High-performance computing resource evaluation method based on AI technology

Abstract

The present invention discloses a high-performance computing power resource evaluation method based on AI technology, which relates to the technical field of high-performance computing power resource evaluation. It solves the technical problem in the existing technology that it is impossible to overcome the influence of all parameters when adjusting the computing power resource evaluation, thereby reducing the processing performance of the computing power resources. Specifically, the computing power resource characteristic evaluation is carried out, and multi-dimensional indicator collection and detection are performed according to the various types of hardware involved in the computing power resources; after completing the multi-dimensional indicator detection, the historical operation process of the computing power resources is detected by AI technology; computing power parameter evaluation is carried out, and weight analysis is performed according to the multi-dimensional indicators of the various types of hardware involved in the computing power resources, and real-time computing power high-performance parameters and computing power low-performance parameters are obtained according to the proportion of the influence weight; AI prediction is carried out, and the load of the computing power resources is detected, the characteristics are input, and the computing power demand range is obtained according to the historical output to perform real-time evaluation of the computing power resources.

Description

Technical Field

[0001] The present invention relates to the technical field of high-performance computing power resource evaluation, and specifically to a high-performance computing power resource evaluation method based on AI technology. Background Art

[0002] The evaluation of high-performance computing resources based on AI technology requires the integration of multi-dimensional indicators such as hardware performance, software adaptation, application scenarios and energy efficiency, and the combination of machine learning, deep learning and other technologies to achieve dynamic and accurate evaluation.

[0003] However, in the existing technology, when high-performance computing power resources are used, it is impossible to collect and detect multi-dimensional indicators, perform adaptability analysis, and adjust the indicator parameters in a timely manner. It is also impossible to monitor the impact of multi-dimensional indicators based on weight analysis. It is impossible to overcome the impact of all parameters when evaluating and adjusting computing power resources, which reduces the processing performance of computing power resources.

[0004] In view of the above technical defects, a solution is now proposed. Summary of the Invention

[0005] The purpose of this invention is to solve the above-mentioned problems and to propose a high-performance computing resource evaluation method based on AI technology.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] As a preferred embodiment of the present invention, a high-performance computing resource evaluation method based on AI technology;

[0008] Step 1: Evaluate computing resource characteristics. This involves collecting and testing multi-dimensional metrics based on the various types of hardware involved in computing resources. Hardware includes processors, GPUs, memory, and other hardware devices. Multi-dimensional metrics include computing power, storage performance, and network communication.

[0009] Inferring whether the corresponding characteristics of the computing resources built by hardware operation are efficient, so as to evaluate the characteristics of computing resources through hardware characteristic analysis. When computing resources cannot be adapted, hardware indicators can be adjusted in time to improve the computing performance of computing resources themselves.

[0010] After completing the multi-dimensional indicator detection, AI technology is used to detect the historical operation process of computing power resources. Time node statistics are collected on the fluctuation of the corresponding multi-dimensional indicators of various types of hardware. The hardware adaptation efficiency is inferred based on the time node analysis. After obtaining the adaptation efficiency analysis results, real-time fluctuation prediction is performed based on the real-time regulation of multi-dimensional indicator parameters to infer the real-time operating performance of the current computing power resources.

[0011] Step 2: Evaluate computing power parameters. Perform weight analysis based on the multi-dimensional indicators of various types of hardware involved in computing power resources, and divide them into real-time high-performance and low-performance parameters based on the impact weight ratio. Step 3: AI prediction. Detect the load of computing power resources, input features, and obtain the computing power demand range based on historical output. Perform impact detection on input features based on the historical load processing process, and perform impact analysis on real-time input features based on the impact detection results to conduct real-time evaluation of computing power resources.

