High-load scene-oriented computing power server system layer optimization method and system

Abstract

The invention relates to the technical field of data processing, and discloses a computing power server system layer optimization method and system oriented to a high-load scene. The method comprises the steps of collecting micro performance indexes such as the CPU instruction cycle number and the page table missing frequency, constructing a three-layer causal directed acyclic graph through Granger causal inspection, reducing a parameter search space based on bottleneck node reverse backtracking, generating an interpretable optimization decision with a causal path and contribution degree quantification, and carrying out optimization on the basis of the interpretable optimization decision. The problems that the performance bottleneck root cause cannot be accurately positioned and the optimization result lacks transparency in the prior art are solved. According to the method, bottleneck node reverse backtracking and parameter space pruning are performed based on the causal atlas, so that the problem of low optimization efficiency caused by incapability of accurately positioning a performance bottleneck root cause and blind exploration of a parameter space in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a computing power server system layer optimization method and system for high-load scenarios. BACKGROUND

[0002] With the rapid development of cloud computing, big data, artificial intelligence and other applications, the load pressure faced by computing power servers continues to rise. High-load scenarios such as high-concurrency access, large-scale data processing and intensive computing tasks pose severe challenges to the performance of server system layers. In the prior art, system layer optimization is usually performed by manually adjusting Linux kernel parameters. Operation and maintenance personnel adjust kernel parameter configurations such as vm.swappiness, vm.dirty_ratio and net.core.netdev_max_backlog, or switch I / O scheduler types and adjust NUMA node affinity, according to monitoring data and experience. These optimization methods can alleviate performance bottlenecks to some extent, but they depend on the in-depth understanding of system underlying mechanisms and long-term accumulated experience of operation and maintenance personnel. Some technical solutions use automatic scripts to periodically collect macro performance indicators such as CPU usage, memory occupancy and disk IO. When a certain indicator exceeds a preset threshold, a parameter adjustment action is triggered. For example, when CPU usage is continuously higher than 80%, the process scheduling granularity is reduced, and when memory usage is higher than 90%, the swap partition usage tendency is increased. Such a scheme realizes partial automation optimization through a threshold triggering mechanism.

[0003] The system layer optimization method in the prior art has many deficiencies. First, the manual adjustment method highly depends on the professional ability of operation and maintenance personnel, and it is difficult to fully grasp the complex system parameter combination and the mutual influence relationship between parameters, which may cause other performance problems after adjusting a certain parameter. Second, the automatic scheme based on a threshold trigger only focuses on the numerical change of a single performance indicator, and cannot identify the causal relationship between different performance indicators. For example, the increase in CPU usage may be caused by insufficient memory bandwidth, resulting in a large number of waiting memory accesses rather than computing-intensive. At this time, adjusting only the CPU-related parameters cannot solve the root problem. Third, the existing optimization method usually uses static configuration or simple rule triggering mechanism, and lacks dynamic adaptation ability to the system running state. When the business load mode changes, the original parameter configuration may be invalid or even have a negative impact. In addition, the existing scheme blindly explores the entire parameter space when adjusting parameters, without considering the causal relationship between parameters and performance bottlenecks, resulting in low optimization efficiency and difficulty in converging to the optimal configuration.

[0004] Further analysis based on the deficiencies of the prior art shows that the core difficulty of system layer optimization lies in accurately identifying the root cause of performance bottlenecks and locating related adjustable parameters. The macro performance indicators collected by the prior art, such as CPU usage and memory occupancy, only reflect the appearance of resource usage and cannot reveal the deep system behavior that causes performance bottlenecks. For example, high CPU usage may be caused by frequent context switching, cache miss, branch prediction failure, and other micro factors. Without collecting these micro indicators, it is difficult to determine the true bottleneck. Meanwhile, the prior art lacks quantitative analysis of the causal relationship between performance indicators, making it difficult to distinguish between which indicators change as a result of performance decline and which as a cause. This leads to parameter adjustment targeting result indicators rather than cause indicators. Furthermore, even if the performance bottleneck is identified, the prior art does not establish a causal mapping relationship between bottleneck indicators and adjustable parameters, and can only search for the optimal configuration by traversing or heuristically searching the entire parameter space. This blind search method is extremely inefficient in high-dimensional parameter space, and the parameter optimization results of the prior art lack explainability, making it difficult for operation and maintenance personnel to understand why adjusting a certain parameter can improve performance. This limits the credibility and maintainability of optimization solutions in production environments. SUMMARY

[0005] The present application provides a high-load scenario-oriented computing power server system layer optimization method and system, which constructs a three-layer causal directed acyclic graph based on Granger causality test, establishes a complete causal chain from kernel parameters to micro performance indicators to macro business indicators, and performs bottleneck node reverse backtracking and parameter space pruning based on the causal graph, solving the problems of the prior art that cannot accurately locate the root cause of performance bottlenecks and blindly explore the parameter space, resulting in low optimization efficiency. By generating an explainable optimization decision with a causal path and contribution quantification, the present application solves the problem of the prior art that optimization results lack transparency and credibility.

[0006] In a first aspect, the present application provides a high-load scenario-oriented computing power server system layer optimization method, which comprises:

[0007] Step S1: Collecting the CPU instruction cycle number, cache hit rate, page table miss frequency, memory access delay, and block device queue depth of the computing power server to obtain a system performance dataset;

[0008] Step S2: Performing causal relationship analysis on the system performance dataset by a Granger causality test algorithm to obtain a causal adjacency matrix, and constructing a three-layer causal directed acyclic graph based on the causal adjacency matrix;

[0009] Step S3: Locating a bottleneck node with an abnormality degree exceeding a threshold value in the three-layer causal directed acyclic graph, and performing reverse backtracking from the bottleneck node to the parameter configuration layer to obtain an effective adjustment set;

[0010] Step S4: setting the parameters not in the effective adjustment set to zero to obtain a reduced action space, performing parameter adjustment iteratively in the reduced action space to obtain a target parameter configuration vector;

[0011] Step S5: performing forward traversal through the three-layer causal directed acyclic graph from the adjustment parameters of the target parameter configuration vector to obtain a causal path sequence, and calculating the product of the causal edge weight and the parameter adjustment amplitude in the causal path sequence to obtain a causal contribution degree.

[0012] In a second aspect, the present application provides a computing power server system layer optimization system for a high-load scenario, comprising:

[0013] A collection module is configured to collect CPU instruction cycle number, cache hit rate, page table miss times, memory access delay and block device queue depth of the computing power server to obtain a system performance dataset;

[0014] An analysis module is configured to perform causal relationship analysis on the system performance dataset by using a Granger causality test algorithm to obtain a causal adjacency matrix, and construct a three-layer causal directed acyclic graph according to the causal adjacency matrix;

[0015] A backtracking module is configured to locate a bottleneck node with an abnormality degree exceeding a threshold in the three-layer causal directed acyclic graph, and perform reverse backtracking from the bottleneck node to a parameter configuration layer to obtain an effective adjustment set;

[0016] An adjustment module is configured to set the parameters not in the effective adjustment set to zero to obtain a reduced action space, and perform parameter adjustment iteratively in the reduced action space to obtain a target parameter configuration vector;

[0017] A traversal module is configured to perform forward traversal through the three-layer causal directed acyclic graph from the adjustment parameters of the target parameter configuration vector to obtain a causal path sequence, and calculate the product of the causal edge weight and the parameter adjustment amplitude in the causal path sequence to obtain a causal contribution degree.

[0018] In a third aspect, a computing power server system layer optimization device for a high-load scenario is provided, comprising a memory and at least one processor, wherein the memory stores instructions; and the at least one processor invokes the instructions in the memory to enable the computing power server system layer optimization device for a high-load scenario to perform the above-mentioned computing power server system layer optimization method for a high-load scenario.

[0019] In a fourth aspect, a computer readable storage medium is provided, having stored therein instructions which, when executed on a computer, cause the computer to perform the above-mentioned method for optimizing system layer of computing power server for high-load scenarios.

