristic prediction, employing the following technical solution: A load balancing method for managed server clusters based on business characteristics prediction includes the following steps: The system acquires CPU utilization time-series data for each node in the server cluster, calculates load burstability based on this data to reflect burst characteristics, and constructs a load burstability sequence. It then extracts autocorrelation characteristics from the load burstability sequence to construct a time-series clustering sequence. Finally, it extracts the measured time-series clustering sequence for the current moment and a historical time-series clustering reference sequence, assesses prediction reliability based on the deviation between the measured and reference sequences, and constructs periodic consistency. Based on the periodic consistency, it dynamically adjusts the fusion weight parameters of historical periodic predictions and real-time short-term predictions to obtain the fused predicted load, and schedules business requests to the corresponding nodes according to the fused predicted load. The method for constructing periodic consistency is as follows: extract the root mean square error of the corresponding segments of the measured sequence and the reference sequence. Periodic consistency is negatively correlated with root mean square error.

[0007] Compared to traditional load balancing algorithms that rely on instantaneous states for passive response, this algorithm mines the deep characteristics of business requests by constructing load burstability and temporal clustering. Then, before the business peak arrives, it adaptively adjusts the prediction weights using periodic consistency, achieving forward-looking resource pre-allocation. This effectively avoids the problems of local cluster overload and sharp increase in service response latency caused by scheduling lag.

[0008] Preferably, the method for calculating the load burst degree, which reflects burst characteristics, based on CPU utilization time-series data is as follows: within a preset window, extract the maximum value, mean, standard deviation, and number of abnormal sampling points exceeding the threshold of the CPU utilization time-series data; calculate the load burst degree based on the difference between the maximum value and the mean, the standard deviation, and the ratio of the number of abnormal sampling points to the total number of sampling points; the load burst degree is positively correlated with the number of abnormal sampling points and positively correlated with the calculated difference.

[0009] Compared to methods that judge load solely based on a single instantaneous utilization threshold, this solution integrates two dimensions—peak amplitude and burst frequency—in business scenarios such as e-commerce promotions that cause short-term, drastic fluctuations. This allows for a more accurate distinction between normal, short-term spikes and sustained high loads, avoiding scheduling misjudgments caused by traditional algorithms' inability to identify genuine bursts.

[0010] Preferably, the expression for load burst degree is: the quotient is obtained by dividing the difference between the maximum value and the mean by the standard deviation, and the load burst degree is obtained by multiplying the ratio of the number of abnormal sampling points to the total number of sampling points by the quotient.

[0011] Compared to conventional qualitative analysis or simple linear assessment, by combining standard deviation and the proportion of outlier sampling points to construct a dimensionless quantitative indicator, it is possible to capture the dramatic changes in cluster load caused by a surge in high-concurrency requests such as financial settlement, further improving the quantitative accuracy of emergency assessment and providing a reliable basis for the accurate allocation of subsequent computing resources.

[0012] Preferably, the steps for constructing the temporal clustering sequence are as follows: Set a sliding analysis window and extract the first-order autocorrelation coefficient of the corresponding load burst degree subsequence within the sliding analysis window; count the longest continuous time step in which the load burst degree sequence continuously exceeds the set burst threshold within the sliding analysis window; add one to the longest continuous time step and perform a logarithmic transformation, and multiply the result of the logarithmic transformation with the first-order autocorrelation coefficient to obtain the temporal clustering degree.

[0013] Compared to traditional statistical methods such as moving averages, which are insensitive to time-series structures, the introduction of autocorrelation coefficients and burst consecutive steps can effectively distinguish between occasional isolated request spikes and continuous clustered bursts caused by business session backlogs. When faced with complex concurrent operations such as a large number of users placing orders continuously, it improves the ability to identify complex business load patterns and avoids subsequent resource misallocation or waste.

[0014] Preferably, the method for calculating the burst threshold is as follows: the burst threshold is automatically determined by applying the Otsu threshold segmentation method to the historical load burst degree sequence.