[0012] As a preferred embodiment of the present invention, the multi-dimensional indicator collection and detection process is as follows:

[0013] The multi-dimensional indicators corresponding to each type of hardware involved in computing resources are computing power, storage performance, and network communication. The multi-dimensional indicators are quantified, where computing power is expressed as the numerical ratio of the execution buffer time of the hardware for data task processing to the data task processing error rate in the current time; storage performance is expressed as the numerical ratio of the actual occupancy rate of each storage space of the hardware to the delay duration of data calls in multiple storage spaces; network communication is expressed as the numerical ratio of the actual throughput of the network covered by the hardware to the actual network delay frequency. It should be explained that the numerical ratios are all the numerical ratios of two collected data, without considering the dimensionality of the collected data, and only the floating impact analysis of the data values ​​is performed;

[0014] Perform a numerical comparison of the quantified multi-dimensional indicators, that is, compare the corresponding quantitative parameters of the multi-dimensional indicators with the corresponding set parameter threshold ranges. If the corresponding quantitative parameters exceed the set parameter threshold ranges, it is inferred that the hardware of the current computing power resources has high-performance characteristics. The complexity of the actual data processing task is analyzed. If the data processing task complexity is high, the current computing power resources are set as adaptive resources; conversely, if the data processing task complexity is low, the current computing power resources are set as excess resources.

[0015] If the corresponding quantitative parameter is within the set parameter threshold range, the computing power resource is inferred to be a cost-effective resource; conversely, if the corresponding quantitative parameter is lower than the set parameter threshold range, the computing power resource is inferred to have low performance characteristics, and the hardware multi-dimensional index control of the computing power resource is performed;

[0016] If high-performance characteristics are detected, the multi-dimensional indicators are tested to be qualified. If there are excess resources, the use of computing resources will be re-planned.

[0017] As a preferred embodiment of the present invention, the historical operation detection process is as follows:

[0018] Obtain the historical operating period of the current computing power resource and collect quantitative parameters of multi-dimensional indicators within the historical operating period. Before the computing power resource executes the data processing task, the quantitative parameters at the current moment are set as input features, and the type of data processing task is marked as task features, where task features are represented by characteristic parameters such as task type, data processing volume, and data setting buffer length;

[0019] When the task characteristics are constant, record the time when the value of the input feature fluctuates and the abnormal time of the data processing task. If the floating time and the abnormal time are in the same task processing cycle, the numerical fluctuation of the quantitative parameter of the input feature is statistically analyzed. The overlapping probability of the value floating time and the task abnormal time is used as the impact level classification standard, and the impact setting is performed on the characteristic vector of the numerical fluctuation, such as the characteristic vector of the numerical fluctuation span, the duration of the numerical fluctuation or the frequency of the numerical fluctuation.

[0020] If the overlapping probability exceeds the set probability threshold, the characteristic vector that appears to float at the corresponding floating moment will be set as a high-probability influence vector. Conversely, if the overlapping probability does not exceed the set probability threshold, the characteristic vector that appears to float at the corresponding floating moment will be set as a low-probability influence vector. Real-time fluctuation analysis will be performed based on each type of characteristic vector and the corresponding vector type.

[0021] As a preferred embodiment of the present invention, the real-time fluctuation analysis process is as follows:

[0022] Compare the task characteristics of the current task with those of tasks in each historical period. Tasks with task characteristic parameters in the same range are regarded as tasks of the same type. Collect the floating quantitative parameter feature vectors that appear when the same type of tasks are executed, and identify the feature vector type according to the historical operating period of the same type of tasks. If it is a high-probability impact vector, issue a data processing task anomaly warning and use the historical data processing task anomaly type as the current warning object, and adjust the feature vector according to the warning object. If it is a low-probability impact vector, identify the floating trend of the feature vector in the historical operating period, monitor the floating trend of the current feature vector of the same type, and make timely adjustments when the trend is synchronized.

[0023] As a preferred embodiment of the present invention, the computing power parameter evaluation process in step 2 is as follows:

[0024] Analyze the multi-dimensional indicators of various types of hardware involved in computing power resources, conduct numerical floating trend statistics on the quantitative parameters of the multi-dimensional indicators, and analyze the floating trend of the quantitative parameters when the computing power resources are executing data processing tasks. When the floating trend of the quantitative parameters has not stabilized, that is, there is a floating trend and it fluctuates back and forth, obtain the processing speed drop frequency and the speed recovery buffer time after the processing speed drops for the data processing tasks in the current stage. If any of the processing speed drop frequency and the speed recovery buffer time after the processing speed drops exceeds the corresponding set threshold, it is inferred that the current quantitative parameter is a high-risk parameter; conversely, if no data exists in the processing speed drop frequency and the speed recovery buffer time after the processing speed drops that exceeds the corresponding set threshold, it is inferred that the current quantitative parameter is a low-risk parameter.