[0020] In the technical scheme provided in the application, the system performance dataset is obtained by collecting the CPU instruction cycle number, cache hit rate, page table miss times, memory access delay and block device queue depth of the computing power server. Compared with the prior art which only collects macro resource usage indicators such as CPU usage and memory occupancy, the micro performance indicators collected by the application can deeply reveal the resource competition and scheduling conflict status of the system bottom layer. The CPU instruction cycle number reflects the actual amount of effective calculation completed by the processor rather than the simple busy degree. The page table miss times directly reflect the pressure of virtual memory management. The memory access delay between NUMA nodes exposes the performance loss caused by cross-node access. The block device queue depth reveals the load status of the storage subsystem. These micro indicators are closer to the real root cause of the performance bottleneck than the macro indicators. The three-layer causal directed acyclic graph is constructed by performing causal relationship analysis on the system performance dataset by using the Granger causality test algorithm. Compared with the prior art which determines the parameter influence relationship based on correlation analysis or expert experience, the Granger causality test can strictly verify whether one time series indicator has predictive ability for another time series indicator from a statistical point of view, so as to determine whether there is a real causal relationship between the two rather than a false correlation. The three-layer causal directed acyclic graph establishes a complete causal propagation path from the kernel parameter node of the parameter configuration layer to the business indicator node of the performance indicator layer through the performance indicator node of the intermediate state layer, clearly showing how parameter adjustment affects the final performance through the causal chain. In the three-layer causal directed acyclic graph, the bottleneck node whose abnormal degree exceeds the threshold value is located, and an effective adjustment set is obtained by backtracking from the bottleneck node to the parameter configuration layer. Compared with the prior art which blindly adjusts all parameters or only triggers parameter adjustment according to a single indicator threshold value, the reverse backtracking mechanism based on the causal graph in the application can accurately trace the upstream causal source of the bottleneck node, filter out a few key parameters that have a causal relationship with the current performance bottleneck, set the dimensions of parameters not in the effective adjustment set to zero to obtain a reduced action space, greatly reduce the dimension of the parameter search space, and iteratively perform parameter adjustment in the reduced action space to avoid invalid exploration of irrelevant parameters, thereby accelerating the convergence speed of the optimization process. The adjustment parameter of the target parameter configuration vector performs forward traversal through the three-layer causal directed acyclic graph to obtain a causal path sequence and calculate a causal contribution degree. Compared with the black box optimization result of the prior art which only gives parameter configuration values and cannot explain the adjustment reason, the application generates a complete causal propagation path for each adjustment parameter, shows how the parameter finally affects the business indicator of the performance indicator layer through multiple links of the intermediate state layer, calculates the product of the causal edge weight and the parameter adjustment amplitude in the causal path sequence to obtain the causal contribution degree, quantifies the actual contribution size of each parameter adjustment to performance improvement, and enables operation and maintenance personnel to clearly understand the causal logic and importance ranking of parameter adjustment, thereby enhancing the transparency and credibility of optimization decisions, and facilitating the review and maintenance of optimization schemes in a production environment.

[0021] The application of the Granger causality test algorithm in the present application makes a key contribution to the technical solution. The algorithm compares the prediction accuracy of an autoregressive model using only the historical data of the explained index with that of an extended autoregressive model using the historical data of both the explained index and the explanation index, to determine whether the explanation index contains useful information for predicting the explained index. When the residual sum of squares of the extended model is significantly smaller than that of the autoregressive model, it indicates that the explanation index has a causal effect on the explained index. This causal determination method can effectively distinguish between real causal relationships and false correlations compared to simple correlation coefficient calculation, avoiding the misjudgment of two indexes as having a causal relationship as a result of a certain potential factor. In the performance server system layer optimization scenario, there are complex mutual influences and common influences of upstream parameters between performance indicators. The Granger causality test determines the causal relationship through the prediction ability of time series and uses the time sequence characteristic that the cause must occur before the result to accurately identify which changes in performance indicators are the real causes of changes in other indicators. The causal adjacency matrix laid a foundation for the establishment of a three-layer causal directed acyclic graph. The bottleneck node backtracking and parameter space pruning based on the causal graph directly rely on the accurate causal relationship information provided by the Granger causality test. If correlation analysis is used instead of causality test, the highly correlated but non-causal index pair may be incorrectly connected, causing the backtracking to incorrect upstream parameters and deviating the parameter adjustment direction from the real bottleneck root cause. The introduction of the Granger causality test algorithm enables the present application to accurately locate the causal source of the performance bottleneck in a high-load scenario, and adjust the relevant parameters accordingly, avoiding the invalid optimization and even negative impact caused by misjudgment of causal relationship in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Figure 1 An embodiment schematic diagram of the algorithm power server system layer optimization method in the embodiment of the present application facing the high-load scenario;

[0024] Figure 2 An embodiment schematic diagram of the algorithm power server system layer optimization system in the embodiment of the present application facing the high-load scenario;

[0025] Figure 3 An embodiment schematic diagram of the algorithm power server system layer optimization system in the embodiment of the present application facing the high-load scenario; DETAILED DESCRIPTION

[0026] The embodiment of the present application provides a computing power server system layer optimization method and system for a high-load scene. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0027] For ease of understanding, the specific process of the embodiment of the present application is described below. Please refer to Figure 1 One embodiment of the computing power server system layer optimization method for a high-load scene in the embodiment of the present application includes the following steps.

[0028] Step S1: Collecting CPU instruction cycle number, cache hit rate, page table miss times, memory access delay and block device queue depth of the computing power server to obtain a system performance data set;

[0029] Step S2: Performing causal relationship analysis on the system performance data set by using a Granger causality test algorithm to obtain a causal adjacency matrix, and constructing a three-layer causal directed acyclic graph according to the causal adjacency matrix;

[0030] Step S3: Locating a bottleneck node with an abnormality degree exceeding a threshold value in the three-layer causal directed acyclic graph, and backtracking from the bottleneck node to a parameter configuration layer to obtain an effective adjustment set;

[0031] Step S4: Setting zero to a parameter dimension not in the effective adjustment set to obtain a reduced action space, and iteratively performing parameter adjustment in the reduced action space to obtain a target parameter configuration vector;

[0032] Step S5: Performing forward traversal through the three-layer causal directed acyclic graph from an adjustment parameter of the target parameter configuration vector to obtain a causal path sequence, and calculating a product of a causal edge weight and a parameter adjustment amplitude in the causal path sequence to obtain a causal contribution degree.

[0033] It can be understood that the execution subject of the present application can be a computing power server system layer optimization system for a high-load scene, and can also be a terminal or a server, which is not limited here. The embodiment of the present application takes a server as an execution subject for example.

[0034] Specifically, by deploying a performance monitoring module in the operating system kernel layer of the computing power server, system running data is continuously read at a sampling period of 100 milliseconds. Specifically, the number of CPU instruction cycles is obtained by reading the instruction counter value in the / proc / stat file, which reflects the efficiency of the processor executing instructions. The higher the value, the more instructions are completed per unit time. At the same time, the number of page table misses is obtained by reading the page_fault statistical item in the / proc / meminfo file, which records the number of times that the virtual memory address conversion fails, resulting in the need to load pages from the disk. The numa_hit and numa_miss count values of each NUMA node under the / sys / devices / system / node directory are read, and the memory access delay is obtained by calculating the ratio of the difference between the two values to the total access amount. The larger the ratio, the more frequent the cross-node access, resulting in increased delay. The ninth column value of the / sys / block / sda / stat file is read to obtain the block device queue depth, which represents the number of IO requests waiting to be processed. After arranging the above collected raw values according to the timestamp to form a sampling sequence, the arithmetic mean value in the 5-second window is calculated by dividing the sum of all sampling point values by the number of sampling points. The variance value is calculated by dividing the square sum of the difference between each sampling point and the mean value by the number of sampling points. The change rate is calculated by dividing the difference between the end value of the window and the starting value of the window by the starting value. The same statistical calculation is performed on the 30-second window and the 300-second window. The nine statistical values (three means, three variances, and three change rates) of each performance indicator in the three time windows are arranged in a row, and the statistical values of all performance indicators are stacked in rows to form a feature matrix. Each column of the feature matrix is standardized by calculating the mean and standard deviation of all values in the column. Each original value is subtracted from the mean and divided by the standard deviation to obtain the standardized value. The mean of the standardized feature matrix is zero and the standard deviation is one, eliminating the influence of the dimension difference of different performance indicators. Finally, a system performance dataset is formed.