[0015] Compared to methods that rely on human experience to preset fixed thresholds, the Otsu threshold segmentation method can adaptively determine the segmentation criteria based on the actual distribution of historical load fluctuations. When facing different types of Internet business scenarios, it eliminates the deviation of human parameter settings and improves the automation and objective accuracy of sudden state identification.

[0016] Preferably, the method for calculating period consistency is as follows: extract the Pearson correlation coefficient between the corresponding segments of the measured sequence and the reference sequence, and calculate the amplitude consistency penalty factor corresponding to the root mean square error; multiply the Pearson correlation coefficient and the amplitude consistency penalty factor to obtain the period consistency.

[0017] Compared to algorithms that rely solely on historical averages as a prediction benchmark, this algorithm, by comprehensively evaluating the similarity in form and amplitude between real-time load and historical data, can more accurately capture aperiodic abnormal disturbances in the actual business environment. In the event of abnormal traffic attacks or sudden marketing events, it effectively reduces the risk of inaccurate predictions caused by relying on historical data.

[0018] Preferably, the calculation method of the amplitude consistency penalty factor is as follows: extract the arithmetic mean of the corresponding segment of the reference sequence, divide the root mean square error by the arithmetic mean to obtain the normalized deviation; construct a factor with the natural constant as the base and the negative number of the normalized deviation as the exponent as the amplitude consistency penalty factor.

[0019] Compared to the approach of directly using absolute deviation to measure periodic differences, this approach effectively eliminates the interference of absolute load amplitude differences between business systems of different sizes by introducing the arithmetic mean for normalization and constructing an exponential penalty factor.

[0020] Preferably, the step of dynamically adjusting the fusion weight parameters of historical periodic forecasts and real-time short-term forecasts based on periodic consistency to obtain the fused forecast load is as follows: The periodic consistency is used as the prediction confidence coefficient; the prediction confidence coefficient is used as the first weight of the predicted value of the historical periodic pattern, and the difference between the prediction confidence coefficient and the prediction confidence coefficient is used as the second weight of the short-term predicted value of the real-time state; the first weight is used to weight the historical periodic predicted value obtained based on the linear regression method, and the second weight is used to weight the real-time short-term predicted value obtained based on the exponential smoothing method, and the two are added together to obtain the fused prediction load.

[0021] Compared to methods that use fixed weights for prediction, this method smoothly switches between historical pattern prediction and real-time short-term prediction based on the consistency of dynamic cycles. When business is stable, it makes full use of the forward-looking advantage of historical patterns, while quickly switching to a responsive real-time conservative scheduling mode during non-periodic emergencies. This achieves the best adaptive fit of the prediction strategy under the impact of large and varied business traffic.

[0022] Preferably, the steps for scheduling service requests to corresponding nodes based on the converged predicted load are as follows: newly arriving service requests are routed to the node with the largest remaining predicted capacity according to the principle of minimum predicted load priority; when there is a node whose converged predicted load exceeds the first proportion of the preset rated capacity, the task is preemptively hot-migrated to a standby node with both current and predicted loads being low; when there is a node whose converged predicted load is continuously lower than the second proportion of the preset rated capacity, the node is triggered to hibernate or reduce its frequency.

[0023] Compared to static traffic distribution solutions that lack lifecycle management, by introducing a control mechanism that links proportional thresholds, it can not only proactively trigger hot migration to ensure the continuity of core businesses when high load warnings are issued, but also reduce the frequency of idle nodes to hibernate during off-peak periods. This ensures the overall high availability of large-scale clusters while reducing energy consumption costs in long-term operation.

[0024] Preferably, after obtaining the CPU utilization time-series data of each node in the server cluster, the process further includes preprocessing the CPU utilization time-series data. The preprocessing steps are: for missing sampling points, linear interpolation is used to complete them; the CPU utilization time-series data is Z-score standardized; and... The criteria detect and replace outliers, and the moving median is used to replace outliers.

