A business hall computing power resource dynamic monitoring and intelligent scheduling method

Abstract

The present application relates to the technical field of computing power resource scheduling, and particularly relates to a business hall computing power resource dynamic monitoring and intelligent scheduling method, which comprises the following steps: collecting operation data of each node, extracting operation data of a current analysis period and calling historical benchmark data; for the current analysis period, using the jitter degree of CPU utilization of each node to weight the mean value of the CPU utilization, to obtain the load degree of each node in CPU utilization performance; extracting the repetition rate and duration of each node process, network delay and packet loss rate, comparing and evaluating the abnormality to correct and obtain the effective task load degree of each node, combining the average value of the task real-time level of each node and the effective concurrent load factor obtained by boundary prediction to calculate the computing power resource allocation priority of each node; based on the proportion of the priority of each node in the global total priority, calculating the CPU resource allocation quota of each node to dynamically allocate CPU resources. The present application effectively distinguishes pseudo high load and avoids computing power mismatch.

Description

Technical Field

[0001] This invention relates to the field of computing power resource scheduling technology, and in particular to a method for dynamic monitoring and intelligent scheduling of computing power resources in a business hall. Background Technology

[0002] With the intelligent development of bank and telecommunications service halls, digital services such as video surveillance analysis, OCR document recognition, and self-service terminals are becoming increasingly common. When faced with random fluctuations such as peak business hours and sudden increases in customer traffic, the CPU resource utilization of computing units such as counter terminals and edge servers in service halls will rise rapidly. Therefore, dynamic monitoring and intelligent scheduling of computing resources within service halls are important means to ensure business processing efficiency and user service experience.

[0003] Existing computing resource scheduling methods typically use a single CPU utilization rate or task queue length as the allocation basis, such as allocating more computing power when CPU utilization exceeds a threshold. However, in multi-service concurrent scenarios in business halls, this traditional strategy has the following obvious drawbacks: First, business hall operations are highly dependent on network coordination. When network communication experiences delays or congestion, or when the underlying system experiences abnormal lag or deadlocks, business request responses are delayed. At this time, the CPU is not actually processing effective tasks, but the apparent utilization rate is extremely high, presenting a false accumulation of tasks. Traditional methods are prone to misjudging this as insufficient computing power and blindly allocating resources. This not only fails to improve the efficiency of faulty nodes but also preempts and affects other healthy nodes. Firstly, the computing power of nodes is limited. Secondly, service halls often face sudden, short-term, high-frequency tasks, causing rapid fluctuations in CPU resources. Traditional methods lack the ability to extract these fluctuation characteristics, easily misjudging healthy system responses as system bottlenecks. Thirdly, adopting a single, fixed priority preemption strategy fails to perceive the global concurrent customer flow, easily leading to complex background tasks being shelved for extended periods during peak business hours, resulting in resource starvation. In summary, existing technologies cannot effectively distinguish between the mismatch in computing power demand caused by network latency, system anomalies, and normal transient fluctuations, and cannot take into account the dynamic balance of global tasks. Ultimately, this results in the overall computing power allocation in service halls not meeting actual needs, leading to decreased system throughput efficiency and service experience. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method for dynamic monitoring and intelligent scheduling of computing resources in business halls, in order to solve the above-mentioned technical problems in multi-service concurrent scenarios, such as the inability to distinguish between network latency, system anomalies and transient fluctuations leading to CPU resource scheduling mismatch, and the lack of global customer flow perception causing secondary tasks to be unable to obtain resources.

[0005] This invention provides a method for dynamic monitoring and intelligent scheduling of computing resources in a business hall, comprising the following steps: collecting operational data from each node, including CPU utilization, network status data, process list, and pending task list; extracting operational data for the current analysis period and retrieving corresponding historical benchmark data; for the current analysis period: calculating the jitter of CPU utilization for each node, weighting the average CPU utilization of each node using the jitter to obtain the load level of each node in terms of CPU utilization performance; extracting the repetition rate and duration of various processes for each node in the process list, and comparing them with historical benchmark data to obtain the possibility of system operation anomalies; and comparing network latency and packet loss rate in the network status data with corresponding historical benchmark data. The network anomaly level is obtained by analyzing changes in historical benchmark data. The load level is then adjusted based on the probability of system malfunctions and the degree of network anomalies to obtain the effective task load level of each node. The average real-time performance level of each node is obtained based on the list of tasks to be processed, and the total number of tasks to be processed in the system is compared with the maximum total number of tasks in historical benchmark data for boundary prediction to obtain the effective concurrent load factor. The computing resource allocation priority of each node is calculated by combining the effective task load level, the average real-time performance level, and the effective concurrent load factor. Based on the proportion of each node's computing resource allocation priority in the global total priority, the CPU resource allocation quota of each node is calculated in combination with the total schedulable CPU resources to dynamically allocate CPU resources.

