Large model service platform for high-efficiency concurrent computing

Abstract

The invention discloses a large model service platform for high-efficiency concurrent computing, particularly relates to the technical field of data analysis, and realizes optimal management of the large model service platform for high-efficiency concurrent computing through a task allocation module, a node health assessment module, a node division module, a dynamic adjustment module and a fault identification module. The task distribution module reasonably distributes computing resources according to task priorities and real-time monitoring data; the node health assessment module accurately assesses node health conditions through preprocessing and feature extraction; the node division module classifies the nodes according to the health assessment result; the dynamic adjustment module dynamically adjusts a resource allocation strategy by using fuzzy logic; the fault recognition module recognizes and repairs fault nodes in time by further analyzing the health conditions of the high-load nodes; the overall performance and the resource utilization rate of the system are improved, efficient and reliable execution of intelligent financial services is ensured, and the stability and the reliability of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and more specifically, to a large model service platform for high-efficiency concurrent computing. Background Art

[0002] High-performance concurrent computing refers to the technology of executing multiple computing tasks simultaneously in a computer system to improve the overall performance and resource utilization of the system. Through concurrent computing, multiple tasks can share processors, memory and other resources in the same time period, thereby reducing the waiting time and processing time of tasks. The key to high-performance concurrent computing is to reasonably schedule and manage tasks to avoid resource conflicts and deadlocks, and ensure that tasks can be executed quickly and efficiently. This technology is widely used in multi-core processors, distributed computing and cloud computing to meet the needs of high-performance computing and large-scale data processing.

[0003] Provide intelligent financial services by leveraging pre-trained large-scale AI models. These services include intelligent customer service, risk management, personalized recommendations, and automated business processes. Through large models, banks can analyze large amounts of customer data, predict credit risk and fraud, provide personalized product recommendations, and optimize processes such as loan approval, thereby improving customer experience and operational efficiency. Large model services are provided through APIs or cloud platforms, enabling banks to quickly apply advanced AI technologies to improve service quality and decision-making.

[0004] The prior art has the following disadvantages: When using pre-trained large-scale artificial intelligence models to provide intelligent financial services, the large-model service platform with high-efficiency concurrent computing faces the challenge of handling a large number of concurrent tasks. The platform must be able to dynamically allocate and adjust computing resources to ensure efficient and reliable execution of intelligent customer service, risk management, personalized recommendations, and automated business process tasks. However, with the increase in concurrent tasks, resource competition between computing nodes intensifies, causing some nodes to be overloaded while other nodes are idle, seriously affecting the overall performance and resource utilization of the system. In addition, in a distributed computing environment, the load data of the nodes changes in real time. Traditional health assessment methods are difficult to accurately reflect the health status of each node in a timely manner, which may lead to unreasonable task allocation and further increase the pressure on high-load nodes. At the same time, in the face of real-time changes in load and node health status, there is a lack of effective dynamic adjustment mechanisms and fault identification methods, and it is impossible to adjust resource allocation strategies in a timely manner and identify and handle faulty nodes, resulting in reduced system stability and reliability. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a large model service platform with high-efficiency concurrent computing to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: A large model service platform for high-efficiency concurrent computing, including a task allocation module, a node health assessment module, a node partitioning module, a dynamic adjustment module and a fault identification module; Task allocation module: Receives concurrent task requests from customer service content through the cloud platform, classifies tasks into different priorities, allocates tasks to computing nodes according to the priorities, and monitors the load data of each node in real time through the load monitor; Node health assessment module: pre-processes the load data of each node, extracts features from the pre-processed load data, and assesses the health status of each node based on the extracted frequency variation features of threads competing for resource locks and node power efficiency fluctuation features; Node classification module: Based on the evaluation results, the health status of each node is classified into different categories, namely high-load nodes and low-load nodes; Dynamic adjustment module: According to the health status of nodes and task priority, the resource allocation strategy is dynamically adjusted through fuzzy logic to optimize the execution order of tasks; Fault identification module: reallocates the tasks of each computing node according to the adjusted resource allocation strategy, and further analyzes the health status of high-load nodes within a fixed time period. According to the analysis results, the faulty nodes are identified in time and repaired.

[0007] Preferably, in the task allocation module, the task and node information are recorded in a task allocation table, which includes the task ID, the allocated node, the estimated completion time, and the allocated time. The estimated completion time of the task is estimated based on the task type and the processing capacity of the node; the task allocation information is sent to the selected computing node to notify the node to start task processing. After receiving the task, the computing node starts the task execution process and allocates corresponding computing resources according to the requirements of the task; the execution of the task is monitored in real time through the load monitor, including task progress, resource consumption, and execution time, and the task status in the task allocation table is updated regularly; after the task is completed, the computing node returns the result to the cloud platform, notifies the task completion status, and stores the task result in the database of the cloud platform for subsequent query and analysis.