[0035] The time series data of any two performance indicators in the system performance dataset is selected to form an indicator pair, the first time series is set as the explained indicator marked as Y sequence, and the second time series is set as the explaining indicator marked as X sequence, two regression models are constructed for the indicator pair, the first autoregressive model only uses the historical data of the Y sequence itself to predict the current value, specifically, the value of the Y sequence at time t is expressed as a linear combination of the values of the Y sequence at ten historical time points t minus 1, t minus 2, and so on to t minus 10, each historical value is multiplied by the corresponding regression coefficient and then summed, the regression coefficients are fitted by the least squares method, and the residual sum of squares between the predicted value and the true value of the model is denoted as RSS1, the second extended autoregressive model adds the historical data of the X sequence as an additional prediction variable on the basis of the autoregressive model, the value of the Y sequence at time t is expressed as a linear combination of the ten historical values of the Y sequence plus a linear combination of the ten historical values of the X sequence, all regression coefficients are fitted by the least squares method, and the residual sum of squares is denoted as RSS2, the F statistic is equal to the difference between RSS1 and RSS2 divided by RSS2, multiplied by the sample size minus 20 and then divided by 10, the F statistic obeys the F distribution with degrees of freedom of 10 and sample size minus 20, the critical value at a significance level of 0.05 is looked up in the F distribution critical value table, if the calculated F statistic is greater than the critical value, it is determined that there is a Granger causality relationship between the X sequence and the Y sequence, the element at the position of the row index corresponding to the X sequence and the column index corresponding to the Y sequence in the causal adjacency matrix is assigned a value of 1 to represent the causal relationship, otherwise, it is assigned a value of 0 to represent the absence of causal relationship, all possible indicator pair combinations in the system performance dataset are traversed to complete the filling of the entire causal adjacency matrix, a three-layer causal directed acyclic graph is constructed based on the causal adjacency matrix, the first layer of parameter configuration layer includes 48 adjustable Linux kernel parameter nodes such as vm.swappiness, vm.dirty_ratio and net.core.netdev_max_backlog, the second layer of intermediate state layer includes 120 performance indicator nodes such as CPU instruction cycle number and page table miss number in the system performance dataset, and the third layer of performance indicator layer includes two business concerned indicator nodes of system throughput TPS and average response delay, a directed edge is drawn between the corresponding nodes according to the position of the element with a value of 1 in the causal adjacency matrix, the starting point of the edge is the node corresponding to the explaining indicator, the terminal point of the edge is the node corresponding to the explained indicator, and the weight of the edge is set as the value of the F statistic normalized to the interval of 0 to 1 after subtracting the critical value, which reflects the strength of the causal influence.

[0036] The real-time monitoring values of each performance indicator node are extracted from the middle state layer of the three-layer causal directed acyclic graph, and the mean and standard deviation of each performance indicator in the past 24 hours are calculated from the historical running data. For the CPU instruction cycle number node, assuming that the current real-time monitoring value is 15000, the historical mean is 12000, and the historical standard deviation is 800, the abnormality degree of this node is calculated as 15000 minus 12000 divided by 800, which is 3.75. The same abnormality degree calculation is performed on all 120 performance indicator nodes in the middle state layer, and the nodes with an abnormality degree greater than 2 are selected as the bottleneck node set. Assuming that three bottleneck nodes, page table miss count, memory access delay, and block device queue depth, are selected, starting from the page table miss count node, find all directed edges pointing to this node in the three-layer causal directed acyclic graph, identify the starting nodes of these edges and record them in the reverse search path, continue to backtrack from these starting nodes to find their upstream nodes, and repeat the process until the kernel parameter nodes of the parameter configuration layer are reached. Assuming that the upstream node of the page table miss count node is found to be a memory management related performance indicator by backtracking, continue to backtrack to find the vm.swappiness parameter node, and add this parameter node to the effective adjustment set. The same reverse breadth-first search process is performed on the memory access delay and block device queue depth bottleneck nodes. Assuming that the memory access delay node backtracks to the vm.dirty_ratio and kernel.numa_balancing parameter nodes, and the block device queue depth node backtracks to the block device I / O scheduler type parameter node, the parameter nodes obtained by backtracking are de-duplicated to obtain an effective adjustment set containing the vm.swappiness, vm.dirty_ratio, kernel.numa_balancing, and I / O scheduler type parameter nodes. The number of parameters in the effective adjustment set is 4, and the total number of parameter nodes in the parameter configuration layer is 48. The parameter space reduction ratio is calculated as 1 minus 4 divided by 48, which is approximately 0.92, indicating that the parameter space to be adjusted is reduced from 48 dimensions to 4 dimensions through causal backtracking.

[0037] The initial action space is constructed as a 48-dimensional vector, each dimension of the vector corresponding to a kernel parameter node of the parameter configuration layer. Each dimension of the vector is traversed to check if the corresponding parameter node is in the valid adjustment set. For parameter dimensions in the valid adjustment set, their adjustable state is maintained. For parameter dimensions not in the valid adjustment set, their adjustment amplitude is forcibly assigned a value of zero. Assuming the valid adjustment set contains parameter nodes at positions 3, 15, 27, and 40, only the values at positions 3, 15, 27, and 40 in the 48-dimensional action vector are non-zero, and the remaining 44 positions are zero. After this operation, the reduced action space is obtained. The real-time values of the current 120 performance indicators are read from the middle state layer of the three-layer causal directed acyclic graph to form a system state vector. The 120-dimensional state vector is input into the policy network. The policy network is composed of three fully connected neural network layers. The input layer has 120 neurons to receive the values of each dimension of the state vector. The first hidden layer has 256 neurons, the second hidden layer has 256 neurons, and the third hidden layer has 128 neurons. The output layer has 48 neurons corresponding to the adjustment amplitudes of the 48 parameters. After the policy network outputs a 48-dimensional vector, the vector is multiplied element-wise with the mask vector of the reduced action space, so that the output values of parameter dimensions not in the valid adjustment set are set to zero, obtaining the parameter adjustment action of the current iteration. Assuming that the output value at position 3 is positive 0.15, it means that the corresponding parameter will be increased by 15% of the current value. According to the parameter adjustment action, the system parameter modification operation is performed. The vm.swappiness parameter is adjusted from the current value 60 to 69 through the sysctl command. After waiting for 60 seconds for the system to reach steady state, the performance indicators are re-collected. Assuming that the system throughput before adjustment is 8500 TPS and the average delay is 120 milliseconds, the system throughput after adjustment becomes 9200 TPS and the average delay becomes 105 milliseconds. The immediate reward value is calculated as the throughput change value 700 divided by the baseline value 8500 multiplied by the weight 0.5 plus the delay change value negative 15 divided by the baseline value 120 multiplied by the weight 0.5. The immediate reward value is approximately equal to 0.0411 plus negative 0.0625, which is equal to negative 0.0214. The system state vector before adjustment, the parameter adjustment action vector, the immediate reward scalar value, and the system state vector after adjustment are packaged into an experience tuple and stored in the experience replay buffer. When the buffer accumulates more than 1000 experiences, 128 experiences are randomly selected to form a training batch. The gradients of the policy network and the value network are calculated using the batch data to update the network parameters. The iteration cycle of parameter adjustment, reward calculation, experience storage, and network training is repeatedly executed. After each iteration, the average reward value of the current 100 iterations is calculated. When the absolute value of the average reward change rate of 20 consecutive iterations is less than 0.01, convergence is determined. At this time, the parameter adjustment vector output by the policy network is the target parameter configuration vector.

[0038] The 48 dimensions of the target parameter configuration vector are checked to identify the parameter dimensions with non-zero adjustment amplitudes. Assuming that the adjustment amplitudes of the 3rd, 15th, and 27th dimensions are 0.15, -0.08, and 0.12 respectively, and the adjustment amplitudes of the remaining dimensions are all zero, the parameter nodes corresponding to these three dimensions are extracted as adjustment parameter nodes. Starting from the vm.swappiness parameter node corresponding to the 3rd dimension, all directed edges leading from this node are searched in the three-layer causal directed acyclic graph. Assuming that the node is connected to the page table miss count node of the intermediate state layer, the directed edges are further traversed forward from the page table miss count node to the memory access latency node, and then from the memory access latency node to the average response latency node of the performance indicator layer. The complete path is recorded as vm.swappiness node to page table miss count node to memory access latency node to average response latency node. This path includes three directed edges, and the weight values stored in the causal graph for these three edges are assumed to be 0.42, 0.35, and 0.28. The three weight values are arranged in the order of the path to form an edge weight sequence, and a path weight vector is constructed as a vector containing 0.42, 0.35, and 0.28. The path causal strength is calculated as 0.42 multiplied by 0.35 multiplied by 0.28, which is approximately equal to 0.041. The adjustment amplitude of the vm.swappiness parameter node is extracted as 0.15, and the causal contribution degree of this parameter node is calculated as 0.041 multiplied by 0.15, which is approximately equal to 0.00615. The same forward traversal and calculation process is performed for the adjustment parameter nodes of the 15th and 27th dimensions. Assuming that the path causal strength of the 15th parameter node is 0.038 and the adjustment amplitude is -0.08, the causal contribution degree is -0.00304. The path causal strength of the 27th parameter node is 0.052 and the adjustment amplitude is 0.12, resulting in a causal contribution degree of 0.00624. The causal contribution degrees of the three adjustment parameter nodes are arranged in descending order of absolute value to obtain a parameter importance list. The 27th parameter is ranked first, indicating that it has the greatest contribution to performance improvement. The 3rd parameter is ranked second, and the 15th parameter is ranked third. The importance list, combined with the complete paths of each parameter node in the causal graph, constitutes an interpretable basis for system layer optimization decisions.