[0025] The influence weight of high-risk parameters is set to a high influence weight, and the weight factor is set to a low numerical factor to avoid the influence of high-risk parameters being infinitely amplified relative to the influence of low-risk parameters, so that the influence of low-risk parameters is ignored. It should be explained that the value of the weight factor is the difference between the low numerical factor and the high numerical factor; the influence weight of high-risk parameters is set to a low influence weight, and the weight factor is set to a high numerical factor;

[0026] The quantitative parameters of multi-dimensional indicators are divided into types, that is, high-impact weights and corresponding weight factors are uniformly marked as high-performance computing power parameters, and low-impact weights and corresponding weight factors are uniformly marked as low-performance computing power parameters.

[0027] As a preferred embodiment of the present invention, the AI ​​prediction process is as follows:

[0028] Collect multi-dimensional indicators and corresponding quantitative parameters of current computing resources, monitor quantitative parameters during the execution phase of real-time data processing tasks, set up a resource load analysis model, record the floating feature vectors of quantitative parameters during the execution phase, and mark them as model startup features. When model startup features are generated, evaluate computing resource utilization based on the type of model startup features, i.e., high-probability impact features and low-probability impact features.

[0029] Collect high-risk parameters and low-risk parameters and corresponding weight factors in the quantitative parameters, and mark the high-risk parameters and low-risk parameters as C respectively. 高 and C 低 ; The weight factors are Q 高 and Q 低 ; Through the formula The influence coefficient S of the high-probability impact feature is obtained. It should be explained that the values ​​of the high-risk parameter and the low-risk parameter are the values ​​of the corresponding feature vectors, while the weight factor is artificially set through continuous analysis of historical periods. At the same time, when the units of the feature vectors are not uniform, the weight factor can be used to directly de-dimensionalize them.

[0030] If the impact coefficient S exceeds the coefficient threshold and continues to increase, it indicates that the utilization rate of the current multi-dimensional indicators of the computing power resources continues to decrease. The data processing task anomaly type is obtained based on the execution process of the same type of tasks in the historical operating period. After identifying the current processing task anomaly type, the corresponding feature vector control of the data processing task anomaly type is obtained, and the feature vector control range in the historical process is marked as the computing power demand range corresponding to the quantitative parameters of the current computing power resources. The current multi-dimensional indicators are controlled according to the computing power demand range. Conversely, if the impact coefficient S does not exceed the coefficient threshold and remains constant, it indicates that the utilization rate of the current multi-dimensional indicators of the computing power resources is constant.

[0031] When the model startup feature is a high-probability impact feature, the maximum floating span of high-risk parameters and the number of parameter types that fluctuate in low-risk parameters are collected. If any of these parameters increases, the computing power requirement range needs to be expanded based on the regulation trend. Conversely, if no parameter increases, regulation is performed within the current computing power requirement range.

[0032] When the model startup feature is a low-probability impact feature, the numerical increase span of the high-risk parameter floating frequency and the cumulative span of the continuous floating of the low-risk parameter are collected. If any parameter of the numerical increase span of the high-risk parameter floating frequency and the cumulative span of the continuous floating of the low-risk parameter exceeds the set numerical threshold, the current computing power demand range is used as the control peak, and the computing power demand range is controlled in stages and steps; if any parameter of the numerical increase span of the high-risk parameter floating frequency and the cumulative span of the continuous floating of the low-risk parameter does not exceed the set numerical threshold, the current computing power demand range is narrowed to reduce the control span of the characteristic vector; it should be explained that the premise of the control of the computing power demand range is that the fluctuation of the quantitative parameter characteristic vector does not reach the value of the characteristic vector when the data processing task is abnormal.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. In the present invention, multi-dimensional indicator collection and detection are performed based on various types of hardware involved in computing resources;

[0035] Infer whether the corresponding characteristics of the computing power resources constructed by hardware operation are efficient, so as to evaluate the characteristics of computing power resources through hardware characteristic analysis, and timely adjust hardware indicators when computing power resources cannot be adapted to improve the computing power performance of computing power resources themselves; infer the hardware adaptation efficiency based on time node analysis, and after obtaining the adaptation efficiency analysis results, perform real-time fluctuation prediction based on the regulation of real-time multi-dimensional indicator parameters to infer the real-time operation performance of current computing power resources.