[0008] Preferably, in the node health assessment module, a resource contention frequency anomaly index is generated according to the extracted frequency change characteristics of threads competing for resource locks of each node, and the method for obtaining the resource contention frequency anomaly index is: Obtain the resource contention frequency data within the time period t, and record the resource contention frequency data at N time points as R(t); select the fractional order parameter α, which determines the order of the transformation, and perform a fractional Fourier transform on the resource contention frequency data. The calculation formula is: ; In the formula, The result of fractional Fourier transform of frequency data is obtained by resource competition. is the kernel function of the fractional Fourier transform. Through the fractional Fourier transform, the spectrum components of different fractional orders are obtained. , ; Perform inverse transform according to the fractional order parameter α to reconstruct the signal, and the transformation expression is: ; In the formula, is the result of fractional Fourier inverse transform, and the resource contention frequency anomaly index is calculated. The specific calculation expression is: ; In the formula, Resource contention frequency anomaly index.

[0009] Preferably, in the node health assessment module, a power efficiency fluctuation index is generated according to the extracted power efficiency fluctuation characteristics of each node, and the method for obtaining the power efficiency fluctuation index is: The power consumption value of each node is obtained in real time through the power sensor, and the obtained real-time power consumption value is used to construct a time reference sequence Q= ; and obtain the corresponding historical power usage value time target sequence C= ; Compare and analyze the time series of current sensor output values ​​and historical data time series to build a The matrix D is n, n is the length of the reference sequence, and m is the length of the target sequence; where D(i, j) represents the distance between the i-th point in Q and the j-th point in C. The similarity between the real-time power usage value and the historical power usage value is calculated. The specific calculation expression is: D(i, j) = ; Where D(i,j) is the similarity between the real-time power usage value and the historical power usage value; Construct an n×m cumulative matrix M, where M(i,j) represents the minimum cumulative distance from (1,1) to (i,j). The specific calculation expression is: ; Start from M(n,m) and backtrack to find the optimal alignment path, which represents the alignment method with the minimum cumulative distance. The steps of the backtracking path are: start from (n,m) and look forward for the minimum value: from M(i,j) to ; Repeat this step until reaching (1,1); Calculate the total cumulative distance of the optimal alignment path, that is, calculate the power efficiency fluctuation index.

[0010] Preferably, in the node health assessment module, the resource contention frequency anomaly index and the power efficiency fluctuation index are normalized, and the health assessment coefficient of each node is calculated based on the normalized resource contention frequency anomaly index and the power efficiency fluctuation index.

[0011] Preferably, in the node division module, the health status of each node is divided into different categories according to the evaluation results, and the nodes are divided into high-load nodes and low-load nodes; The obtained health assessment coefficient of each node is compared with the health assessment coefficient reference threshold. If the health assessment coefficient of the node is greater than or equal to the health assessment coefficient reference threshold, the node is in a high load state, it is divided into a high load node, and a high load node signal is generated; if the health assessment coefficient of the node is less than the health assessment coefficient reference threshold, the node is in a low load state, it is divided into a low load node, and a low load node signal is generated.

[0012] Preferably, in the dynamic adjustment module, according to the health status of the node and the task priority, the resource allocation strategy is dynamically adjusted through fuzzy logic to optimize the execution order of the tasks, specifically: The health assessment coefficient and task priority are used as the input items of fuzzy logic, and the execution order of tasks is used as the output item of fuzzy logic; Define the fuzzy set of health assessment coefficient KM: low, medium, high; define the fuzzy set of task priority P: low, medium, high; define the fuzzy set of task execution order S: immediate execution, later execution, delayed execution; Establish fuzzy rules based on health assessment coefficients and task priorities; Fuzzify the input items KM and P, calculate the membership of each fuzzy set, apply fuzzy rules for reasoning, and calculate the membership of each output item; The result of fuzzy reasoning is defuzzified to obtain the specific task execution order value S, and the defuzzified value is calculated using the weighted average method.

[0013] Preferably, in the fault identification module, the tasks of each computing node are reallocated according to the adjusted resource allocation strategy, and the health status of the high-load node within a fixed time period is further analyzed, specifically: Further analysis is conducted on the health status of high-load nodes within a fixed time period. When a node is a high-load node, that is, the health assessment coefficient generated within a fixed time period is greater than or equal to the health assessment coefficient reference threshold, the health assessment coefficients generated in subsequent fixed time periods that are greater than or equal to the health assessment coefficient reference threshold are collected, and a corresponding data set is established. The mean and standard deviation of the data set are calculated. After analyzing the data set, the faulty nodes are identified in a timely manner according to the analysis results, and they are repaired.