[0006] Preferably, obtaining the load level of each node in terms of CPU utilization performance includes: calculating the absolute value of the difference in CPU utilization of each node at all adjacent monitoring times within the current analysis period, summing and averaging the values, and standardizing the summation and averaging results to obtain the jitter level of CPU utilization of each node; calculating the difference between constant 1 and the jitter level as a negative weighting coefficient, and multiplying the mean of CPU utilization of each node within the current analysis period by the negative weighting coefficient to obtain the load level of each node in terms of CPU utilization performance.

[0007] Preferably, the probability of the system malfunctioning satisfies the expression: In the formula, Indicates the number of times within the current analysis period. The possibility of abnormal system operation at each node; Indicates the number of times within the current analysis period. The total number of processes on each node; Indicates the number of times within the current analysis period. The degree of abnormal frequency of process occurrence on each node; Indicates the number of times within the current analysis period. The degree of abnormality in the process duration of each node; This represents the standard normalization function.

[0008] Preferably, the method for obtaining the degree of abnormality in the frequency of process occurrence of the node is as follows: the ratio of the number of occurrences of each type of process in each node to the total number of processes in that node is taken as the repetition rate; the repetition rate fitting value is obtained by performing time-series decomposition based on the repetition rate of the same type of process in the same node in the same period of historical benchmark data; the absolute difference between the repetition rate and the repetition rate fitting value is calculated, and the ratio of the absolute difference to the repetition rate fitting value is negatively taken and input into an exponential function for mapping to obtain the probability that the repetition rate is normal; the difference between the constant 1 and the probability that the repetition rate is normal is calculated as the abnormality penalty weight, the repetition rate is multiplied by the abnormality penalty weight and subjected to standard normalization processing to obtain the frequency abnormality score of each process, and the maximum value of the frequency abnormality score is extracted as the degree of abnormality in the frequency of process occurrence.

[0009] Preferably, the method for obtaining the abnormality degree of process duration is as follows: obtain the duration of each occurrence of the same node and process in the current analysis period from historical benchmark data, and the average duration of all occurrences; calculate the difference between the duration of each node and each process in the current analysis period and the duration of each occurrence, divide it by the average duration of all occurrences, sum the quotients and take the average, and then perform standard normalization to obtain the abnormality score of the duration of each process, and extract the maximum value of the abnormality score of the duration as the abnormality degree of process duration.

[0010] Preferably, obtaining the network anomaly level includes: extracting the mean network latency, the standard deviation of network latency, and the mean packet loss rate of each node within the current analysis period, dividing each by the maximum value in the corresponding historical benchmark data, and then summing them.

[0011] Preferably, the effective task load of each node satisfies the expression: In the formula, Indicates the number of times within the current analysis period. Effective task load level of each node; Indicates the number of times within the current analysis period. The load level of each node in terms of CPU utilization; Indicates the number of times within the current analysis period. The possibility of abnormal system operation at each node; Indicates the number of times within the current analysis period. The degree of network anomaly at each node; This represents the natural exponential function.

[0012] Preferably, the step of comparing the total number of tasks to be processed in the system with the maximum total number of tasks in the historical benchmark data and making boundary predictions to obtain the effective concurrent load factor includes: calculating the ratio of the total number of tasks to be processed in the system to the maximum total number of tasks in the historical benchmark data and performing standard normalization to obtain the initial macro concurrent load factor; if the initial macro concurrent load factor is greater than a set upper boundary or less than a set lower boundary, then the effective concurrent load factor is fixed to the corresponding upper boundary or lower boundary; if it is within the set lower boundary and upper boundary interval, then the initial macro concurrent load factor is directly used as the effective concurrent load factor.

[0013] Preferably, the priority of computing resource allocation for each node satisfies the expression: In the formula, Indicates the number of times within the current analysis period. Priority of computing resource allocation for each node; Indicates the number of times within the current analysis period. Effective task load level of each node; Indicates the number of times within the current analysis period. The average real-time performance level of each node; This indicates the maximum real-time performance level set by the system. This represents the effective concurrent load factor obtained after boundary prediction.

[0014] Preferably, the step of calculating the CPU resource allocation quota for each node in combination with the total schedulable CPU resources includes: summing the computing power resource allocation priorities of each node to obtain the global total priority; dividing the computing power resource allocation priority of each node by the global total priority to obtain the computing power allocation ratio of each node in the current scheduling cycle; and multiplying the computing power allocation ratio by the total schedulable CPU resources to obtain the CPU resource allocation quota for each node.