[0036] 2. In the present invention, a weight analysis is performed based on the multi-dimensional indicators of various types of hardware involved in computing power resources, and the real-time computing power high-performance parameters and computing power low-performance parameters are obtained according to the proportion of the influence weights; the parameter influence weights can be effectively set by dividing the performance parameters to ensure that the influence of each parameter can be collected and counted in real time, to avoid the influence of different parameters being monitored when computing power resources are used, and to avoid the large deviation in the corresponding influence of the two parameters causing the low-impact parameter value to lose its influence, so that when the computing power demand range is controlled, there is a control deviation due to the lack of the influencing parameter, and the deviation cannot be reduced to a minimum. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] To facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0038] Figure 1 A flow chart of the overall method of the present invention;

[0039] Figure 2 This is a flow chart of the method for collecting and detecting multi-dimensional indicators in the present invention. DETAILED DESCRIPTION

[0040] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of 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. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0041] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute a separate or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0042] See also Figure 1 As shown in the figure, a high-performance computing resource evaluation method based on AI technology;

[0043] Step 1: Evaluate computing resource characteristics. This involves collecting and testing multi-dimensional metrics based on the various types of hardware involved in computing resources. Hardware includes processors, GPUs, memory, and other hardware devices. Multi-dimensional metrics include computing power, storage performance, and network communication.

[0044] Inferring whether the corresponding characteristics of the computing resources built by hardware operation are efficient, so as to evaluate the characteristics of computing resources through hardware characteristic analysis. When computing resources cannot be adapted, hardware indicators can be adjusted in time to improve the computing performance of computing resources themselves.

[0045] After completing the multi-dimensional indicator detection, AI technology is used to detect the historical operation process of computing power resources. Time node statistics are collected on the fluctuation of the corresponding multi-dimensional indicators of various types of hardware. The hardware adaptation efficiency is inferred based on the time node analysis. After obtaining the adaptation efficiency analysis results, real-time fluctuation prediction is performed based on the real-time regulation of multi-dimensional indicator parameters to infer the real-time operating performance of the current computing power resources.

[0046] Step 2: Evaluate computing power parameters. Perform a weighted analysis based on the multi-dimensional indicators of various types of hardware involved in computing power resources, and divide them into high-performance and low-performance parameters based on the influence weight ratio.

[0047] Step 3: AI prediction: Detect the load of computing resources, input features, and obtain the computing power demand range based on historical output. Perform impact detection on the input features based on the historical load processing process. Based on the impact detection results, perform impact analysis on the real-time input features to conduct real-time evaluation of computing power resources.

[0048] See also Figure 2 As shown, the multi-dimensional indicator collection and detection process is as follows:

[0049] The multi-dimensional indicators corresponding to each type of hardware involved in computing resources are computing power, storage performance, and network communication. The multi-dimensional indicators are quantified, where computing power is expressed as the numerical ratio of the execution buffer time of the hardware for data task processing to the data task processing error rate in the current time; storage performance is expressed as the numerical ratio of the actual occupancy rate of each storage space of the hardware to the delay duration of data calls in multiple storage spaces; network communication is expressed as the numerical ratio of the actual throughput of the network covered by the hardware to the actual network delay frequency. It should be explained that the numerical ratios are all the numerical ratios of two collected data, without considering the dimensionality of the collected data, and only the floating impact analysis of the data values ​​is performed;

[0050] Perform a numerical comparison of the quantified multi-dimensional indicators, that is, compare the corresponding quantitative parameters of the multi-dimensional indicators with the corresponding set parameter threshold ranges. If the corresponding quantitative parameters exceed the set parameter threshold ranges, it is inferred that the hardware of the current computing power resources has high-performance characteristics. The complexity of the actual data processing task is analyzed. If the data processing task complexity is high, the current computing power resources are set as adaptive resources; conversely, if the data processing task complexity is low, the current computing power resources are set as excess resources.