[0014] Preferably, the mean and standard deviation of the health assessment coefficient in the data set are compared with the corresponding reference thresholds respectively. If the mean of the health assessment coefficient in the data set is greater than or equal to the reference threshold of the mean of the health assessment coefficient, and the standard deviation of the health assessment coefficient is less than the reference threshold of the standard deviation of the health assessment coefficient, it is marked as a faulty node and a secondary warning signal is generated, requiring further monitoring and additional resources; If the mean value of the health assessment coefficient is greater than or equal to the reference threshold of the mean value of the health assessment coefficient, and the standard deviation of the health assessment coefficient is greater than or equal to the reference threshold of the standard deviation of the health assessment coefficient, it will be marked as a faulty node and a first-level warning signal will be generated. It is necessary to immediately check and repair potential faults and reallocate tasks; If the mean value of the health assessment coefficient is less than the reference threshold value of the mean value of the health assessment coefficient, and the standard deviation of the health assessment coefficient is greater than or equal to the reference threshold value of the standard deviation of the health assessment coefficient, it is marked as a faulty node and a level 3 warning signal is generated. It is necessary to observe whether there are accidental events and adjust the resource scheduling strategy; If the mean of the health assessment coefficient is less than the reference threshold of the mean of the health assessment coefficient, and the standard deviation of the health assessment coefficient is less than the reference threshold of the standard deviation of the health assessment coefficient, it is in normal operation and is marked as a normal node. No warning signal is generated and no measures are required.

[0015] Technical effects and advantages of the present invention: 1. The present invention realizes efficient management and optimization of a large number of concurrent tasks in intelligent financial services through the collaborative work of task allocation module, node health assessment module, node division module, dynamic adjustment module and fault identification module. The platform can dynamically allocate and adjust computing resources to ensure efficient and reliable execution of intelligent customer service, risk management, personalized recommendations and automated business process tasks. By real-time monitoring of node load data and preprocessing and feature extraction, the system can accurately evaluate the health status of each node and divide it into high-load nodes and low-load nodes, thereby optimizing resource utilization, reducing resource competition and idleness, and improving overall system performance.

[0016] 2. The present invention uses fuzzy logic to dynamically adjust the resource allocation strategy according to the node health status and task priority, ensuring that high-priority tasks are processed first and optimizing the task execution order. The fault identification module further analyzes the high-load nodes within a fixed time period, and timely identifies and repairs the faulty nodes by calculating the mean and standard deviation of the health assessment coefficient to ensure the stability and reliability of the system. In summary, this solution not only improves the efficiency and resource utilization of the system in processing concurrent tasks, but also enhances the stability and reliability of the system, ensuring the efficient and stable operation of intelligent financial services. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings; Figure 1 This is a module diagram of a large model service platform for high-efficiency concurrent computing according to the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Embodiment: The present invention provides as Figure 1 A high-efficiency concurrent computing large model service platform is shown, comprising a task allocation module, a node health assessment module, a node partitioning module, a dynamic adjustment module and a fault identification module; Task allocation module: Receives concurrent task requests from customer service content through the cloud platform, classifies tasks into different priorities, allocates tasks to computing nodes according to the priorities, and monitors the load data of each node in real time through the load monitor; Node health assessment module: pre-processes the load data of each node, extracts features from the pre-processed load data, and assesses the health status of each node based on the extracted frequency variation features of threads competing for resource locks and node power efficiency fluctuation features; Node classification module: Based on the evaluation results, the health status of each node is classified into different categories, namely high-load nodes and low-load nodes; Dynamic adjustment module: According to the health status of nodes and task priority, the resource allocation strategy is dynamically adjusted through fuzzy logic to optimize the execution order of tasks; Fault identification module: reallocates the tasks of each computing node according to the adjusted resource allocation strategy, and further analyzes the health status of high-load nodes within a fixed time period. According to the analysis results, the faulty nodes are identified in time and repaired.

[0020] Among them, in the task allocation module, concurrent task requests from customer service content are received through the cloud platform, and the tasks are classified into different priorities. The tasks are allocated to computing nodes according to the priorities, and the load data of each node is monitored in real time through the load monitor.

[0021] The cloud platform receives service requests from customers, which can be concurrent and of various types of tasks, such as query, data processing, analysis, etc. For each customer request, a corresponding task object is created, recording detailed information of the task, such as task type, request time, required resources, etc.