[0015] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows: This invention, based on monitoring the average CPU utilization, introduces CPU utilization jitter characteristics to reduce the interference of transient high load illusions caused by healthy short-term high-concurrency tasks. Simultaneously, by analyzing the abnormal repetition rate and duration of node processes and comparing with historical benchmarks, it suppresses invalid load caused by system deadlocks. Furthermore, it uses network state anomaly correction to obtain the effective task load of nodes. This mechanism improves the accuracy of computing power demand assessment and reduces the consumption of global resources by invalid scheduling. Based on a dynamic adjustment mechanism of the concurrent load factor and setting a safety boundary interval, it can adaptively adjust the allocation weight of real-time front-end services and long-term back-end services according to the real-time customer flow pressure in the business hall. This ensures the response speed of front-end interactions during peak business periods and avoids the phenomenon of complex back-end tasks stagnating for extended periods under extreme traffic through a safety net mechanism, achieving a smooth flow of computing resources among nodes. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for dynamic monitoring and intelligent scheduling of computing resources in a business hall, provided in Embodiment 1 of the present invention. Detailed Implementation

[0018] Embodiments of this disclosure are described in detail below, with examples of these embodiments illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting it.

[0019] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0020] This invention provides a method for dynamic monitoring and intelligent scheduling of computing resources in a business hall, such as... Figure 1 As shown, the method includes the following steps: Step S101: Collect the operating data of each node, including CPU utilization, network status data, process list and pending task list, extract the operating data of the current analysis period and retrieve the corresponding historical benchmark data.

[0021] It should be noted that nodes within a service hall typically include computing units with independent task processing capabilities, such as counter terminals, edge servers, and self-service terminals. Each node usually corresponds to an independent operating system. Through the underlying operating system, the physical and logical resource consumption status can be directly monitored, providing basic data support for subsequent judgments on network congestion, system deadlock, or healthy high concurrency. Due to the rapid changes in service hall operations, high-frequency monitoring is needed to capture instantaneous data spikes. However, if the resource scheduling frequency is too high, it will lead to frequent CPU context switching, causing system instability. Therefore, it is necessary to decouple the data collection frequency from the resource scheduling frequency and aggregate monitoring data by defining specific time windows.

[0022] Specifically, by calling system commands through the operating system of each node in the business hall, the underlying operating data of each node is monitored and collected in real time according to the preset collection cycle. The collected data includes: CPU utilization, network status data including network latency and packet loss rate, process list information including the number of times a process appears and its duration, and a list of pending tasks including the total number of tasks and the real-time level of the tasks.

[0023] To ensure system stability, the system is configured to perform CPU resource scheduling calculation once according to a preset scheduling cycle. The current analysis period is defined here: that is, each time the scheduling calculation is triggered, the system will extract a set of data from multiple consecutive preset collection cycles, including the current time, as a complete time window to participate in the subsequent evaluation.

[0024] After acquiring the data for the current analysis period, in order to eliminate the interference of different units on subsequent calculations, the CPU utilization, network latency, and packet loss rate data for the current analysis period are standardized and normalized respectively; and historical benchmark data of similar nodes for anomaly comparison are retrieved simultaneously.

[0025] It should be added that the preset acquisition period is used to capture transient business changes, and the preset scheduling period is used to smooth scheduling actions to avoid kernel oscillation. In this embodiment, the preset acquisition period is 1 second to ensure that it can capture short-lived tasks such as face verification and OCR recognition with fine precision. The preset scheduling period is 5 seconds. The value of 5 seconds is based on the fact that the average life cycle of routine short-term tasks in the business hall, such as single document recognition, interface call or network addressing, is usually within a few seconds. The 5-second scheduling window can accommodate a complete business fluctuation cycle to verify whether the CPU is in a healthy fluctuation state, and will not cause system kernel burden due to excessively frequent issuance of scheduling instructions. Implementers can adjust it according to the physical performance of the actual computing unit and the monitoring accuracy requirements. The historical benchmark data refers to the statistical values ​​of the status characteristics of the node during a period of normal operation in the past. In this embodiment, the process repetition rate, process duration, network latency extreme value and maximum total number of tasks of the same type of node in the previous day are taken to ensure that the comparison benchmark data can fully cover and reflect the operating rules of the business hall under a complete natural customer flow cycle including morning and evening business peaks and off-peak hours.

[0026] At this point, we have obtained normalized CPU and network characteristic data, node process and task status lists, and corresponding historical benchmark reference data for the current analysis period.

[0027] Step S102: For the current analysis period, calculate the fluctuation of CPU utilization of each node, and use the fluctuation to weight the average CPU utilization of each node to obtain the load level of each node in terms of CPU utilization performance.