[0039] In an embodiment, step S1 comprises:

[0040] The performance monitoring module deployed in the operating system kernel layer synchronously reads system files to obtain CPU instruction cycle count, page table miss count, memory access latency, and block device queue depth at a preset sampling period to obtain raw performance indicator sampling data.

[0041] The original performance index sampling data is respectively subjected to sliding statistics according to a short time window, a medium time window and a long time window, and the arithmetic mean value, the variance value and the change rate of each performance index in each time window are calculated to obtain multi-time window statistical characteristics;

[0042] The statistical values of each performance index in the multi-time window statistical characteristics in the three time windows are spliced by rows to construct a feature matrix to obtain the feature matrix;

[0043] Each dimension characteristic value in the feature matrix is subjected to standardization processing to obtain a system performance data set.

[0044] Specifically, the performance monitoring module is executed once every 100 milliseconds as a kernel state daemon, opens the / proc / stat file to extract the sum of user state time and system state time as effective instruction execution time, divides the increment of the time between two samplings by the sampling interval to obtain the CPU instruction cycle number, opens the / proc / meminfo file to find the pgfault line to extract the page table miss number, traverses the numastat file of each NUMA node under the / sys / devices / system / node directory to extract the numa_foreign and numa_hit values to calculate the cross-node access ratio as a memory access delay index, opens the stat file of the block device under the / sys / block directory to extract the 9th field value as the block device queue depth, and packs the four values obtained at each sampling together with the timestamp into a record to form a sequence of original performance index sampling data.

[0045] Three sliding time windows are established for each performance indicator in the original performance indicator sampling data, the short time window is set to 5 seconds containing 50 sampling points, the medium time window is set to 30 seconds containing 300 sampling points, and the long time window is set to 300 seconds containing 3000 sampling points. At the current time, the sampling points in the corresponding time range are extracted from the original data. The arithmetic mean value is obtained by accumulating and summing these values and then dividing by the number of sampling points. The variance value is obtained by subtracting the mean value from each sampling point value, squaring the sum, and then dividing by the number of sampling points. The change rate is obtained by subtracting the starting time value from the ending time value and then dividing by the starting value. Each performance indicator generates three statistical values, including mean, variance, and change rate, in three time windows, a total of thirty-six statistical characteristic values, forming a multi-time window statistical characteristic. The nine statistical values of each performance indicator are arranged in the order of short window mean, short window variance, short window change rate, medium window mean, medium window variance, medium window change rate, long window mean, long window variance, and long window change rate, forming a four-row nine-column feature matrix. The mean and standard deviation of each column of the feature matrix are calculated. The normalized value is obtained by subtracting the mean value of each original value in the column and then dividing by the standard deviation of the column. After normalization, the mean value of each column is zero and the standard deviation is one, eliminating the dimension difference of different performance indicators to form a system performance data set.

[0046] In a specific embodiment, step S2 includes:

[0047] From the system performance data set, any two time series indicators are extracted as an indicator pair. The first time series indicator in the indicator pair is defined as the explained indicator, and the second time series indicator in the indicator pair is defined as the explanatory indicator. An autoregressive model and an extended autoregressive model are constructed, and the residual sum of squares is calculated.

[0048] According to the residual sum of squares, the F statistic is calculated. When the F statistic is greater than a preset critical value, the corresponding element in the causal adjacency matrix is set to 1. All indicator pairs are traversed to obtain the causal adjacency matrix.

[0049] According to the causal adjacency matrix, a parameter configuration layer, an intermediate state layer, and a performance indicator layer are constructed. The parameter configuration layer includes a kernel parameter node, the intermediate state layer includes a performance indicator node, and the performance indicator layer includes a throughput node and a response delay node.

[0050] According to the position of the element with a value of 1 in the causal adjacency matrix, a directed edge is established between the corresponding nodes to obtain a three-layer causal directed acyclic graph.

[0051] Specifically, time series data of any two performance indicators are selected from the system performance dataset to form an indicator pair. The first time series is designated as the explained indicator, denoted as the Y series, and the second time series is designated as the explained indicator, denoted as the X series. Two regression models are constructed for this indicator pair to test causality. The first autoregressive model uses only the historical data of the Y series itself to predict the current value. The value of the Y series at time t is represented as a linear combination of the values ​​of the Y series at ten historical times from t-1, t-2 to t-10. Each historical value is multiplied by the corresponding regression coefficient and summed to obtain the predicted value. The combination of regression coefficients that minimizes the sum of squared errors between the predicted and actual values ​​is calculated using the least squares method. The residual sum of squares of this model is denoted as RSS1. The second extended autoregressive model adds historical data of the X series as an additional factor to the autoregressive model. The predictor variable is represented by the value of the Y sequence at time t as a linear combination of ten historical values ​​of the Y sequence plus a linear combination of ten historical values ​​of the X sequence. All regression coefficients are obtained by fitting using the least squares method, and the sum of squared residuals (RSS2) is calculated. If the sum of squared residuals significantly decreases after expanding the model to include X sequence information, it indicates that the X sequence contains useful information for predicting the Y sequence. The F-statistic is calculated as the difference between RSS1 and RSS2, divided by RSS2, multiplied by the sample size minus 20, and then divided by 10. This F-statistic follows an F-distribution with 10 degrees of freedom and a sample size minus 20. Under the null hypothesis that the X sequence does not affect the Y sequence, the F-statistic should be small. The critical value at a significance level of 0.05 is consulted in the F-distribution critical value table. If the calculated F-statistic is greater than this critical value, the null hypothesis is rejected, and a Granger causal relationship exists between the X sequence and the Y sequence.

[0052] In the causal adjacency matrix, the elements in the position of the row index corresponding to the X sequence and the column index corresponding to the Y sequence are assigned a value of 1 to represent the existence of a causal relationship, otherwise a value of 0 is assigned to represent the non-existence of a causal relationship. The above Granger causality test process is repeated for all possible combinations of index pairs in the system performance data set to complete the filling of the entire causal adjacency matrix. The number of rows and columns of the causal adjacency matrix is equal to the number of performance indicators. The element in the ith row and jth column of the matrix is 1, indicating that the ith indicator has a causal effect on the jth indicator. Based on the causal adjacency matrix, a three-layer causal directed acyclic graph is constructed. The first layer is the parameter configuration layer, which includes 48 adjustable Linux kernel parameter nodes such as vm.swappiness, which controls the tendency of swap partition usage, vm.dirty_ratio, which controls the dirty page flushing threshold, and net.core.netdev_max_backlog, which controls the network device backlog queue length. The second layer is the intermediate state layer, which includes performance indicator nodes such as CPU instruction cycle count, page table miss count, memory access delay, and block device queue depth in the system performance data set. The third layer is the performance indicator layer, which includes two business focus indicator nodes, system throughput TPS, which represents the number of transactions per second, and average response delay, which represents the average time from request initiation to completion.