[0022] Categorize tasks into different priorities based on their urgency, customer rating, and task type. For example, urgent query tasks may be assigned high priority, while routine data processing tasks may be assigned low priority. Mark each task with its priority to facilitate the subsequent scheduling and allocation process.

[0023] The load monitor monitors the load data of all computing nodes in real time, collects and summarizes the load data of each node, and provides real-time information for task allocation and resource scheduling. Tasks are allocated to computing nodes according to priority, and the priority information of tasks (such as high, medium, and low) is obtained. The computing resources required for the task (such as CPU, memory, and network bandwidth) are determined.

[0024] Records the unique identifier of the task. Records detailed information about the task, including task type, priority, resource requirements, etc. Records selected compute node information, including node ID and node name.

[0025] Record the task and node information in the task allocation table. The table should contain information such as task ID, allocated node, estimated completion time, allocated time, etc. Estimate the estimated completion time of the task based on the task type and the processing capacity of the node.

[0026] The task allocation information is sent to the selected computing node to notify the node to start task processing. After receiving the task, the computing node starts the task execution process and allocates the corresponding computing resources according to the task requirements.

[0027] Monitor the execution of tasks in real time through the load monitor, including task progress, resource consumption, execution time, etc. Regularly update the task status in the task allocation table and record the execution progress of the task and any abnormal situations.

[0028] After the task is completed, the computing node returns the result to the cloud platform and notifies the task completion status. The task results are stored in the cloud platform database for subsequent query and analysis.

[0029] Node health assessment module: pre-processes the load data of each node, extracts features from the pre-processed load data, and assesses the health status of each node based on the extracted frequency variation characteristics of threads competing for resource locks and node power efficiency fluctuation characteristics.

[0030] Collect the load data of each node in real time from the load monitor, including CPU usage, memory usage, network traffic, disk I / O, etc. Store the collected load data in the database for subsequent processing and analysis.

[0031] Check whether there are missing values ​​in the data, and fill in the missing data (such as filling with the mean or interpolation method) or delete it. Identify and handle outliers, such as data mutations and extreme values. Statistical methods (such as the 3σ rule) or model-based methods can be used to detect and handle outliers.

[0032] Standardize data of different dimensions to make them have the same scale. Common methods include Z-score standardization or Min-Max normalization. Time-align data from different sources to ensure that data at the same time point can be compared and analyzed.

[0033] Resource contention characteristics: extract the frequency change characteristics of thread contention for resource locks, and calculate the number of resource lock contention and change trend. Power usage characteristics: extract the fluctuation characteristics of node power usage efficiency, and calculate the mean, variance and fluctuation range of power usage efficiency.

[0034] The resource contention frequency abnormality index is generated according to the extracted frequency change characteristics of the threads competing for resource locks of each node. The method for obtaining the resource contention frequency abnormality index is:

[0035] Obtain resource contention frequency data within a time period of t, and record the resource contention frequency data at N time points as R(t); select the fractional order parameter α, which determines the order of the transformation, and α usually varies between 0 and 2π.

[0036] Perform fractional Fourier transform on resource contention frequency data, and the calculation formula is: ; In the formula, The result of fractional Fourier transform of frequency data is obtained by resource competition. is the kernel function of the fractional Fourier transform. Through the fractional Fourier transform, the spectrum components of different fractional orders are obtained. , ; Perform inverse transform according to the fractional order parameter α to reconstruct the signal, and the transformation expression is: ; In the formula, is the result of fractional Fourier inverse transform, and the resource contention frequency anomaly index is calculated. The specific calculation expression is: ; In the formula, Resource contention frequency anomaly index.

[0037] The larger the resource contention frequency anomaly index is, the more unstable the node's health is. A high anomaly index indicates that the resource contention frequency on the node changes dramatically, which is usually due to frequent changes in the number or type of tasks running on the node, resulting in uneven resource allocation and use. This situation may cause the node to process tasks less efficiently and increase the risk of task delays and failures.

[0038] The sharp fluctuations in the frequency of node resource contention may also indicate potential performance bottlenecks or resource contention issues. For example, when multiple tasks compete for CPU, memory, or I / O resources at the same time, the system may not be able to effectively coordinate these contentions, resulting in reduced resource utilization and increased response time. This not only affects the performance of the node itself, but may also have a negative impact on the stability and reliability of the entire system.