[0028] It's important to note that in traditional CPU resource scheduling methods, when a node's CPU utilization exceeds a set threshold, it's typically considered to have insufficient computing power, triggering scheduling. However, in intelligent business hall scenarios, nodes often need to handle sudden, short-term concurrent tasks such as OCR document recognition and facial verification. When such tasks are triggered, the node's CPU utilization spikes rapidly, but resources are quickly released after the task is completed. This dramatic fluctuation is a normal dynamic response of the system handling transient tasks. If the CPU utilization remains high and fluctuates dramatically, it indicates that the processor can quickly process the task flow and complete context switching, and its processing capacity can meet the current business needs without additional computing power allocation. Conversely, if the CPU utilization remains high for a long time with minimal fluctuation, showing a stable state of continuous occupancy, it indicates that the node has a real computing power bottleneck or is experiencing process blocking. Therefore, it's necessary to extract the fluctuation characteristics of CPU utilization and use them as an adjustment factor to correct the overall utilization, in order to more accurately assess the node's true computing power requirements.

[0029] Specifically, the fluctuation level of CPU utilization for each node within the current analysis period is obtained as follows: The absolute values ​​of the differences in CPU utilization at each node across all adjacent monitoring times within the current analysis period are summed and averaged to obtain the average result. This average result is then normalized to obtain the fluctuation level of CPU utilization for each node within the current analysis period. This fluctuation level is then used to weight the mean CPU utilization of each node within the current analysis period to obtain the load level of each node in terms of CPU utilization performance within the current analysis period. The specific calculation formula is as follows: In the formula, Indicates the number of times within the current analysis period. The load level of each node in terms of CPU utilization; Indicates the number of times within the current analysis period. Average CPU utilization of each node; Indicates the number of times within the current analysis period. The degree of fluctuation in CPU utilization of each node.

[0030] in, This represents the negative weighting coefficient constructed based on CPU utilization jitter characteristics. The larger it is, the more likely it is to be the first The higher the fluctuation in CPU utilization of the first node, the more likely it is to cause problems. The stronger the ability of each node to handle transient tasks, the smaller the corresponding negative weighting coefficient value, thereby suppressing the interference of normal transient high load on the overall load assessment results; The larger it is, the more likely it is to be the first The high average CPU utilization and low jitter of the nth node mean that the nth node... The node exhibits a high load state with a lack of resource release cycles during the current analysis period, which means that the node is currently in a state of computing power shortage and needs the scheduling system to allocate more resources to it.

[0031] At this point, the load level of each node in terms of CPU utilization during the current analysis period has been obtained.

[0032] Step S103: Analyze the possibility of system operation anomalies based on the node process list and historical benchmark reference data, and make corrections based on the degree of network anomalies to obtain the effective task load of each node in the current analysis period.

[0033] It should be noted that while the load levels of nodes in terms of CPU utilization obtained from the aforementioned steps exclude the impact of health transient fluctuations, this seemingly stable and continuous high load may still be misleading. In actual business hall operation scenarios, in addition to the massive volume of real business concurrency, underlying software logic errors can also lead to increased resource consumption. For example, when a printing service component crashes, the daemon process may get stuck in a loop of restarting and reporting errors; or when a database query deadlocks, the processing thread will be suspended for a long time and continuously occupy CPU resources. In these cases of system malfunction, the high load exhibited by the node is not for processing actual effective business. In this case, even if the scheduling center allocates more computing power, it cannot solve the service lag problem, but will instead consume overall resources. Therefore, it is necessary to analyze the internal process state of the node, assess the possibility of system malfunction, and thus filter the load of basic tasks to ensure the reliability of the load assessment results.

[0034] Specifically, the probability that the repetition rate of each type of process at each node is normal within the current analysis period is obtained as follows: the ratio of the number of occurrences of each type of process at each node to the total number of processes at each node within the current analysis period is used as the repetition rate of each type of process at each node within the current analysis period. Time-series data is constructed by retrieving the repetition rates of the same type of process at the same node within the same time period from historical benchmark data. A known time-trend decomposition procedure based on Loess (STL) is used to obtain the time-repetition rate trend curve and periodic curve. By adding the segment trend term and periodic term, the fitted value of the repetition rate of each type of process at each node within the current analysis period is obtained. The absolute difference between the repetition rate of each type of process at each node within the current analysis period and the fitted value of the repetition rate is calculated. The ratio of the absolute difference to the fitted value of the repetition rate is negative and input into an exponential function for mapping to obtain the probability that the repetition rate of each type of process at each node is normal.

[0035] Furthermore, the repetition rates of various processes at each node are weighted based on the probability that the repetition rate is normal within the current analysis period, resulting in an anomaly score for each type of process at each node within the current analysis period; the specific calculation formula is as follows: In the formula, Indicates the number of times within the current analysis period. The node of the first Abnormal scores for process classes; Indicates the number of times within the current analysis period. The node of the first The repetition rate of process classes; Indicates the number of times within the current analysis period. The node of the first The likelihood that the repetition rate of the process class is normal; This represents the standard normalization function.

[0036] Extract the maximum abnormal score of all types of processes for each node within the current analysis period as the degree of abnormality in the frequency of process occurrence for that node.