[0053] According to the position of the element with a value of 1 in the causal adjacency matrix, a directed edge is established between the corresponding nodes. The starting point of the edge is the node corresponding to the explained indicator, and the end point of the edge is the node corresponding to the explained indicator. The direction of the edge represents the propagation direction of the causal effect. The weight of the edge is set to the value obtained by normalizing the F statistic minus the critical value by the maximum value of all significant F statistics to obtain a weight value in the interval [0, 1]. This weight value reflects the strength of the causal effect. The larger the F statistic, the greater the predictive contribution of the explained indicator to the explained indicator, and the stronger the causal effect. By establishing a weighted directed edge between the three layers of nodes, a three-layer causal directed acyclic graph is formed. This graph structure clearly shows how the kernel parameter nodes in the parameter configuration layer affect the performance indicator nodes in the intermediate state layer through causal chains, and further affect the business indicator nodes in the performance indicator layer. The existence of a directed path from any node to another node in the graph indicates that the former has a direct or indirect causal effect on the latter.

[0054] In a specific embodiment, step S3 includes:

[0055] From the intermediate state layer of the three-layer causal directed acyclic graph, the current values of each performance indicator node are extracted. The abnormality of each performance indicator node is calculated as the current value minus the historical mean divided by the historical standard deviation. The performance indicator nodes with an abnormality exceeding a pre-set threshold are selected as bottleneck nodes.

[0056] The reverse breadth-first search is performed along the directed edges of the three-layer causal directed acyclic graph from the bottleneck node, all nodes passed in the search path are recorded, the search is stopped when the parameter configuration layer is reached, and a reverse search path set is obtained;

[0057] All kernel parameter nodes located in the parameter configuration layer are extracted from the reverse search path set, and the identifiers of the kernel parameter nodes are stored in a set container to obtain an effective adjustment set.

[0058] The number of kernel parameter nodes in the effective adjustment set is counted, and the number of kernel parameter nodes is compared with the total number of parameter nodes in the parameter configuration layer to calculate the parameter space reduction ratio, and the parameter space reduction ratio is obtained.

[0059] Specifically, the current real-time monitoring values of each performance indicator node are extracted from the middle state layer of the three-layer causal directed acyclic graph, and the mean and standard deviation of each performance indicator in the past 24 hours are calculated from historical running data. For the CPU instruction cycle number node, the current real-time monitoring value is assumed to be 15000, the historical mean value of all sampling points of the indicator in the past 24 hours is obtained from the historical database, the sum of all historical values is divided by the number of sampling points to obtain the historical mean value of 12000, the square of the difference between each historical value and the mean value is summed and then divided by the number of sampling points to obtain the historical standard deviation of 800, and the abnormality degree of the node is calculated as the current value 15000 minus the historical mean value 12000 to obtain 3000, which is divided by the historical standard deviation 800 to obtain an abnormality degree of 3.75. The abnormality degree value reflects the degree of deviation of the current state from the normal level. The abnormality degree is positive, indicating that the current value is higher than the historical normal level. The abnormality degree is negative, indicating that the current value is lower than the historical normal level. The greater the absolute value of the abnormality degree, the more serious the deviation. The same abnormality degree calculation is performed on all performance indicator nodes in the middle state layer. A preset threshold of 2 is set, which means that a deviation of more than two standard deviations is considered abnormal. The performance indicator nodes with an absolute abnormality degree greater than 2 are selected as the bottleneck node set. It is assumed that the page table miss count node has an abnormality degree of 2.8, the memory access delay node has an abnormality degree of 3.2, and the block device queue depth node has an abnormality degree of 2.5.

[0060] From the page table miss count node, all directed edges pointing to the node are searched in the three-layer causal directed acyclic graph, the direction of the directed edge from the starting point to the terminal point indicates the causal influence direction, the reverse search, that is, backtracking from the terminal point to the starting point in the opposite direction of the edge, identifies all starting point nodes pointing to the page table miss count node and records them into the reverse search path, and the upstream nodes of these starting point nodes are searched upwards, and the process is repeated until the kernel parameter nodes of the parameter configuration layer are reached. The parameter configuration layer nodes have no upstream nodes and thus serve as the termination condition for backtracking. Assuming that the upstream node of the page table miss count node is found to be a certain intermediate state indicator node related to memory management through backtracking, the vm.swappiness parameter node is found by continuing to backtrack upwards, and the parameter node is added to the reverse search path set. The same reverse breadth-first search process is performed on the memory access delay node and the block device queue depth node. The breadth-first search refers to traversing all nodes in the current layer before entering the previous layer during the backtracking process. Assuming that the memory access delay node backtracks to the vm.dirty_ratio and kernel.numa_balancing parameter nodes, and the block device queue depth node backtracks to the block device I / O scheduler type parameter node, all backtracked parameter nodes are collected into the reverse search path set.

[0061] All kernel parameter nodes located in the parameter configuration layer are extracted from the reverse search path set. Each node in the reverse search path set is traversed, and the level to which the node belongs is determined. If the node belongs to the parameter configuration layer, the identifier of the node is extracted. The identifier is a parameter name string such as vm.swappiness, vm.dirty_ratio, etc. These parameter name strings are stored in a set container. The set container automatically removes duplicates to ensure that each parameter node appears only once. Assuming that the reverse search path set contains the vm.swappiness, vm.dirty_ratio, kernel.numa_balancing, and I / O scheduler type parameter nodes, the four parameter names are stored in the set container to obtain the effective adjustment set. The number of elements in the effective adjustment set is 4, indicating that the number of kernel parameters that need to be adjusted is 4. The parameter configuration layer contains a total of 48 adjustable kernel parameter nodes. The parameter space reduction ratio is equal to 1 minus the number of effective adjustment sets divided by the total number of parameters, that is, 1 minus 4 divided by 48, which is approximately equal to 0.917. The parameter space reduction ratio reflects that the parameter space that needs to be adjusted is reduced from 48 dimensions to 4 dimensions through causal backtracking, and only a few key parameters that have a causal relationship with the current bottleneck node need to be focused on.

[0062] In a specific embodiment, step S4 comprises:

[0063] The initial action space is constructed to contain all kernel parameter nodes of the parameter configuration layer, each parameter dimension in the initial action space is traversed, and it is determined whether the kernel parameter node corresponding to each parameter dimension is in the effective adjustment set. The adjustment amplitude of the parameter dimension not in the effective adjustment set is assigned as zero to obtain a reduced action space;

[0064] The current system state vector is read from the middle state layer of the three-layer causal directed acyclic graph, and the system state vector is input into the policy network to generate a parameter adjustment action. Each dimension of the parameter adjustment action is valued within the non-zero dimension range of the reduced action space to obtain a parameter adjustment action of the current iteration;

[0065] The corresponding kernel parameter value is modified according to the parameter adjustment action, and a new system state vector is collected after the system enters a steady state. The weighted sum of the throughput change value and the delay change value is calculated as an immediate reward value. The system state vector, the parameter adjustment action and the immediate reward value are stored in an experience replay buffer;

[0066] A batch of data is randomly sampled from the experience replay buffer to train the policy network and the value network. When the average reward change rate of consecutive multiple iterations is less than a preset convergence threshold, the iteration is terminated. The parameter adjustment vector output by the policy network is extracted as a target parameter configuration vector.

[0067] Specifically, the initial action space is a vector containing 48 elements. Each element of the vector corresponds to a kernel parameter node of the parameter configuration layer. The numerical value of the element represents the adjustment amplitude of the parameter. Each element position of the vector is traversed to check whether the kernel parameter node corresponding to the position is in the effective adjustment set. Assuming that the effective adjustment set contains vm.swappiness, vm.dirty_ratio, kernel.numa_balancing, and I / O scheduler type four parameter nodes, the element values of the fourth positions of the third, fifteenth, twenty-seventh, and fortieth positions of the vector are maintained in an adjustable state. For the remaining 44 parameter nodes not in the effective adjustment set, the element values of the vector positions corresponding to these parameter nodes are forcibly assigned as zero to indicate that these parameters are not adjusted. After the operation, the reduced action space is obtained. The vector of the reduced action space is still 48-dimensional, but only four dimensions have non-zero values and the remaining 44 dimensions are zero. The reduced action space limits parameter adjustment to only a few key parameters that have a causal relationship with the current bottleneck node.