[0039] In order to maintain the efficient operation and stability of the system, nodes with abnormal fluctuation index must be evaluated and optimized in a timely manner. The volatility of resource contention frequency can be reduced by dynamically adjusting task scheduling strategies, optimizing resource allocation, or increasing hardware resources, thereby improving the health of nodes and ensuring the continuous and stable operation of the system under high load and complex environments.

[0040] The power efficiency fluctuation index is generated according to the extracted power efficiency fluctuation characteristics of each node. The method for obtaining the power efficiency fluctuation index is: The power consumption value of each node is obtained in real time through the power sensor, and the obtained real-time power consumption value is used to construct a time reference sequence Q= ; and obtain the corresponding historical power usage value time target sequence C= ; Compare and analyze the time series of current sensor output values ​​and historical data time series to build a The matrix D is n, n is the length of the reference sequence, and m is the length of the target sequence; where D(i, j) represents the distance between the i-th point in Q and the j-th point in C. The similarity between the real-time power usage value and the historical power usage value is calculated. The specific calculation expression is: D(i, j) = ; Where D(i,j) is the similarity between the real-time power usage value and the historical power usage value; Construct an n×m cumulative matrix M, where M(i,j) represents the minimum cumulative distance from (1,1) to (i,j). The specific calculation expression is: ; Start from M(n,m) and backtrack to find the optimal alignment path, which represents the alignment method with the minimum cumulative distance. The steps of the backtracking path are: start from (n,m) and look forward for the minimum value: from M(i,j) to ; Repeat this step until reaching (1,1); Calculate the total cumulative distance of the optimal alignment path, that is, calculate the power efficiency fluctuation index.

[0041] The larger the power efficiency fluctuation index, the more unstable or poor the node's health is. A high fluctuation index indicates a large change in power efficiency, which may be caused by factors such as unstable power supply, severe load fluctuations, or hardware aging. This instability may cause the node to experience performance degradation, task delays, or even failures when performing tasks, thus affecting the reliability and efficiency of the entire system.

[0042] A continuously high power efficiency fluctuation index may indicate potential hardware problems or unreasonable system configuration of the node. In order to ensure efficient operation and stability of the system, further diagnosis and optimization measures must be taken on these nodes with high fluctuation indexes, such as replacing hardware, optimizing load distribution strategies, or upgrading the power management system to reduce the fluctuation index and improve the health of the node.

[0043] The resource contention frequency anomaly index and the power efficiency fluctuation index are normalized, and the health assessment coefficient of each node is calculated based on the normalized resource contention frequency anomaly index and the power efficiency fluctuation index.

[0044] For example, the present invention can use the following formula to calculate the health assessment coefficient of each node, and the calculation expression is: ; In the formula, is the health assessment coefficient, The resource contention frequency anomaly index, is the power efficiency fluctuation index, is the proportional coefficient of the resource contention frequency anomaly index and the power efficiency fluctuation index, and 0.

[0045] Node classification module: Based on the evaluation results, the health status of each node is classified into different categories, namely high-load nodes and low-load nodes.

[0046] The obtained health assessment coefficient of each node is compared with the health assessment coefficient reference threshold. If the health assessment coefficient of the node is greater than or equal to the health assessment coefficient reference threshold, it means that the node is in a high-load state, and there may be resource shortages and performance degradation. It is classified as a high-load node and a high-load node signal is generated. If the health assessment coefficient of the node is less than the health assessment coefficient reference threshold, it means that the node is in a low-load state, with sufficient resources and stable performance. It is classified as a low-load node and a low-load node signal is generated.

[0047] Dynamic adjustment module: According to the health status of the node and the priority of the task, the resource allocation strategy is dynamically adjusted through fuzzy logic to optimize the execution order of the tasks.

[0048] The health assessment coefficient and task priority are used as the input items of fuzzy logic, and the execution order of tasks is used as the output item of fuzzy logic; Define the fuzzy set of health assessment coefficient KM: low, medium, high; define the fuzzy set of task priority P: low, medium, high; define the fuzzy set of task execution order S: immediate execution, later execution, delayed execution; Fuzzy rules are established based on the health assessment coefficient and task priority: If KM is low and P is high, S is executed immediately; if KM is medium and P is high, S is executed later. If KM is high and P is high, S is executed later. If KM is low and P is medium, S is executed immediately. If KM is medium and P is medium, S is executed later. If KM is high and P is medium, S is executed later. If KM is low and P is low, S is executed later. If KM is medium and P is low, S is executed later. If KM is medium and P is low, S is executed later. If KM is high and P is low, S is executed later.

[0049] The input items KM and P are fuzzified and the membership of each fuzzy set is calculated. Fuzzy rules are applied for reasoning and the membership of each output item is calculated.