[0037] After obtaining the frequency of abnormal occurrences of processes on this node, it should be noted that considering only the frequent repetition of processes is not enough to cover all failure modes. In some hidden deadlock scenarios, such as inter-process resource mutual exclusion waiting, abnormal processes will not crash and restart, but will remain in memory and run blocked, causing their life cycle to be abnormally prolonged. Therefore, it is necessary to supplement the abnormal monitoring of process duration.

[0038] Obtain the duration of each type of process for each node within the current analysis period. The duration refers to the time from the start time of the corresponding process to the last moment of the current analysis period. If the process has ended within the current analysis period, the duration is calculated up to the moment the process ends.

[0039] Based on the difference between the duration of each type of process at each node during the current analysis period and the duration of the same type of process at the same node in historical benchmark data, the probability of abnormal durations of each type of process at each node during the current analysis period is calculated: The specific calculation formula is as follows: In the formula, Indicates the number of times within the current analysis period. Each node process The possibility of abnormal duration; This indicates the first [number]th [item] in the current analysis period within the historical baseline data. Each node process Total number of occurrences; Indicates the number of times within the current analysis period. Each node process The duration; This indicates the first [number]th [item] in the current analysis period within the historical baseline data. Each node process The first The duration of each; This indicates the first [number]th [item] in the current analysis period within the historical baseline data. Each node process Mean duration of occurrence This represents the normalization function.

[0040] in, Reflecting the first [number]th [period] within the current analysis period Each node process The degree to which the duration deviates from its historical average duration; the larger the value, the stronger the process. The duration is significantly longer than the historical average, indicating that the process may have encountered a bottleneck, deadlock, or resource contention during execution, resulting in an abnormally extended cycle. This means that the process may be in an abnormal blocking or abnormal operation state, rather than a normal business processing flow.

[0041] Extract the maximum probability of abnormal process duration for each node within the current analysis period as the degree of abnormality in process duration for that node.

[0042] Considering the varying impacts of different deadlock types on the system process ecosystem, and taking into account the abnormality in process frequency and duration for each node during the current analysis period, the probability of system malfunction for each node during the current analysis period is calculated using the following formula: In the formula, Indicates the number of times within the current analysis period. The possibility of abnormal system operation at each node; Indicates the number of times within the current analysis period. The total number of processes on each node; Indicates the number of times within the current analysis period. The degree of abnormal frequency of process occurrence on each node; Indicates the number of times within the current analysis period. The degree of abnormality in the process duration of each node; This represents the standard normalization function.

[0043] in, Its complement reflects the current concurrent environment of the node, when Smaller A large value indicates that the total number of processes within the node is low. This suggests that due to some prolonged blocking, system resources are locked, preventing new processes from being generated normally. In this state of minimal parallel processes, the duration of a single process is abnormally high. The reference value of a process should be given greater weight; conversely, if the total number of processes is large, the focus should be on examining the frequency and degree of abnormal process awakenings. ; The larger the value, the more likely it is that the CPU resources of that node are being used by invalid processes, and the apparent high load is an anomaly.

[0044] It should be further explained that, after ruling out system malfunctions, there is an external interference situation in this distributed collaborative business hall scenario: a limited network environment. The business hall's business system heavily relies on data interaction via LAN or cloud. When network congestion, packet loss, or jitter occurs, the uplink and downlink transmission of business requests is delayed, causing processing threads within nodes to be suspended and wait for data responses. At this time, the task backlog and high CPU usage are caused by network bottlenecks, not by a lack of local computing power. If computing power is not allocated in a differentiated manner, it will lead to the misallocation of computing power to nodes with network obstruction. Therefore, network anomaly indicators are introduced for final load correction.

[0045] The network anomaly level of each node within the current analysis period is obtained. The network anomaly level is equal to the mean network latency, the standard deviation of network latency, and the mean packet loss rate of each node within the current analysis period, which are obtained by dividing each by the corresponding maximum value in the historical benchmark data and then summing them. The reason for proportionalizing and adding these three indicators is that the mean network latency reflects the lag in basic transmission, the standard deviation reveals the severity of network fluctuations, and the packet loss rate directly reflects the risk of connection breakage. These three dimensions independently and complementaryly constitute an overall characterization of network communication limitations. By proportionalizing, the dimensions of different physical quantities are unified, and the summation can comprehensively reflect the overall deterioration level of the network environment in which the current node is located.

[0046] Considering the likelihood of system malfunctions and the severity of network anomalies, the load levels of nodes in terms of CPU utilization are adjusted, and the effective task load level of each node in the current analysis period is calculated using the following formula: In the formula, Indicates the number of times within the current analysis period. Effective task load level of each node; Indicates the number of times within the current analysis period. The load level of each node in terms of CPU utilization; Indicates the number of times within the current analysis period. The possibility of abnormal system operation at each node; Indicates the number of times within the current analysis period. The degree of network anomaly at each node; This represents the natural exponential function.