[0068] The current real-time values of all performance indicator nodes in the middle state layer of the three-layer causal directed acyclic graph are read to form a system state vector, assuming that the middle state layer contains 120 performance indicator nodes, the system state vector is a 120-dimensional vector, which is input into the policy network. The policy network is composed of multiple layers of fully connected neural networks. The input layer receives a 120-dimensional state vector, and after nonlinear transformation through multiple hidden layers, the output layer produces a 48-dimensional parameter adjustment action vector. The numerical value of each dimension of the parameter adjustment action vector output by the policy network represents the recommended adjustment amplitude of the corresponding kernel parameter. The 48-dimensional vector output by the policy network is multiplied element by element with the mask vector that reduces the action space. The mask vector is 1 at the dimension position corresponding to the valid adjustment set and 0 at the remaining positions. After multiplication, the output value of the parameter dimension not in the valid adjustment set is forced to zero, obtaining the parameter adjustment action vector of the current iteration. Assuming that the 3rd dimension output value is positive 0.15, indicating that the vm.swappiness parameter is increased by 15% of the current value, the 15th dimension output value is negative 0.08, indicating that the vm.dirty_ratio parameter is reduced by 8% of the current value, and the 27th dimension output value is positive 0.12, indicating that the kernel.numa_balancing parameter is increased by 12% of the current value.

[0069] According to the parameter adjustment action vector, the system parameter modification operation is performed. For the vm.swappiness parameter, the current value is 60, and after increasing by 15%, it is adjusted to 69. The parameter value is written into the kernel through the system command sysctl. The same modification operation is performed on other non-zero dimension parameters. After the parameter modification is completed, 60 seconds are waited for the system to reach a steady state. The steady state refers to a state in which the system performance indicators no longer fluctuate sharply and reach a relatively stable running state under the new parameter configuration. New system state vectors are collected from the middle state layer. Assuming that the system throughput before adjustment is 8500 transactions per second, and the average delay is 120 milliseconds, the system throughput after adjustment is 9200 transactions per second, and the average delay is 105 milliseconds. The throughput change value is 9200 minus 8500, which is 700. The delay change value is 105 minus 120, which is negative 15. The throughput change value is divided by the baseline value 8500, multiplied by the weight coefficient 0.5. The delay change value is divided by the baseline value 120, multiplied by the weight coefficient 0.5. The sum of the two is the immediate reward value. The increase in throughput produces a positive reward, and the decrease in delay also produces a positive reward. The system state vector before adjustment, the parameter adjustment action vector, the immediate reward value, and the system state vector after adjustment are packaged into an experience tuple and stored in the experience replay buffer. The experience replay buffer uses a first-in-first-out queue structure. When the buffer capacity reaches the upper limit, the oldest experience tuple is overwritten by the new experience tuple.

[0070] Randomly sample 128 experience tuples from the experience replay buffer to form a training batch, use the batch data to calculate the loss function of the policy network and the value network, the goal of the policy network is to maximize the cumulative reward, the goal of the value network is to accurately estimate the state value, calculate the gradient of the loss function to the network parameters through the back propagation algorithm, update the network parameters using the gradient descent method, repeat the iteration cycle of parameter adjustment, reward calculation, experience storage and network training, record the current instantaneous reward value after each iteration, calculate the average reward value of the last 100 iterations, when the absolute value of the average reward change rate of 20 consecutive iterations is less than 0.01, it is determined that the optimization process converges, at this time the policy network has learned the mapping relationship from the system state to the optimal parameter adjustment, extract the parameter adjustment vector output by the policy network as the target parameter configuration vector, the vector contains the optimal adjustment amplitude of each kernel parameter, apply the vector to the actual production environment to complete the system layer optimization.

[0071] In a specific embodiment, step S5 comprises:

[0072] Extract the parameter nodes with non-zero adjustment amplitudes from the target parameter configuration vector as adjustment parameter nodes, perform forward breadth-first traversal in the three-layer causal directed acyclic graph starting from each adjustment parameter node, record the intermediate state layer nodes passed through and the performance indicator layer nodes reached during the traversal process, and obtain a causal path sequence;

[0073] Extract the weight values of each directed edge in the causal path sequence, and perform a multiplication operation on the weight values of all directed edges in the same path to obtain a path causal strength;

[0074] Extract the parameter adjustment amplitudes corresponding to each adjustment parameter node in the target parameter configuration vector, multiply the path causal strength and the corresponding parameter adjustment amplitude to obtain a causal contribution degree;

[0075] Sort the causal contribution degrees of all adjustment parameter nodes in descending order according to the numerical value to generate a parameter importance list as an interpretable basis for system layer optimization decision-making.

[0076] Specifically, the adjustment amplitude values of all 48 parameter dimensions are checked from the target parameter configuration vector, the parameter dimensions with non-zero adjustment amplitudes are identified, it is assumed that the adjustment amplitude of the 3rd dimension corresponding to the vm.swappiness parameter is positive 0.15, the adjustment amplitude of the 15th dimension corresponding to the vm.dirty_ratio parameter is negative 0.08, the adjustment amplitude of the 27th dimension corresponding to the kernel.numa_balancing parameter is positive 0.12, and the adjustment amplitudes of the remaining 45 dimensions are all zero, the parameter nodes corresponding to the three parameter dimensions with non-zero adjustment amplitudes are extracted as adjustment parameter nodes, the forward breadth-first traversal is performed in the three-layer causal directed acyclic graph starting from the vm.swappiness parameter node, the forward traversal refers to walking from the starting point to the terminal point along the positive direction of the directed edge, all directed edges leading from the parameter node are searched, it is assumed that the parameter node is connected to the page table miss frequency node of the intermediate state layer, the page table miss frequency node is continuously traversed forward along the directed edge to find the downstream node, it is assumed that the downstream node is the memory access latency node, the average response delay node of the performance indicator layer is further traversed forward from the memory access latency node, and the complete traversal path vm.swappiness node-page table miss frequency node-memory access latency node-average response delay node is recorded. The path includes three directed edges connecting four nodes, the same forward breadth-first traversal process is performed on the vm.dirty_ratio parameter node and the kernel.numa_balancing parameter node, it is assumed that the traversal path of the vm.dirty_ratio node is the parameter node-disk write frequency node-block device queue depth node-system throughput node, and the traversal path of the kernel.numa_balancing node is the parameter node-cross-node memory access proportion node-memory bandwidth occupancy rate node-average response delay node, and the traversal paths of all adjustment parameter nodes are collected to obtain a causal path sequence.

[0077] The directed edges contained in each path in the sequence of causal paths are extracted, and the weight value stored in the three-layer causal directed acyclic graph for each directed edge is read. The weight value is calculated according to the F statistic of Granger causality test when the causal graph is constructed, and reflects the strength of causal influence. Assuming that the path of the vm.swappiness node contains three edges, the weight value of the first edge from the vm.swappiness node to the page table miss count node is 0.42, the weight value of the second edge from the page table miss count node to the memory access latency node is 0.35, and the weight value of the third edge from the memory access latency node to the average response latency node is 0.28. The three weight values are multiplied in sequence according to the path to obtain the path causal strength of the path, which is 0.041. The path causal strength reflects the comprehensive influence degree of the parameter node on the performance indicator node through the causal path. The longer the path and the more intermediate nodes, the smaller the path causal strength after multiplication. Assuming that the path causal strength of the vm.dirty_ratio node is 0.038, and the path causal strength of the kernel.numa_balancing node is 0.052.

[0078] The parameter adjustment amplitude value corresponding to each adjustment parameter node in the target parameter configuration vector is extracted, the adjustment amplitude of the vm.swappiness node is 0.15, the adjustment amplitude of the vm.dirty_ratio node is -0.08, and the adjustment amplitude of the kernel.numa_balancing node is 0.12, the path causal strength of each node is multiplied by the corresponding parameter adjustment amplitude, the causal contribution degree of the vm.swappiness node is equal to the path causal strength 0.041 multiplied by the adjustment amplitude 0.15, which is equal to 0.00615, the causal contribution degree of the vm.dirty_ratio node is equal to the path causal strength 0.038 multiplied by the adjustment amplitude -0.08, which is equal to -0.00304, and the causal contribution degree of the kernel.numa_balancing node is equal to the path causal strength 0.052 multiplied by the adjustment amplitude 0.12, which is equal to 0.00624. The causal contribution degree comprehensively considers the parameter adjustment amplitude and the causal influence strength, and the larger the value is, the greater the contribution of the parameter adjustment to the performance improvement is. The causal contribution degrees of all the adjustment parameter nodes are arranged in descending order according to the absolute values, the absolute value of the causal contribution degree of the kernel.numa_balancing node is 0.00624, which is ranked first, the absolute value of the causal contribution degree of the vm.swappiness node is 0.00615, which is ranked second, and the absolute value of the causal contribution degree of the vm.dirty_ratio node is 0.00304, which is ranked third. An importance list of parameters is generated, which combines the complete paths of each parameter node in the causal graph to form an interpretable basis for system layer optimization decision-making. The operation and maintenance personnel can clearly understand the causal logic and contribution size of each parameter adjustment by checking the list.