[0050] The result of fuzzy reasoning is defuzzified to obtain the specific task execution sequence value S.

[0051] Compute the defuzzified value using a weighted average: ; In the formula, is the membership value of the fuzzy set, indicating the degree to which an input item (such as the health assessment coefficient KM and the task priority P) belongs to a fuzzy set. The membership value is between 0 and 1, indicating the degree of membership of a certain input value in a fuzzy set. For example, a health assessment coefficient may have a membership of 0.2 to the "low" fuzzy set, a membership of 0.8 to the "medium" fuzzy set, and a membership of 0 to the "high" fuzzy set.

[0052] is the membership value The corresponding task execution order value. The task execution order value is also a member of the fuzzy set. For example, the task execution order can be defined as "immediate execution", "later execution" and "deferred execution". These values ​​will be given specific numerical representations (such as 0.0, 0.5, 1.0) during the defuzzification process.

[0053] The dynamic adjustment module uses fuzzy logic to define the corresponding fuzzy sets and membership functions by taking the node's health assessment coefficient and task priority as input items, and establishes fuzzy rules to fuzzify the input. The membership of each output item is obtained through fuzzy reasoning, and the weighted average method is used for defuzzification, and finally the specific task execution order is output. In this way, the system can dynamically adjust the resource allocation strategy according to the node health status and task priority, optimize the execution order of tasks, and thus improve the overall performance and stability of the system.

[0054] Fault identification module: reallocates the tasks of each computing node according to the adjusted resource allocation strategy, and further analyzes the health status of high-load nodes within a fixed time period. According to the analysis results, the faulty nodes are identified in time and repaired.

[0055] According to the adjusted resource allocation strategy, the system will reallocate the tasks of each computing node, prioritize high-priority tasks to nodes with better health, and low-priority tasks to nodes with lighter loads or average health, to ensure that high-priority tasks are processed in a timely manner, while optimizing overall resource utilization, reducing system bottlenecks and improving task processing efficiency.

[0056] Further analysis is conducted on the health status of high-load nodes within a fixed time period. When a node is a high-load node, that is, the health assessment coefficient generated within a fixed time period is greater than or equal to the health assessment coefficient reference threshold, the health assessment coefficients generated in subsequent fixed time periods that are greater than or equal to the health assessment coefficient reference threshold are collected, and a corresponding data set is established. The mean and standard deviation of the data set are calculated. After analyzing the data set, the faulty nodes are identified in a timely manner according to the analysis results, and they are repaired.

[0057] Compare the mean and standard deviation of the health assessment coefficient in the data set with the corresponding reference thresholds. If the mean of the health assessment coefficient in the data set is greater than or equal to the reference threshold of the mean of the health assessment coefficient, and the standard deviation of the health assessment coefficient is less than the reference threshold of the standard deviation of the health assessment coefficient, it means that the health of the node is continuously poor but the fluctuation is small, and there may be a stable high load situation. It is marked as a faulty node and a secondary warning signal is generated. Further monitoring is required and consideration should be given to increasing resources or optimizing the load. If the mean value of the health assessment coefficient is greater than or equal to the reference threshold of the mean value of the health assessment coefficient, and the standard deviation of the health assessment coefficient is greater than or equal to the reference threshold of the standard deviation of the health assessment coefficient, it means that the health status of the node is continuously poor and fluctuates greatly. It is marked as a faulty node and a first-level warning signal is generated. There may be unstable factors or intermittent faults. Potential faults need to be checked and repaired immediately, and tasks may need to be reallocated. If the mean value of the health assessment coefficient is less than the reference threshold value of the mean value of the health assessment coefficient, and the standard deviation of the health assessment coefficient is greater than or equal to the reference threshold value of the standard deviation of the health assessment coefficient, it means that the overall health status of the node is good but fluctuates greatly, and there may be temporary load peaks. It is marked as a faulty node and a third-level warning signal is generated. It is necessary to observe whether there are accidental events and adjust the resource scheduling strategy; If the mean of the health assessment coefficient is less than the reference threshold of the mean of the health assessment coefficient, and the standard deviation of the health assessment coefficient is less than the reference threshold of the standard deviation of the health assessment coefficient, it means that the overall health status of the node is good and stable, and it is in normal operation. It is marked as a normal node, no warning signal is generated, and no special measures are required.

[0058] It should be noted here that the management level of the first-level warning signal is higher than that of the second-level warning signal, and the management level of the second-level warning signal is higher than that of the third-level warning signal. Relevant personnel can perform corresponding node repairs based on different levels of warning signals.