[0047] in, As a linear filtering factor, it indicates that the more normal the system operation of a node, the higher the reliability of its load. The monotonically decreasing property of a negative exponent smoothly characterizes the decreasing trend in effective service conversion rate within the CPU as the network anomaly level increases; if The larger the value, the more likely the node is to be operating under continuous high load, and the system kernel is healthy and network communication is smooth. This indicates that the data accumulation is caused by the increased computing power demand due to real business volume. The effective task load level can objectively reflect the degree to which the node urgently needs resource scheduling.

[0048] At this point, the effective task load level of each node within the current analysis period has been obtained.

[0049] Step S104: Obtain the average real-time level of the task and predict the effective concurrent load factor. Combined with the effective task load level, calculate the priority of computing resource allocation for each node in the current analysis period.

[0050] It should be noted that in multi-service concurrent scenarios in the business hall, different types of services have significantly different tolerances for response latency. When allocating resources, priority should be given to processing high real-time tasks to maintain the basic service experience. However, if a single priority preemption strategy is implemented during peak business periods, a large number of high real-time tasks will occupy computing power for a long time, causing long-running background tasks with lower real-time requirements but complex computational logic to be unable to obtain CPU time slices, thus leading to task backlog and resource constraints. Conversely, if allocation is still based on real-time priority during off-peak business hours, redundant computing power cannot be fully utilized to clear the backlog of complex services. Therefore, the total number of concurrent tasks in the current business hall is introduced as a macro-environmental awareness factor to dynamically adjust the weight of effective task load and real-time priority in computing power allocation. At the same time, statistical analysis of the business hall's historical long-tail data reveals that when this concurrency load factor tends to be excessively extreme, it can easily cause secondary tasks to stagnate. To avoid excessive computing power occupation under extreme traffic conditions, boundary prediction is performed before priority calculation to achieve an adaptive balance between front-end response and back-end throughput.

[0051] Specifically, firstly, based on the response latency standards for each service in the business hall, the real-time requirements of each type of service are mapped to a preset level scale. In this embodiment, levels 1 to 4 are used. The scenario setting is based on the following: business hall services can generally be clearly divided into four typical tiers: Level 1 is backend non-real-time data synchronization or log archiving, with a relatively high latency tolerance; Level 2 is internal system status polling and resource monitoring, with a moderately high latency tolerance; Level 3 is general business queries at self-service terminals, with a low latency tolerance; and Level 4 is interactive front-end services such as face verification and OCR document recognition directly facing customers, with an extremely low latency tolerance and requiring rapid response. Using this scale, the real-time level of all tasks in the task queue of each node within the current analysis period is obtained, and their average value is calculated as the average real-time level of the tasks at that node.

[0052] Furthermore, the total number of tasks to be processed in the system during the current analysis period and the maximum total number of tasks to be processed in all periods of the historical baseline data are extracted; the ratio of the two is calculated and normalized to obtain the initial macroscopic concurrent load factor.

[0053] Before calculating the priority of computing resource allocation, a safety boundary prediction is performed on the initial macro-level concurrent load factor to obtain the effective concurrent load factor: If the initial macro-level concurrent load factor is greater than the set upper boundary, it indicates that the system is in a peak business period. To prevent complex background tasks from being unable to obtain resources, the effective concurrent load factor is fixed at this upper boundary. If the initial macro-level concurrent load factor is less than the set lower boundary, it indicates that the system is in a low-traffic period. To ensure that high-real-time front-end tasks maintain a basic response priority, the effective concurrent load factor is fixed at this lower boundary. If the initial macro-level concurrent load factor is within the safe boundary range formed by the lower and upper boundaries, then the initial macro-level concurrent load factor is directly used as the effective concurrent load factor.

[0054] It should be added that the upper and lower boundaries of the parameters are used to constrain the extreme degree of resource skew during peak and off-peak periods, constructing a bottom-level backup scheduling mechanism. In this implementation, the upper boundary is set to 0.8 and the lower boundary to 0.2. The basis for these values ​​is as follows: In the underlying interaction of the distributed system in the business hall, if the background daemon process, long connections such as database session synchronization, and node heartbeat packets do not obtain time slices for a long time, it is easy to trigger timeout disconnection or node unresponsiveness. Forcibly reserving at least 20% weight as a backup throughput line can maintain the continuous connection between the system heartbeat and the underlying data callback without excessively depriving the front-end of the computing power required for high concurrency. Setting the lower limit of 0.2 is to reserve basic preemption weight for sudden sporadic front-end interactions during idle periods, such as customers suddenly entering the self-service terminal, to ensure timely business response. Implementation personnel can adjust the values ​​according to the actual heartbeat timeout threshold of the business hall and the physical performance of the server.