[0079] In a specific embodiment, the weight values of each directed edge in the causal path sequence are extracted, and the weight values of all directed edges on the same path are multiplied to obtain the path causal strength, including:

[0080] Each directed edge is extracted in topological order from the causal path sequence, and the weight values of each directed edge stored in the three-layer causal directed acyclic graph are read to obtain an edge weight sequence;

[0081] All weight values belonging to the same causal path in the edge weight sequence are indexed in the path direction order to obtain a path weight vector;

[0082] Each weight value in the path weight vector is multiplied to calculate the multiplication result equal to the first weight value multiplied by the second weight value and then multiplied by all subsequent weight values to obtain the path causal strength;

[0083] The path causality strength corresponding to each adjustment parameter node is stored as a key-value pair mapping relationship, where the key is the adjustment parameter node identifier and the value is the path causality strength value, thus obtaining the path causality strength mapping table.

[0084] Specifically, directed edges are extracted from the causal path sequence in topological order, meaning they are processed sequentially from upstream to downstream nodes according to the order of causal propagation. Assuming the causal path of the `vm.swappiness` parameter node contains three directed edges: the first edge connects the `vm.swappiness` node to the page table miss count node; the second edge connects the page table miss count node to the memory access latency node; and the third edge connects the memory access latency node to the average response latency node. These three edges are extracted sequentially according to the causal propagation order. The weight of the first edge stored in the three-layer causal directed acyclic graph is read as 0.42. The weight of the second edge is read as... The weight of the first edge is 0.35, and the weight of the third edge is 0.28. These three weight values ​​are arranged in the order of extraction to form an edge weight sequence. The same operation is performed on the causal path of the vm.dirty_ratio parameter node. Assuming that the path contains three edges with weight values ​​of 0.38, 0.31, and 0.32 respectively, the same edge weight sequence is extracted for the causal path of the kernel.numa_balancing parameter node. Assuming that the path contains three edges with weight values ​​of 0.45, 0.36, and 0.32 respectively, the edge weight sequences corresponding to the causal paths of all adjusted parameter nodes are collected to obtain a complete set of edge weight sequences.

[0085] The path weight vector is constructed by indexing all weight values ​​belonging to the same causal path in the edge weight sequence according to the path direction. For the causal path of the vm.swappiness parameter node, its edge weight sequence contains three weight values: 0.42, 0.35, and 0.28. Following the path direction from the parameter configuration layer to the intermediate state layer and then to the performance index layer, the first weight value 0.42 is marked as the first element of the vector, the second weight value 0.35 is marked as the second element, and the third weight value 0.28 is marked as the third element. After construction, a path weight vector containing three elements is obtained. For the vm.dirty_ratio parameter node, the path weight vector contains three elements: 0.38, 0.31, and 0.32. For the kernel.numa_balancing parameter node, the path weight vector contains three elements: 0.45, 0.36, and 0.32. The path weight vector maintains the order of the weight values ​​in the causal path to ensure that subsequent cumulative multiplication operations are performed in the correct causal propagation direction.

[0086] The multiplication operation is performed on each weight value in the path weight vector, and for the path weight vector of the vm.swappiness parameter node, the first weight value 0.42 is taken as the initial value of the multiplication, the initial value 0.42 is multiplied by the second weight value 0.35 to obtain an intermediate result 0.147, and the intermediate result 0.147 is continuously multiplied by the third weight value 0.28 to obtain a final multiplication result 0.041, which is the path causal strength of the vm.swappiness parameter node. The path causal strength value reflects the comprehensive influence degree of the parameter node on the performance indicator node through the causal path. The multiplication operation is performed on the path weight vector of the vm.dirty_ratio parameter node, the first weight value 0.38 is multiplied by the second weight value 0.31 to obtain 0.118, and then multiplied by the third weight value 0.32 to obtain the path causal strength 0.038. The multiplication operation is performed on the path weight vector of the kernel.numa_balancing parameter node, the first weight value 0.45 is multiplied by the second weight value 0.36 to obtain 0.162, and then multiplied by the third weight value 0.32 to obtain the path causal strength 0.052. The multiplication operation synthesizes the strengths of multiple causal influence links on the path, and the longer the path, the more edges it contains, and the path causal strength after multiplication is usually smaller.

[0087] The path causal strength corresponding to each adjustment parameter node is stored as a key-value mapping relationship to construct a path causal strength mapping table. The mapping table adopts a hash table data structure, the key is the identifier of the adjustment parameter node, that is, the parameter name string, and the value is the path causal strength value. A first record is created in the mapping table, the key is set as the vm.swappiness string, and the value is set as 0.041. A second record is created, the key is set as the vm.dirty_ratio string, and the value is set as 0.038. A third record is created, the key is set as the kernel.numa_balancing string, and the value is set as 0.052. The mapping table supports fast query of the corresponding path causal strength value through the parameter name. The operation and maintenance personnel can input the parameter name to obtain the influence strength of the parameter on the system performance through the causal path. The mapping table serves as a data basis for subsequent calculation of causal contribution degree. The path causal strength is multiplied by the parameter adjustment amplitude to obtain the actual contribution of each parameter to performance improvement.

[0088] The power server system layer optimization method for the high-load scenario in the embodiments of the present application is described above, and the power server system layer optimization system for the high-load scenario in the embodiments of the present application is described below. Please refer to Figure 2 The power server system layer optimization system for the high-load scenario in the embodiments of the present application includes one embodiment:

[0089] The data acquisition module is used to collect data on the CPU instruction cycle count, cache hit rate, page table miss count, memory access latency, and block device queue depth of the computing server to obtain a system performance dataset.

[0090] The analysis module is used to perform causal relationship analysis on the system performance dataset using the Granger causality test algorithm to obtain a causal adjacency matrix, and to construct a three-layer causal directed acyclic graph based on the causal adjacency matrix;

[0091] The backtracking module is used to locate bottleneck nodes whose anomalies exceed a threshold in the three-layer causal directed acyclic graph, and backtrack from the bottleneck node to the parameter configuration layer to obtain an effective set of adjustments.

[0092] The adjustment module is used to set the dimension of parameters that are not in the effective adjustment set to zero to reduce the action space, and iteratively perform parameter adjustment within the reduced action space to obtain the target parameter configuration vector;

[0093] The traversal module is used to perform a forward traversal of the adjustment parameters of the target parameter configuration vector through the three-layer causal directed acyclic graph to obtain the causal path sequence, and calculate the causal contribution by multiplying the causal edge weights and parameter adjustment magnitudes in the causal path sequence.

[0094] above Figure 2 The optimization system for high-load scenarios of the computing server system layer in this embodiment of the invention is described in detail from the perspective of modular functional entities. The optimization device for high-load scenarios of the computing server system layer in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0095] Reference Figure 3 This invention also provides a computing server system-level optimization device for high-load scenarios. This device can be a server, and its internal structure can be as follows: Figure 3 As shown, this computing server system-level optimization device for high-load scenarios includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor in this computer design provides computing and control capabilities. The memory of this computing server system-level optimization device for high-load scenarios includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of this computing server system-level optimization device for high-load scenarios stores the data corresponding to this embodiment. The network interface of this computing server system-level optimization device for high-load scenarios is used for communication with external terminals via network connection. When the computer program is executed by the processor, it implements the above-described method.

[0096] Those skilled in the art can understand that, Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the application, and does not constitute a limitation on the high-load scenario oriented computing power server system layer optimization equipment to which the scheme of the application is applied.

[0097] The application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium or a volatile computer readable storage medium, and the computer readable storage medium stores instructions, and when the instructions run on a computer, the computer executes the steps of the high-load scenario oriented computing power server system layer optimization method.

[0098] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-mentioned system, system and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0099] The integrated unit, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical scheme of the application or the whole or part of the technical scheme that contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a high-load scenario oriented computing power server system layer optimization equipment (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0100] The above embodiments are only used to illustrate the technical scheme of the application, rather than limit it; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical scheme recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical scheme deviate from the spirit and scope of the technical scheme of the embodiments of the application.