[0051] If the corresponding quantitative parameter is within the set parameter threshold range, the computing power resource is inferred to be a cost-effective resource; conversely, if the corresponding quantitative parameter is lower than the set parameter threshold range, the computing power resource is inferred to have low performance characteristics, and the hardware multi-dimensional index control of the computing power resource is performed;

[0052] If high-performance characteristics are detected, the multi-dimensional indicators are tested to be qualified. If there are excess resources, the use of computing resources will be re-planned.

[0053] Furthermore, the historical operation detection process is as follows:

[0054] Obtain the historical operating period of the current computing power resource and collect quantitative parameters of multi-dimensional indicators within the historical operating period. Before the computing power resource executes the data processing task, the quantitative parameters at the current moment are set as input features, and the type of data processing task is marked as task features, where task features are represented by characteristic parameters such as task type, data processing volume, and data setting buffer length;

[0055] When the task characteristics are constant, record the time when the value of the input feature fluctuates and the abnormal time of the data processing task. If the floating time and the abnormal time are in the same task processing cycle, the numerical fluctuation of the quantitative parameter of the input feature is statistically analyzed. The overlapping probability of the value floating time and the task abnormal time is used as the impact level classification standard, and the impact setting is performed on the characteristic vector of the numerical fluctuation, such as the characteristic vector of the numerical fluctuation span, the duration of the numerical fluctuation or the frequency of the numerical fluctuation.

[0056] If the overlapping probability exceeds the set probability threshold, the characteristic vector that appears to float at the corresponding floating moment will be set as a high-probability influence vector. Conversely, if the overlapping probability does not exceed the set probability threshold, the characteristic vector that appears to float at the corresponding floating moment will be set as a low-probability influence vector. Real-time fluctuation analysis will be performed based on each type of characteristic vector and the corresponding vector type.

[0057] Furthermore, the real-time fluctuation analysis process is as follows:

[0058] Compare the task characteristics of the current task with those of tasks in each historical period. Tasks with task characteristic parameters in the same range are regarded as tasks of the same type. Collect the floating quantitative parameter feature vectors that appear when the same type of tasks are executed, and identify the feature vector type according to the historical operating period of the same type of tasks. If it is a high-probability impact vector, issue a data processing task anomaly warning and use the historical data processing task anomaly type as the current warning object, and adjust the feature vector according to the warning object. If it is a low-probability impact vector, identify the floating trend of the feature vector in the historical operating period, monitor the floating trend of the current feature vector of the same type, and make timely adjustments when the trend is synchronized.

[0059] Furthermore, the second step of computing power parameter evaluation process is as follows:

[0060] Analyze the multi-dimensional indicators of various types of hardware involved in computing power resources, conduct numerical floating trend statistics on the quantitative parameters of the multi-dimensional indicators, and analyze the floating trend of the quantitative parameters when the computing power resources are executing data processing tasks. When the floating trend of the quantitative parameters has not stabilized, that is, there is a floating trend and it fluctuates back and forth, obtain the processing speed drop frequency and the speed recovery buffer time after the processing speed drops for the data processing tasks in the current stage. If any of the processing speed drop frequency and the speed recovery buffer time after the processing speed drops exceeds the corresponding set threshold, it is inferred that the current quantitative parameter is a high-risk parameter; conversely, if no data exists in the processing speed drop frequency and the speed recovery buffer time after the processing speed drops that exceeds the corresponding set threshold, it is inferred that the current quantitative parameter is a low-risk parameter.