[0055] Based on the effective task load level, average task real-time performance level, and effective concurrent load factor of each node obtained in the previous steps during the current analysis period, the priority of computing resource allocation for each node during the current analysis period is calculated. The specific calculation formula is as follows: In the formula, Indicates the number of times within the current analysis period. Priority of computing resource allocation for each node; Indicates the number of times within the current analysis period. Effective task load level of each node; Indicates the number of times within the current analysis period. The average real-time performance level of each node; This indicates the maximum real-time performance level set by the system. This represents the effective concurrent load factor obtained after boundary prediction.

[0056] in, This reflects the current state of customer flow and concurrent user pressure in the overall system of the service hall. The allocation weights, representing the actual resource demand of nodes, are composed of... The driving force means that during idle periods when the macro-level concurrency pressure is lower, the weight of this part is greater. The scheduling system tends to allocate computing power to nodes with a high load of effective tasks, that is, nodes with a large backlog of complex and long-term business, thereby accelerating the clearing of the backlog in the background. The weighting of tasks based on the front-end business service experience is directly driven by the effective concurrent load factor. This means that as the system approaches a high-concurrency state, this weight increases, and the system allocates computing power to user interaction nodes with higher average task real-time performance, thereby ensuring a better core service experience during peak periods. In summary... The larger the value, the higher the urgency of computing power scheduling in the overall planning, after combining the current macro passenger flow pressure with the micro task characteristics of the node.

[0057] At this point, the priority of computing resource allocation for each node within the current analysis period has been obtained.

[0058] Step S105: Calculate the CPU resource allocation quota for each node based on the priority of computing power resource allocation and the total schedulable CPU resources to dynamically allocate CPU resources.

[0059] It should be noted that in the distributed collaborative processing scenario of the business hall, after obtaining the reliable computing power allocation priority of each node, these priority values ​​are converted into computing power scheduling instructions at the physical layer. That is, in the post-processing and execution stage, the single preemptive allocation mode is abandoned. By placing the priority of each node into the global total priority for relative proportional conversion, it is possible to achieve smooth and dynamic flow of resources under the objective condition of limited total computing power resources, thereby maximizing the overall business processing concurrency of the business hall and maintaining system stability.

[0060] Specifically, after obtaining the computing power resource allocation priority of each node within the current analysis period, these priorities are summed to obtain the global total priority of the business hall system for the current analysis period. The computing power resource allocation priority of each node within the current analysis period is then divided by the global total priority to calculate the computing power allocation ratio of that node within the current scheduling cycle. Subsequently, this computing power allocation ratio is multiplied by the total schedulable CPU resources of the business hall to calculate the final CPU resource allocation quota for each node within the current analysis period.

[0061] After obtaining the CPU resource allocation quota for each node, the underlying scheduling system dynamically adjusts and issues physical allocation strategies based on the quota instructions. For example, it limits or relaxes the CPU time slice of the corresponding node through the underlying resource control tools, ensuring that nodes in the high-load and high-real-time foreground state can obtain the corresponding computing power to respond, while ensuring that redundant resources are effectively recovered and smoothly transferred, thus maintaining the stability of the system's business.

[0062] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for dynamic monitoring and intelligent scheduling of computing resources in a business hall, characterized in that, include: Collect operational data from each node, including CPU utilization, network status data, process list, and list of tasks to be processed. Extract operational data for the current analysis period and retrieve corresponding historical baseline data. For the current analysis period: calculate the jitter of CPU utilization of each node, and use the jitter to weight the mean of CPU utilization of each node to obtain the load level of each node in terms of CPU utilization performance. Extract the repetition rate and duration of various processes in each node of the process list, and compare them with historical benchmark data to obtain the possibility of abnormal system operation; The degree of network anomaly is obtained by comparing the changes in network latency and packet loss rate in network status data with corresponding historical baseline data. The load level is adjusted based on the probability of system malfunction and the degree of network anomaly to obtain the effective task load level of each node; The average real-time performance level of each node is obtained from the list of tasks to be processed. The total number of tasks to be processed in the system is compared with the maximum total number of tasks in the historical benchmark data and the boundary is predicted to obtain the effective concurrent load factor. The computing power resource allocation priority of each node is calculated by combining the effective task load level, the average real-time performance level of each task and the effective concurrent load factor. Based on the proportion of each node's computing resource allocation priority in the overall global priority, and combined with the total schedulable CPU resources, the CPU resource allocation quota for each node is calculated to dynamically allocate CPU resources.

2. The method for dynamic monitoring and intelligent scheduling of computing resources in a business hall according to claim 1, characterized in that, The process of obtaining the load level of each node in terms of CPU utilization includes: Calculate the absolute value of the difference in CPU utilization at each node for all adjacent monitoring times within the current analysis period, sum and average them, and then normalize the sum and average results to obtain the jitter level of CPU utilization at each node. Calculate the difference between constant 1 and the jitter level as a negative weighting coefficient, and multiply the mean of CPU utilization at each node within the current analysis period by the negative weighting coefficient to obtain the load level of each node in terms of CPU utilization performance.