[0025] Compared to directly inputting raw monitoring data into the scheduling process, systematic interpolation and outlier cleaning effectively eliminate interference from dirty data caused by network jitter or node restarts. In complex physical node cluster monitoring scenarios, this approach ensures the data quality of subsequent time-series analysis models from the source, improving the stability of the overall scheduling algorithm and the reliability of the final decision execution.

[0026] The present invention has the following technical effects: This invention proposes a load balancing method for managed server clusters based on business characteristic prediction. By constructing load burstability and temporal clustering, it deeply mines business characteristics and constructs periodic consistency based on the deviation between measured and historical sequences. Based on this, it dynamically adjusts the fusion weight of historical period and real-time short-term predictions, outputs the fused predicted load for forward-looking node scheduling. This invention breaks the limitations of passive scheduling, realizes the pre-allocation of resources during peak periods and the adaptive switching of strategies under abnormal disturbances, and effectively balances the high availability guarantee and low energy consumption operation of the cluster system. Attached Figure Description

[0027] Figure 1 This is a flowchart of the hosting server cluster load balancing method based on business characteristic prediction according to the present invention.

[0028] Figure 2 This is a schematic diagram illustrating the effect of the temporal clustering degree of the present invention on distinguishing between two types of load burst modes.

[0029] Figure 3 This is a schematic diagram of the dynamic response curves of the periodicity of the present invention under normal periodic and non-periodic disturbance scenarios. Detailed Implementation

[0030] 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.

[0031] This invention discloses a load balancing method for managed server clusters based on business characteristic prediction, referring to... Figure 1The process includes steps S1-S5, as detailed below: Step S1: Data acquisition and preprocessing.

[0032] To achieve load balancing of the managed server cluster based on business forecasting, it is necessary to collect CPU utilization data of each server node. This is done by deploying PrometheusNodeExporter on each managed server, which periodically pulls the data at a 5-second sampling interval, and then integrating the real-time collected data with historical CPU utilization time-series data into the time-series database InfluxDB for storage management.

[0033] The collected CPU utilization data is preprocessed. Specifically, for missing sampling points caused by network jitter or node restarts, linear interpolation is used to fill in the gaps. Then, the CPU utilization data is Z-score standardized. The criteria detect and replace outliers, filling them with the sliding median.

[0034] Step S2: Build load burstability.

[0035] Because the request arrival process in business systems is not a stationary random process but exhibits significant burst characteristics, according to the M / G / 1 queuing theory model, when the request arrival rate increases sharply in a short period of time, the number of waiting tasks in the server queue will accumulate rapidly in a non-linear manner, and the CPU utilization will rise sharply within a very short time window, forming a sharp load spike. This phenomenon has been recorded in typical scenarios such as e-commerce flash sales and financial market opening settlements. Unlike stable loads, load spikes are characterized by high amplitude, short duration, and significant deviation from the baseline. During the spike, the maximum CPU utilization is significantly higher than the window mean, and the proportion of sampling points exceeding the mean plus 1.5 standard deviations is significantly increased. Existing weighted round-robin load balancing algorithms rely only on instantaneous load state scheduling and cannot distinguish between short-term spikes and continuous high loads, resulting in excessive tasks being scheduled to near-saturated nodes during spikes, triggering a chain reaction of overload. Therefore, it is necessary to construct a load burst degree... This is to reflect the above characteristics.

[0036] For the preprocessed CPU utilization time-series data, a structure of length is constructed along the historical direction with time t as the endpoint. The window contains the average CPU utilization. Maximum value Standard deviation and more Number of sampling points As a measure of load burstiness The basic parameter, the expression for load burstability, is:

[0037] In the formula, for Load burstiness within a time window; for The maximum CPU utilization within the time window; for The average CPU utilization within the time window; for The standard deviation of CPU utilization within the time window, all three are directly calculated from the preprocessed data; for CPU utilization exceeded within the time window The number of sampling points, with a coefficient of 1.5, is taken from the empirical rule of normal distribution, and can be adjusted in the range of 1.0 to 2.0 in practice; This represents the total number of sampling points within the window. In this embodiment, the total number of sampling points is 60, which is used to... The proportion of burst sampling points normalized to dimensionless.