[0059] In this embodiment, concurrent task requests from customer service content are received through the cloud platform, and tasks are classified into different priorities. Tasks are assigned to computing nodes according to the priorities, and the load data of each node is monitored in real time through the load monitor; the load data of each node is preprocessed, and the preprocessed load data is feature extracted. The health status of each node is evaluated according to the extracted frequency change characteristics of the threads competing for resource locks of each node and the fluctuation characteristics of the node power efficiency; according to the evaluation results, the health status of each node is divided into different categories, and it is divided into high-load nodes and low-load nodes; according to the health status of the node and the task priority, the resource allocation strategy is dynamically adjusted through fuzzy logic to optimize the execution order of the tasks; according to the adjusted resource allocation strategy, the tasks of each computing node are reallocated, and the health status of the high-load node within a fixed time period is further analyzed, and the faulty node is timely identified according to the analysis results, and repaired. Not only can concurrent task requests be processed efficiently, but also the resources of each computing node can be effectively managed and optimized to ensure the efficiency, stability and reliability of the system, and ultimately improve customer service quality and customer satisfaction.

[0060] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0061] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0062] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0063] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks or optical disks.

[0064] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A large model service platform with high-performance concurrent computing, characterized by: It includes task allocation module, node health assessment module, node division module, dynamic adjustment module and fault identification module; Task allocation module: Receives concurrent task requests from customer service content through the cloud platform, classifies tasks into different priorities, allocates tasks to computing nodes according to the priorities, and monitors the load data of each node in real time through the load monitor; Node health assessment module: pre-processes the load data of each node, extracts features from the pre-processed load data, and assesses the health status of each node based on the extracted frequency variation features of threads competing for resource locks and node power efficiency fluctuation features; Node classification module: Based on the evaluation results, the health status of each node is classified into different categories, namely high-load nodes and low-load nodes; Dynamic adjustment module: According to the health status of nodes and task priority, the resource allocation strategy is dynamically adjusted through fuzzy logic to optimize the execution order of tasks; Fault identification module: reallocates tasks of each computing node according to the adjusted resource allocation strategy, and further analyzes the health status of high-load nodes within a fixed time period. According to the analysis results, faulty nodes are identified in a timely manner and repaired. In the node health assessment module, the resource contention frequency anomaly index is generated according to the extracted frequency change characteristics of the threads competing for resource locks of each node. The method for obtaining the resource contention frequency anomaly index is: Obtain the resource contention frequency data within the time period t, and record the resource contention frequency data at N time points as R(t); select the fractional order parameter α, which determines the order of the transformation, and perform a fractional Fourier transform on the resource contention frequency data. The calculation formula is: ; In the formula, The result of fractional Fourier transform of frequency data is obtained by resource competition. is the kernel function of the fractional Fourier transform. Through the fractional Fourier transform, the spectrum components of different fractional orders are obtained. , ; Perform inverse transform according to the fractional order parameter α to reconstruct the signal, and the transformation expression is: ; In the formula, is the result of fractional Fourier inverse transform, and the resource contention frequency anomaly index is calculated. The specific calculation expression is: ; In the formula, Anomaly index of frequency of competition for resources; In the node health assessment module, a power efficiency fluctuation index is generated according to the extracted power efficiency fluctuation characteristics of each node. The method for obtaining the power efficiency fluctuation index is: The power usage value of each node is obtained in real time through the power sensor, and the obtained real-time power usage value is used to construct a time reference sequence Q= ; and obtain the corresponding historical power usage value time target sequence C= ; Compare and analyze the time series of the current sensor output value and the time series of historical data to build a The matrix D is n, n is the length of the reference sequence, and m is the length of the target sequence; where D(i, j) represents the distance between the i-th point in Q and the j-th point in C. The similarity between the real-time power usage value and the historical power usage value is calculated. The specific calculation expression is: D(i, j) = ; Where D(i,j) is the similarity between the real-time power usage value and the historical power usage value; Construct an n×m cumulative matrix M, where M(i,j) represents the minimum cumulative distance from (1,1) to (i,j). The specific calculation expression is: ; Start from M(n,m) and backtrack to find the optimal alignment path, which represents the alignment method with the minimum cumulative distance. The steps of the backtracking path are: start from (n,m) and look forward for the minimum value: from M(i,j) to ; Repeat this step until reaching (1,1); Calculate the total cumulative distance of the optimal alignment path, that is, calculate the power efficiency fluctuation index; In the node health assessment module, the resource contention frequency anomaly index and the power efficiency fluctuation index are normalized, and the health assessment coefficient of each node is calculated based on the normalized resource contention frequency anomaly index and the power efficiency fluctuation index.