3. The method for dynamic monitoring and intelligent scheduling of computing resources in a business hall according to claim 1, characterized in that, The probability of system malfunction satisfies the expression: ； In the formula, Indicates the number of times within the current analysis period. The possibility of abnormal system operation at each node; Indicates the number of times within the current analysis period. The total number of processes on each node; Indicates the number of times within the current analysis period. The degree of abnormal frequency of process occurrence on each node; Indicates the number of times within the current analysis period. The degree of abnormality in the process duration of each node; This represents the standard normalization function.

4. The method for dynamic monitoring and intelligent scheduling of computing resources in a business hall according to claim 3, characterized in that, The method for obtaining the degree of abnormality in the frequency of processes occurring in the node is as follows: The repetition rate is the ratio of the number of occurrences of each type of process in each node to the total number of processes in that node. The repetition rate fitting value is obtained by time-series decomposition based on the repetition rate of the same type of process in the same node in the same time period in historical benchmark data. The absolute difference between the repetition rate and the repetition rate fitting value is calculated, and the ratio of the absolute difference to the repetition rate fitting value is negative and then input into an exponential function for mapping to obtain the probability that the repetition rate is normal. The difference between constant 1 and the probability of normal repetition rate is calculated as the abnormality penalty weight. The repetition rate is multiplied by the abnormality penalty weight and then normalized to obtain the frequency abnormality score of each process. The maximum value of the frequency abnormality score is extracted as the degree of frequency abnormality of the process.

5. The method for dynamic monitoring and intelligent scheduling of computing resources in a business hall according to claim 3, characterized in that, The method for obtaining the abnormality level of the process duration is as follows: Obtain the duration of each occurrence of the same node and process within the current analysis period from the historical benchmark data, as well as the average duration of all occurrences; Calculate the difference between the duration of each process at each node within the current analysis period and the duration of each occurrence, divide it by the average duration of all occurrences, sum the resulting quotients, take the average, and then perform standard normalization to obtain the duration anomaly score for each process. Extract the maximum value among the duration anomaly scores as the degree of process duration anomaly.

6. The method for dynamic monitoring and intelligent scheduling of computing resources in a business hall according to claim 1, characterized in that, The acquisition of network anomaly level includes: The mean network latency, standard deviation of network latency, and mean packet loss rate of each node within the current analysis period are extracted, divided by the corresponding maximum value in the historical benchmark data, and then summed.

7. The method for dynamic monitoring and intelligent scheduling of computing resources in a business hall according to claim 1, characterized in that, The effective task load of each node satisfies the expression: ； In the formula, Indicates the number of times within the current analysis period. Effective task load level of each node; Indicates the number of times within the current analysis period. The load level of each node in terms of CPU utilization; Indicates the number of times within the current analysis period. The possibility of abnormal system operation at each node; Indicates the number of times within the current analysis period. The degree of network anomaly at each node; This represents the natural exponential function.

8. The method for dynamic monitoring and intelligent scheduling of computing resources in a business hall according to claim 1, characterized in that, The step of comparing the total number of tasks to be processed in the system with the maximum total number of tasks in historical benchmark data and making boundary predictions to obtain the effective concurrent load factor includes: The ratio of the total number of tasks to be processed in the system to the maximum total number of tasks in the historical baseline data is calculated and normalized to obtain the initial macro-level concurrent load factor. If the initial macro-level concurrent load factor is greater than the set upper boundary or less than the set lower boundary, the effective concurrent load factor is fixed to the corresponding upper or lower boundary. If it is within the range between the set lower and upper boundaries, the initial macro-level concurrent load factor is directly used as the effective concurrent load factor.

9. The method for dynamic monitoring and intelligent scheduling of computing resources in a business hall according to claim 1, characterized in that, The priority of computing resource allocation for each node satisfies the expression: ； In the formula, Indicates the number of times within the current analysis period. Priority of computing resource allocation for each node; Indicates the number of times within the current analysis period. Effective task load level of each node; Indicates the number of times within the current analysis period. The average real-time performance level of each node; This indicates the maximum real-time performance level set by the system. This represents the effective concurrent load factor obtained after boundary prediction.

10. The method for dynamic monitoring and intelligent scheduling of computing resources in a business hall according to claim 1, characterized in that, The calculation of CPU resource allocation quotas for each node based on the total schedulable CPU resources includes: The computing power resource allocation priorities of each node are summed to obtain the global total priority; the computing power resource allocation priority of each node is divided by the global total priority to obtain the computing power allocation ratio of each node in the current scheduling cycle; the computing power allocation ratio is multiplied by the total schedulable CPU resources to obtain the CPU resource allocation quota of each node.

Images