[0038] When the load burst characteristics are more obvious The larger, The higher, The larger the value, the more sensitive it is to responses to increased bursts of power. Compared to existing technologies that only utilize instantaneous utilization, It integrates two dimensions, peak amplitude and burst frequency, and can distinguish between two load modes: short-term peak and continuous high load. This avoids scheduling misjudgment caused by the inability of traditional algorithms to identify bursts, and has higher load state representation accuracy.

[0039] Step S3: Construct temporal clustering degree.

[0040] Business-driven load bursts do not occur randomly and independently, but exhibit significant temporal clustering characteristics. According to autoregressive time series theory, if a high-intensity load burst occurs at a certain moment, due to the session persistence of business requests, such as continuous user browsing, ordering, and payment operations, as well as the backlog effect of internal task queues, the high burst state often persists or reappears in subsequent time steps, manifesting as a significant positive autocorrelation between adjacent high-value points in the load burst degree sequence. Taking e-commerce flash sales as an example, a large number of users' continuous operations will continuously generate high-intensity bursts within tens of seconds to several minutes. The load burst degree sequence during this period shows multiple clusters of consecutive high values, rather than isolated spikes. If scheduling decisions fail to identify that the current burst is in the early stage of a cluster, it will underestimate the subsequent load pressure, leading to insufficient resource allocation; if it is at the end, unnecessary resource expansion will cause waste. Existing methods based on moving averages or exponential weighted averages are not sensitive to this temporal structure and cannot effectively distinguish between clustered bursts and isolated, short-lived bursts. Therefore, it is necessary to improve the load burst degree... Constructing temporal clustering based on In order to capture this deep and unique characteristic.

[0041] At each moment Construct a load burst sequence, use this load burst sequence as input, and set the sliding analysis window length. For each of the 10 time steps, the first-order autocorrelation coefficient of the corresponding load burst degree subsequence is calculated within each sliding analysis window. Through history The threshold is determined by applying Otsu's threshold segmentation method to the sequence. Within the statistics window Continue to exceed Longest continuous time steps ,and Together as a calculation of temporal clustering degree The input parameter, the expression for temporal clustering degree, is:

[0042] In the formula, for The temporal clustering degree of each moment comprehensively reflects the current window. The degree of temporal clustering of burst events in a sequence; for In the time analysis window The first-order autocorrelation coefficient of the subsequence, and its range of values. ,Depend on The subsequence is calculated directly; for Within the time window Continuously exceeding the threshold The longest continuous segment length, in time steps, is a dimensionless quantity; Through history The sequence was automatically obtained using the Otsu threshold segmentation method. right Perform logarithmic transformation to suppress extremely long clusters Excessive amplification, where the constant 1 ensures hour Reset to zero.

[0043] When the time-order clustering characteristics are more obvious, The stronger the positive autocorrelation of the sequence, The closer it is to 1, the more frequent the high-instance phases. The larger, The larger the value, the more accurately it can quantify the increase in the degree of sudden clustering. By introducing a first-order autocorrelation coefficient, it was captured The temporal structure information of the sequence can distinguish between isolated, short-lived bursts and continuous, clustered bursts, providing a more structured feature basis for the prediction and scheduling of subsequent steps, and improving the ability to identify complex business load patterns.

[0044] Step S4: Build cycle consistency.