[0061] The influence weight of high-risk parameters is set to a high influence weight, and the weight factor is set to a low numerical factor to avoid the influence of high-risk parameters being infinitely amplified relative to the influence of low-risk parameters, so that the influence of low-risk parameters is ignored. It should be explained that the value of the weight factor is the difference between the low numerical factor and the high numerical factor; the influence weight of high-risk parameters is set to a low influence weight, and the weight factor is set to a high numerical factor;

[0062] The quantitative parameters of multi-dimensional indicators are divided into types, that is, high-impact weights and corresponding weight factors are uniformly marked as high-performance computing power parameters, and low-impact weights and corresponding weight factors are uniformly marked as low-performance computing power parameters.

[0063] Furthermore, the AI ​​prediction process is as follows:

[0064] Collect multi-dimensional indicators and corresponding quantitative parameters of current computing resources, monitor quantitative parameters during the execution phase of real-time data processing tasks, set up a resource load analysis model, record the floating feature vectors of quantitative parameters during the execution phase, and mark them as model startup features. When model startup features are generated, evaluate computing resource utilization based on the type of model startup features, i.e., high-probability impact features and low-probability impact features.

[0065] Collect high-risk parameters and low-risk parameters and corresponding weight factors in the quantitative parameters, and mark the high-risk parameters and low-risk parameters as C respectively. 高 and C 低 ; The weight factors are Q 高 and Q 低 ; Through the formula The influence coefficient S of the high-probability impact feature is obtained. It should be explained that the values ​​of the high-risk parameter and the low-risk parameter are the values ​​of the corresponding feature vectors, while the weight factor is artificially set through continuous analysis of historical periods. At the same time, when the units of the feature vectors are not uniform, the weight factor can be used to directly de-dimensionalize them.

[0066] If the impact coefficient S exceeds the coefficient threshold and continues to increase, it indicates that the utilization rate of the current multi-dimensional indicators of the computing power resources continues to decrease. The data processing task anomaly type is obtained based on the execution process of the same type of tasks in the historical operating period. After identifying the current processing task anomaly type, the corresponding feature vector control of the data processing task anomaly type is obtained, and the feature vector control range in the historical process is marked as the computing power demand range corresponding to the quantitative parameters of the current computing power resources. The current multi-dimensional indicators are controlled according to the computing power demand range. Conversely, if the impact coefficient S does not exceed the coefficient threshold and remains constant, it indicates that the utilization rate of the current multi-dimensional indicators of the computing power resources is constant.

[0067] When the model startup feature is a high-probability impact feature, the maximum floating span of high-risk parameters and the number of parameter types that fluctuate in low-risk parameters are collected. If any of these parameters increases, the computing power requirement range needs to be expanded based on the regulation trend. Conversely, if no parameter increases, regulation is performed within the current computing power requirement range.

[0068] When the model startup feature is a low-probability impact feature, the numerical increase span of the high-risk parameter floating frequency and the cumulative span of the continuous floating of the low-risk parameter are collected. If any parameter of the numerical increase span of the high-risk parameter floating frequency and the cumulative span of the continuous floating of the low-risk parameter exceeds the set numerical threshold, the current computing power demand range is used as the control peak, and the computing power demand range is controlled in stages and steps; if any parameter of the numerical increase span of the high-risk parameter floating frequency and the cumulative span of the continuous floating of the low-risk parameter does not exceed the set numerical threshold, the current computing power demand range is narrowed to reduce the control span of the characteristic vector; it should be explained that the premise of the control of the computing power demand range is that the fluctuation of the quantitative parameter characteristic vector does not reach the value of the characteristic vector when the data processing task is abnormal.

[0069] When the present invention is in use, the characteristics of computing power resources are evaluated, and multi-dimensional indicators are collected and detected based on the various types of hardware involved in the computing power resources; after completing the multi-dimensional indicator detection, the historical operation process of the computing power resources is detected through AI technology; computing power parameters are evaluated, and weight analysis is performed based on the multi-dimensional indicators of the various types of hardware involved in the computing power resources, and real-time computing power high-performance parameters and computing power low-performance parameters are obtained according to the proportion of the impact weights; AI prediction is performed, and the load of the computing power resources is detected, the characteristics are input, and the computing power demand range is obtained based on the historical output to perform real-time evaluation of the computing power resources.