2. A high-performance concurrent computing large model service platform according to claim 1, characterized in that: In the task allocation module, the task and node information are recorded in the task allocation table, which contains the task ID, the allocated node, the estimated completion time, and the allocated time. The estimated completion time of the task is estimated based on the task type and the processing capacity of the node; The task allocation information is sent to the selected computing node to notify the node to start task processing. After receiving the task, the computing node starts the task execution process and allocates corresponding computing resources according to the task requirements. The load monitor monitors the execution of the task in real time, including task progress, resource consumption, and execution time, and regularly updates the task status in the task allocation table. After the task is completed, the computing node returns the result to the cloud platform and notifies the task completion status, and stores the task result in the cloud platform's database for subsequent query and analysis.

3. A high-performance concurrent computing large model service platform according to claim 1, characterized in that: In the node division module, the health status of each node is divided into different categories according to the evaluation results, and it is divided into high-load nodes and low-load nodes; The obtained health assessment coefficient of each node is compared with the health assessment coefficient reference threshold. If the health assessment coefficient of the node is greater than or equal to the health assessment coefficient reference threshold, the node is in a high-load state, and it is classified as a high-load node, and a high-load node signal is generated; If the health assessment coefficient of the node is less than the health assessment coefficient reference threshold, the node is in a low load state, is classified as a low load node, and generates a low load node signal.

4. A high-performance concurrent computing large model service platform according to claim 1, characterized in that: In the dynamic adjustment module, according to the health status of the node and the priority of the task, the resource allocation strategy is dynamically adjusted through fuzzy logic to optimize the execution order of the tasks. Specifically: The health assessment coefficient and task priority are used as the input items of fuzzy logic, and the execution order of tasks is used as the output item of fuzzy logic; Define the fuzzy set of health assessment coefficient KM: low, medium, high; define the fuzzy set of task priority P: low, medium, high; Define the fuzzy set of task execution order S: immediate execution, later execution, and delayed execution; Establish fuzzy rules based on health assessment coefficients and task priorities; Fuzzify the input items KM and P, calculate the membership of each fuzzy set, apply fuzzy rules for reasoning, and calculate the membership of each output item; The result of fuzzy reasoning is defuzzified to obtain the specific task execution order value S, and the defuzzified value is calculated using the weighted average method.

5. A high-performance concurrent computing large model service platform according to claim 1, characterized in that: In the fault identification module, the tasks of each computing node are reallocated according to the adjusted resource allocation strategy, and the health status of the high-load nodes in a fixed time period is further analyzed, specifically: Further analysis is conducted on the health status of high-load nodes within a fixed time period. When a node is a high-load node, that is, the health assessment coefficient generated within a fixed time period is greater than or equal to the health assessment coefficient reference threshold, the health assessment coefficients generated in subsequent fixed time periods that are greater than or equal to the health assessment coefficient reference threshold are collected, and a corresponding data set is established. The mean and standard deviation of the data set are calculated. After analyzing the data set, the faulty nodes are identified in a timely manner according to the analysis results, and they are repaired.

6. A high-performance concurrent computing large model service platform according to claim 5, characterized in that: Compare the mean and standard deviation of the health assessment coefficient in the data set with the corresponding reference thresholds. If the mean of the health assessment coefficient in the data set is greater than or equal to the reference threshold of the mean of the health assessment coefficient, and the standard deviation of the health assessment coefficient is less than the reference threshold of the standard deviation of the health assessment coefficient, mark it as a faulty node and generate a secondary warning signal, requiring further monitoring and additional resources. If the mean value of the health assessment coefficient is greater than or equal to the reference threshold of the mean value of the health assessment coefficient, and the standard deviation of the health assessment coefficient is greater than or equal to the reference threshold of the standard deviation of the health assessment coefficient, it will be marked as a faulty node and a first-level warning signal will be generated. It is necessary to immediately check and repair potential faults and reallocate tasks; If the mean value of the health assessment coefficient is less than the reference threshold value of the mean value of the health assessment coefficient, and the standard deviation of the health assessment coefficient is greater than or equal to the reference threshold value of the standard deviation of the health assessment coefficient, it is marked as a faulty node and a level 3 warning signal is generated. It is necessary to observe whether there are accidental events and adjust the resource scheduling strategy; If the mean of the health assessment coefficient is less than the reference threshold of the mean of the health assessment coefficient, and the standard deviation of the health assessment coefficient is less than the reference threshold of the standard deviation of the health assessment coefficient, it is in normal operation and is marked as a normal node. No warning signal is generated and no measures are required.

Images