[0045] Cluster load spikes not only exhibit clustering but also show significant periodicity. Based on the STL seasonality decomposition method in time series analysis, business system load can be decomposed into trend, periodic, and residual components. The periodic component reflects rhythmic characteristics such as daily and weekly cycles; for example, e-commerce platforms experience peak loads concentrated between 9-11 AM and 7-9 PM, while financial systems are concentrated around market opening and closing times. Due to the high degree of time-series clustering... Depend on It was calculated that, Driven by business cycles, therefore The sequence also exhibits a high-low alternation pattern synchronized with the business cycle. If the current... Sequence and Historical Period The degree of sequence similarity can determine the extent to which the current load follows historical cyclical patterns, and dynamically assess the reliability of forward-looking predictions using historical cycles. Existing algorithms use the historical load average as a prediction benchmark, failing to utilize sequence morphological similarity or dynamically assess the current applicability of historical patterns, leading to severe prediction inaccuracies when aperiodic disturbances (such as sudden abnormal traffic) occur. Therefore, it is necessary to... Building on the basis of periodic consistency The current prediction reliability of the historical cycle pattern is dynamically quantified and used as a direct input for improving the prediction weight parameters of the algorithm.

[0046] A temporal clustering sequence is constructed using the temporal clustering at each time point. The reference period length is then set as the input. (Daily cycle) Each time step corresponds to the number of time steps in a day when the sampling interval is 5 minutes. Extract the nearest time step aligned with the current time from the historical time series clustering sequence. The historical reference sequence is obtained by taking the average of the contemporaneous subsequences of each period at each time step. For example, for the first h moments of day a, extract the first h historical data from days a-1 to a-7 respectively, and take the average of the data at the same corresponding moment in the historical data to obtain the historical reference sequence; extract the data before the current moment... Each time step Measured sequence ,calculate and Pearson correlation coefficient of the corresponding segment Root mean square error ,as well as Arithmetic mean of the corresponding segment Together as computing The intermediate parameter, the expression for periodic consistency, is:

[0047] In the formula, for The periodicity of time, with a value range of 1. The actual effective interval is ; for and The Pearson correlation coefficient for corresponding segments reflects the morphological similarity between two sequences, and its value ranges from [value range missing]. It is directly calculated from the two sequences; for and The root mean square error of the corresponding segment reflects the magnitude deviation between the two sequences, and is calculated from the root mean square of the point-by-point difference between the two sequences. for The arithmetic mean of the corresponding segment, from history The sequence is obtained directly through calculation; and The division yields a dimensionless normalization bias, eliminating the differences between business systems. The impact of absolute amplitude differences; This is the amplitude consistency penalty factor, with a value range of... The larger the amplitude deviation, the closer the factor approaches zero. The stronger the suppression, the better the joint evaluation of morphological similarity and amplitude consistency can be achieved.

[0048] When the load cycle pattern is more stable, the current The closer the sequence matches the historical period, the better. The closer to 1, The smaller the value, the closer the penalty factor is to 1. The larger the load, the lower the load; conversely, when the load is affected by aperiodic disturbances, the lower the load. Decline Increase The value decreases significantly, approaching zero. Compared to existing fixed-weight historical mean prediction methods, It achieves dynamic quantification of the reliability of historical patterns, endows the scheduling algorithm with the ability to adaptively switch prediction strategies under different business scenarios, and reduces prediction errors in non-periodic disturbance scenarios.

[0049] Step S5: Improve the algorithm and implement cluster load balancing.

[0050] Based on the theory of combined forecasting, and according to the periodic consistency The fusion weight parameters are dynamically and adaptively adjusted during the weighted fusion process of historical periodic forecasts and real-time short-term forecasts.

[0051] Within each scheduling decision cycle, with As a prediction confidence coefficient (Right now value range A weighted fusion strategy is used to combine predicted values ​​based on historical cyclical patterns. Compared with short-term forecasts based on real-time conditions The load is fused to obtain the fused predicted load. The fusion formula is ; CPU utilization data was predicted using linear regression. The CPU utilization data was predicted using the exponential smoothing method.

[0052] The aforementioned weighted fusion strategy is based on combinatorial forecasting theory: when historical cyclical patterns are stable, i.e., the cyclical consistency is high. At higher load levels, the load evolution closely matches historical patterns, and the linear regression's ability to fit periodic patterns results in smaller prediction errors, making its higher weighting reasonable; however, when aperiodic disturbances occur, i.e., periodic consistency... When the load is low, the current load deviates from the historical pattern, linear regression fails, while exponential smoothing relies on recent real-time sequences and responds faster to sudden changes, so assigning it greater weight is reasonable.