[0070] Thresholds, preset values, and preset ranges are set for comparative analysis of results to determine whether they are good or bad. The values ​​are set based on a combination of large-scale model analysis of sample data and manual experience, and can also be adjusted appropriately based on seasonal or common-sense factors.

[0071] The settings of weight ratio coefficients, influencing factors, etc. are assigned specific values ​​according to the influence of each parameter on the result, which ultimately reflects the impact on the result. They are also set and entered into storage through a combination of large-scale model analysis of sample data and manual experience. Appropriate adjustments can also be made based on seasonal or common-sense influencing conditions.

[0072] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A high-performance computing resource evaluation method based on AI technology, characterized by: The evaluation method is as follows: Step 1: Evaluate the characteristics of computing resources. Collect and test multi-dimensional indicators based on the various types of hardware involved in the computing resources. After completing the multi-dimensional indicator testing, use AI technology to test the historical operation process of the computing resources. Perform time node statistics on the fluctuation of the corresponding multi-dimensional indicators of each type of hardware. Infer the hardware adaptation efficiency based on the time node analysis. After obtaining the adaptation efficiency analysis results, perform real-time fluctuation prediction based on the real-time adjustment of the multi-dimensional indicator parameters. The historical operation detection process is as follows: Obtain the historical operating period of the current computing power resource and collect the quantitative parameters of multi-dimensional indicators in the historical operating period. Before the computing power resource executes the data processing task, the quantitative parameters at the current moment are set as input features, and the type of data processing task is marked as the task feature; When the task characteristics are constant, record the floating moments of the input characteristics and the abnormal moments of the data processing tasks. If the floating moments and the abnormal moments are in the same task processing cycle, the floating values ​​of the quantized parameters of the input characteristics are statistically analyzed. The overlapping probability of the floating moments and the abnormal moments of the tasks is used as the criterion for the impact level, and the impact of the floating feature vector is set. If the overlap probability exceeds the set probability threshold, the feature vector corresponding to the floating moment will be set as a high-probability influence vector. Conversely, if the overlap probability does not exceed the set probability threshold, the feature vector corresponding to the floating moment will be set as a low-probability influence vector. Real-time fluctuation analysis will be performed based on each type of feature vector and the corresponding vector type. The real-time fluctuation analysis process is as follows: Compare the task characteristics of the current task with those of tasks in each historical period. Tasks with the same range of task characteristic parameters are considered to be of the same type. Collect the floating quantitative parameter feature vectors when executing tasks of the same type, and identify the feature vector type based on the historical running time of tasks of the same type. If the feature vector is a high-probability impact vector, issue a data processing task anomaly warning and use the historical data processing task anomaly type as the current warning target. Adjust the feature vector based on the warning target. If it is a low-probability impact vector, the floating trend of the feature vector in the historical operation period is identified, and the floating trend of the current feature vector of the same type is monitored and timely regulation is carried out when the trend is synchronized. Step 2: Evaluate computing power parameters. Perform a weighted analysis based on the multi-dimensional indicators of various types of hardware involved in computing power resources, and divide them into high-performance and low-performance parameters based on the influence weight ratio. The evaluation process of computing power parameters in step 2 is as follows: Perform numerical floating trend statistics on the quantitative parameters of multi-dimensional indicators and analyze the floating trend of the quantitative parameters. When the floating trend of the quantitative parameters has not stabilized, obtain the frequency of processing speed reduction of the data processing task in the current stage and the speed recovery buffer length after the processing speed reduction. If there is any data with a value exceeding the corresponding set threshold, it is inferred that the current quantitative parameter is a high-risk parameter; conversely, if there is no data with a value exceeding the corresponding set threshold, it is inferred that the current quantitative parameter is a low-risk parameter. The impact weight of high-risk parameters is set to high impact weight, and the weight factor is set to a low numerical factor; the impact weight of high-risk parameters is set to low impact weight, and the weight factor is set to a high numerical factor; the quantitative parameters of multi-dimensional indicators are classified into types, that is, high impact weights and corresponding weight factors are uniformly marked as high-performance computing power parameters, and low impact weights and corresponding weight factors are uniformly marked as low-performance computing power parameters; Step 3: AI prediction: Detect the load of computing resources, input features, and obtain the computing power demand range based on historical output to perform real-time evaluation of computing power resources. The AI ​​prediction process is as follows: Collect multi-dimensional indicators and corresponding quantitative parameters of current computing resources, set up a resource load analysis model, record the floating feature vectors of the quantitative parameters during the execution phase, and mark them as model startup features. When model startup features are generated, evaluate computing resource utilization based on the type of model startup features. Collect high-risk parameters and low-risk parameters and corresponding weight factors in the quantitative parameters, and mark the high-risk parameters and low-risk parameters as C respectively. 高 and C 低 ; The weight factors are Q 高 and Q 低 ; Through the formula Get the influence coefficient S of the high probability influence feature; If the impact coefficient S exceeds the coefficient threshold and continues to increase, it indicates that the utilization rate of the current multi-dimensional indicators of the computing power resources continues to decrease. The data processing task anomaly type is obtained based on the execution process of the same type of tasks in the historical running period. After identifying the current processing task anomaly type, the corresponding feature vector control of the data processing task anomaly type is obtained. The feature vector control range in the historical process is marked as the computing power demand range corresponding to the quantitative parameters of the current computing power resources. The current multi-dimensional indicators are regulated according to the computing power demand range. On the contrary, if the impact coefficient S does not exceed the coefficient threshold and is constant, it indicates that the utilization rate corresponding to the current multi-dimensional indicators of the computing power resources is constant.