Claims

1. A computing power server system layer optimization method for high-load scenarios, characterized in that, The method comprises: Step S1: collecting CPU instruction cycle number, cache hit rate, page table miss times, memory access delay and block device queue depth of the computing power server to obtain a system performance data set; Step S2: performing causal relationship analysis on the system performance data set by a Granger causality test algorithm to obtain a causal adjacency matrix, and constructing a three-layer causal directed acyclic graph according to the causal adjacency matrix; Step S3: locating a bottleneck node with an abnormality degree exceeding a threshold value in the three-layer causal directed acyclic graph, and backtracking from the bottleneck node to a parameter configuration layer to obtain an effective adjustment set; Step S4: setting parameters in a dimension not in the effective adjustment set to zero to obtain a reduced action space, and iteratively performing parameter adjustment in the reduced action space to obtain a target parameter configuration vector; Step S5: performing forward traversal of the three-layer causal directed acyclic graph from an adjustment parameter of the target parameter configuration vector to obtain a causal path sequence, and calculating a product of a causal edge weight and a parameter adjustment amplitude in the causal path sequence to obtain a causal contribution degree.

2. The high-load-scene-oriented computing server system layer optimization method of claim 1, wherein, The step S1 comprises: A performance monitoring module deployed in the operating system kernel layer synchronously reads system files at a preset sampling period to obtain CPU instruction cycle number, page table miss times, memory access delay and block device queue depth, and obtains original performance index sampling data; The original performance index sampling data is respectively statistically processed in a short time window, a medium time window and a long time window to calculate the arithmetic mean, variance and change rate of each performance index in each time window, and to obtain multi-time window statistical characteristics; The statistical values of each performance index in the three time windows in the multi-time window statistical characteristics are spliced by rows to construct a feature matrix, and a feature matrix is obtained; Each dimension feature value in the feature matrix is standardized to obtain a system performance data set.

3. The high-load-scene-oriented computing server system layer optimization method of claim 1, wherein, The step S2 comprises: Any two time series indexes are extracted from the system performance data set as an index pair, a first time series index in the index pair is defined as an explained index, and a second time series index in the index pair is defined as an explanation index, an autoregressive model and an extended autoregressive model are constructed, and a residual sum of squares is calculated; An F statistic is calculated according to the residual sum of squares, when the F statistic is greater than a preset threshold value, corresponding elements in the causal adjacency matrix are set to 1, and all index pairs are traversed to obtain the causal adjacency matrix; A parameter configuration layer, an intermediate state layer and a performance index layer are constructed according to the causal adjacency matrix, wherein the parameter configuration layer comprises a kernel parameter node, the intermediate state layer comprises a performance index node, and the performance index layer comprises a throughput node and a response delay node; Directed edges are established between corresponding nodes according to positions with an element value of 1 in the causal adjacency matrix to obtain the three-layer causal directed acyclic graph.

4. The high-load-scene-oriented computing server system layer optimization method of claim 1, wherein, The step S3 comprises: Extract the current values of each performance indicator node from the middle state layer of the three-layer causal directed acyclic graph, calculate the abnormality degree of each performance indicator node as the current value minus the historical mean divided by the historical standard deviation, and select the performance indicator nodes with abnormality degrees exceeding the preset threshold as bottleneck nodes; Perform reverse breadth-first search along the directed edges of the three-layer causal directed acyclic graph from the bottleneck nodes, record all nodes passed in the search path, and stop the search when reaching the parameter configuration layer to obtain a reverse search path set; Extract all kernel parameter nodes located in the parameter configuration layer from the reverse search path set, store the identifiers of the kernel parameter nodes into a set container to obtain the effective adjustment set; Count the number of kernel parameter nodes in the effective adjustment set, compare the number of kernel parameter nodes with the total number of parameter nodes in the parameter configuration layer to calculate the parameter space reduction ratio, and obtain the parameter space reduction ratio.

5. The high-load-scene-oriented computing server system layer optimization method of claim 1, wherein, The step S4 comprises: Construct an initial action space containing all kernel parameter nodes in the parameter configuration layer, traverse each parameter dimension in the initial action space, judge whether the kernel parameter node corresponding to each parameter dimension is in the effective adjustment set, and assign the adjustment amplitude of the parameter dimension not in the effective adjustment set to zero to obtain the reduced action space; Read the current system state vector from the middle state layer of the three-layer causal directed acyclic graph, input the system state vector into the policy network to generate a parameter adjustment action, and limit the value of each dimension of the parameter adjustment action within the non-zero dimension range of the reduced action space to obtain the parameter adjustment action of the current iteration; Modify the corresponding kernel parameter value according to the parameter adjustment action and collect a new system state vector after the system enters a steady state, calculate the weighted sum of the throughput change value and the delay change value as the immediate reward value, and store the system state vector, the parameter adjustment action and the immediate reward value into an experience replay buffer; Randomly sample batch data from the experience replay buffer to train the policy network and the value network, terminate the iteration when the average reward change rate of consecutive multiple iterations is less than a preset convergence threshold, and extract the parameter adjustment vector output by the policy network as the target parameter configuration vector.

6. The high-load-scene-oriented computing server system layer optimization method of claim 1, wherein, The step S5 comprises: Extract the parameter nodes with non-zero adjustment amplitudes from the target parameter configuration vector as adjustment parameter nodes, perform forward breadth-first traversal in the three-layer causal directed acyclic graph starting from each adjustment parameter node, record the intermediate state layer nodes passed and the performance indicator layer nodes reached in the traversal process, and obtain the causal path sequence; Extract the weight values of each directed edge in the causal path sequence, perform a continuous multiplication operation on the weight values of all directed edges in the same path to obtain the path causal strength; Extract the parameter adjustment amplitudes corresponding to each adjustment parameter node in the target parameter configuration vector, multiply the path causal strength and the corresponding parameter adjustment amplitude to obtain the causal contribution degree. Ranking the causal contribution degrees of all adjustment parameter nodes in descending order according to numerical values to generate a parameter importance list as an interpretable basis for system layer optimization decision-making.

7. The computing power server system layer optimization method for high-load scenarios according to claim 6, characterized in that, The weight values of each directed edge in the causal path sequence are extracted, and the weight values of all directed edges on the same path are multiplied to obtain a path causal strength, including: Each directed edge in the causal path sequence is extracted in topological order to read the weight values stored in the three-layer causal directed acyclic graph to obtain an edge weight sequence; All weight values belonging to the same causal path in the edge weight sequence are indexed in the order of path direction to obtain a path weight vector; Each weight value in the path weight vector is multiplied to calculate the multiplication result equal to the first weight value multiplied by the second weight value and then multiplied by all subsequent weight values to obtain a path causal strength; The path causal strength corresponding to each adjustment parameter node is stored as a key-value mapping relationship, where the key is the adjustment parameter node identifier and the value is the path causal strength numerical value, to obtain a path causal strength mapping table.

8. A computing power server system layer optimization system for a high-load scenario, characterized in that, The system layer optimization method for high-load scenario-oriented computing power server system according to any one of claims 1-7 comprises: A collection module for collecting CPU instruction cycle count, cache hit rate, page table miss count, memory access delay, and block device queue depth of the computing power server to obtain a system performance dataset; An analysis module for performing causal relationship analysis on the system performance dataset by using the Granger causality test algorithm to obtain a causal adjacency matrix, and constructing a three-layer causal directed acyclic graph according to the causal adjacency matrix; A backtracking module for locating a bottleneck node with an abnormal degree exceeding a threshold value in the three-layer causal directed acyclic graph, and backtracking from the bottleneck node to the parameter configuration layer to obtain an effective adjustment set; An adjustment module for setting parameters outside the effective adjustment set to zero to obtain a reduced action space, and iteratively performing parameter adjustment in the reduced action space to obtain a target parameter configuration vector; A traversal module for performing forward traversal from the adjustment parameters of the target parameter configuration vector through the three-layer causal directed acyclic graph to obtain a causal path sequence, and calculating the product of the causal edge weight and the parameter adjustment amplitude in the causal path sequence to obtain a causal contribution degree.

9. A computing power server system layer optimization device for a high-load scenario, characterized in that, The computer program runs on the processor, and the processor executes the computer program to implement the system layer optimization method for high-load scenario-oriented computing power server system according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program runs on the processor, and the processor executes the computer program to implement the system layer optimization method for high-load scenario-oriented computing power server system according to any one of claims 1-7.

Images