[0053] when At higher levels, Approaching 1, the algorithm mainly relies on periodic patterns for forward-looking resource pre-allocation; when At lower levels, When the value approaches zero, switch to a conservative scheduling mode based primarily on real-time status to avoid resource misallocation caused by the failure of historical patterns. Meanwhile, When the load is large, appropriately increase the urgency of offloading tasks to low-load nodes. The longer the duration of resource reservation, the better the coordinated response to the three types of load characteristics: sudden, concentrated, and periodic.

[0054] Based on the improved algorithm, the system implements cluster load balancing based on business forecasting according to the following process: Within each scheduling decision cycle, the system sequentially completes data collection and preprocessing, and calculates the load burst level. Temporal clustering Periodic consistency and output fused predicted load. In this implementation, the scheduling decision cycle is 30 seconds; the scheduling controller according to... Based on the current actual load of each node, newly arriving service requests are routed to the node with the largest predicted remaining capacity according to the principle of prioritizing the minimum predicted load. For nodes whose predicted load is about to exceed 80% of their rated capacity, task hot migration is triggered in advance to standby nodes with lower current and predicted loads. For nodes whose predicted load is consistently below 60% of their rated capacity, node hibernation or frequency reduction is triggered to save energy. Through the above closed-loop control process, the system can complete proactive resource allocation before the arrival of business peaks and quickly switch to conservative strategies in non-periodic sudden scenarios, effectively avoiding local overload and service degradation problems caused by uneven load distribution.

[0055] Figure 2 Temporal clustering degree A diagram illustrating the differentiation effect between the two types of load burst modes. (For example...) Figure 2 As shown, for isolated, brief outbreaks, Although a peak occurred, It quickly fell back to a low value; for clustered outbreaks, The high level of aggregation throughout the entire aggregation period accurately reflects the sustained aggregation degree of the sudden event, while the traditional moving average method shows no significant difference in output between the two modes, thus verifying the temporal aggregation degree. Advantages in identifying complex load timing structures.

[0056] Figure 3 For periodic consistency Dynamic response curves under normal periodic and non-periodic disturbance scenarios. For example... Figure 3 As shown, on a historically stable normal workday, The algorithm maintains a high level for an extended period, with a relatively large... Value priority relies on historical cycles for forward-looking pre-scheduling; when aperiodic disturbances occur... The price dropped rapidly, triggering a switch to a real-time conservative mode in the scheduling strategy. After the disturbance was eliminated... Automatic recovery verified the consistency of the cycle. The adaptive effect of dynamic quantification of prediction confidence.

[0057] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for load balancing of a cluster of hosting servers based on prediction of traffic characteristics, characterized in that, Including the following steps: The system acquires CPU utilization time-series data for each node in the server cluster, calculates load burstability based on this data to reflect burst characteristics, and constructs a load burstability sequence. It then extracts autocorrelation characteristics from the load burstability sequence to construct a time-series clustering sequence. Finally, it extracts the measured time-series clustering sequence for the current moment and a historical time-series clustering reference sequence, assesses prediction reliability based on the deviation between the measured and reference sequences, and constructs periodic consistency. Based on the periodic consistency, it dynamically adjusts the fusion weight parameters of historical periodic predictions and real-time short-term predictions to obtain the fused predicted load, and schedules business requests to the corresponding nodes according to the fused predicted load. The method for constructing periodic consistency is as follows: extract the root mean square error of the corresponding segments of the measured sequence and the reference sequence. Periodic consistency is negatively correlated with root mean square error.

2. The method for load balancing of a managed server cluster based on business characteristic prediction according to claim 1, characterized in that, The method for calculating the load burst degree, which reflects burst characteristics, based on CPU utilization time-series data is as follows: within a preset window, extract the maximum value, mean, standard deviation, and number of abnormal sampling points exceeding the threshold of the CPU utilization time-series data; calculate the load burst degree based on the difference between the maximum value and the mean, the standard deviation, and the ratio of the number of abnormal sampling points to the total number of sampling points. The load burst rate is positively correlated with the number of abnormal sampling points and with the calculated difference.