2. The high-performance computing resource evaluation method based on AI technology according to claim 1 is characterized in that: The multi-dimensional indicator collection and detection process is as follows: The multi-dimensional indicators corresponding to each type of hardware involved in computing power resources are computing power, storage performance, and network communication; and the multi-dimensional indicators are quantified, among which computing power is expressed as the numerical ratio of the execution buffer time of the hardware for data task processing to the data task processing error rate in the current time; storage performance is expressed as the numerical ratio of the actual occupancy rate of each storage space of the hardware to the delay time of data calls in multiple storage spaces; network communication is expressed as the numerical ratio of the actual throughput of the network covered by the hardware to the actual network delay frequency.

3. The high-performance computing resource evaluation method based on AI technology according to claim 2 is characterized in that: The quantified multi-dimensional indicators are numerically compared. If the corresponding quantified parameters exceed the set parameter threshold range, it is inferred that the hardware of the current computing power resources has high-performance characteristics. The complexity of the actual data processing task is analyzed. If the data processing task complexity is high, the current computing power resources are set as adaptive resources; conversely, if the data processing task complexity is low, the current computing power resources are set as excess resources. If the corresponding quantitative parameter is within the set parameter threshold range, it is inferred that the computing power resource is a cost-effective resource; conversely, if the corresponding quantitative parameter is lower than the set parameter threshold range, it is inferred that the computing power resource has low performance characteristics, and the hardware multi-dimensional indicator control of the computing power resource is performed.

4. The high-performance computing resource evaluation method based on AI technology according to claim 1 is characterized in that: When the model startup feature is a high-probability impact feature, the maximum floating span of the high-risk parameter and the number of parameter types that fluctuate in the low-risk parameter are collected. If any of the parameters increases, the computing power demand range needs to be expanded based on the regulation trend; conversely, if no parameter increases, regulation is performed within the current computing power demand range.

5. The high-performance computing resource evaluation method based on AI technology according to claim 4 is characterized in that: When the model startup feature is a low-probability impact feature, the numerical increase span of the high-risk parameter floating frequency and the cumulative span of the low-risk parameter continuous floating are collected. If any parameter of the numerical increase span of the high-risk parameter floating frequency and the cumulative span of the low-risk parameter continuous floating exceeds the set numerical threshold, the current computing power demand range is used as the control peak, and the computing power demand range is controlled in stages and steps; if any parameter of the numerical increase span of the high-risk parameter floating frequency and the cumulative span of the low-risk parameter continuous floating does not exceed the set numerical threshold, the current computing power demand range is narrowed to reduce the control span of the characteristic vector.

Images