3. The method for load balancing of a managed server cluster based on business characteristic prediction according to claim 2, characterized in that, The expression for load burst is: divide the difference between the maximum value and the mean value by the standard deviation to get the quotient, and multiply the ratio of the number of abnormal sampling points to the total number of sampling points by the quotient to get the load burst.

4. The method for load balancing of a managed server cluster based on business characteristic prediction according to claim 1, characterized in that, The steps to construct a temporal clustering sequence are as follows: Set a sliding analysis window and extract the first-order autocorrelation coefficient of the corresponding load burst degree subsequence within the sliding analysis window; count the longest continuous time step in which the load burst degree sequence continuously exceeds the set burst threshold within the sliding analysis window; add one to the longest continuous time step and perform a logarithmic transformation, and multiply the result of the logarithmic transformation with the first-order autocorrelation coefficient to obtain the temporal clustering degree.

5. The hosting server cluster load balancing method based on business characteristic prediction according to claim 4, characterized in that, The method for calculating the burst threshold is as follows: the burst threshold is automatically determined by applying the Otsu threshold segmentation method to the historical load burst degree sequence.

6. The method for load balancing of a managed server cluster based on business characteristic prediction according to claim 1, characterized in that, The method for calculating period consistency is as follows: extract the Pearson correlation coefficient between the corresponding segments of the measured sequence and the reference sequence, and calculate the amplitude consistency penalty factor corresponding to the root mean square error; multiply the Pearson correlation coefficient and the amplitude consistency penalty factor to obtain the period consistency.

7. The hosting server cluster load balancing method based on business characteristic prediction according to claim 6, characterized in that, The amplitude consistency penalty factor is calculated as follows: extract the arithmetic mean of the corresponding segment of the reference sequence, divide the root mean square error by the arithmetic mean to obtain the normalized deviation; construct a factor with the natural constant as the base and the negative number of the normalized deviation as the exponent as the amplitude consistency penalty factor.

8. The method for load balancing of a managed server cluster based on business characteristic prediction according to claim 1, characterized in that, The steps to dynamically adjust the fusion weight parameters of historical periodic forecasts and real-time short-term forecasts based on periodic consistency to obtain the fused forecast load are as follows: The periodic consistency is used as the prediction confidence coefficient; the prediction confidence coefficient is used as the first weight of the predicted value of the historical periodic pattern, and the difference between the prediction confidence coefficient and the prediction confidence coefficient is used as the second weight of the short-term predicted value of the real-time state; the first weight is used to weight the historical periodic predicted value obtained based on the linear regression method, and the second weight is used to weight the real-time short-term predicted value obtained based on the exponential smoothing method, and the two are added together to obtain the fused prediction load.

9. The method for load balancing of a managed server cluster based on business characteristic prediction according to claim 1, characterized in that, The steps for scheduling service requests to corresponding nodes based on the converged predicted load are as follows: newly arriving service requests are routed to the node with the largest remaining predicted capacity according to the principle of minimum predicted load priority; when there is a node whose converged predicted load exceeds the first proportion of the preset rated capacity, the task is preemptively hot-migrated to a standby node with both current and predicted loads being low; when there is a node whose converged predicted load is continuously lower than the second proportion of the preset rated capacity, the node is triggered to hibernate or reduce its frequency.

10. The method for load balancing of a managed server cluster based on business characteristic prediction according to claim 1, characterized in that, After obtaining the CPU utilization time-series data of each node in the server cluster, the process also includes preprocessing the CPU utilization time-series data. The preprocessing steps are as follows: for missing sampling points, linear interpolation is used to complete them; the CPU utilization time-series data is Z-score standardized; and... The criteria detect and replace outliers, and the moving median is used to replace outliers.

Images