Performance analysis method of high-performance computing platform, electronic equipment and storage medium

Abstract

The invention discloses a performance analysis method of a high-performance computing platform, electronic equipment and a storage medium, and relates to the technical field of high-performance computing, and the performance analysis method comprises the steps that information of multiple first parameters is acquired, and the first parameters are associated with the performance of the high-performance computing platform; in the information of the multiple first parameters, target information of the multiple first parameters meeting a preset rule is obtained, and the preset rule is used for indicating abnormal judgment conditions of the multiple first parameters; according to the type of a first parameter corresponding to the target information, performing aggregation processing on the target information to obtain first description information describing a change state of the first parameter; and analyzing the performance of the high-performance computing platform according to the plurality of pieces of first description information corresponding to the plurality of first parameters. Through the method, the performance of the high-performance computing platform can be automatically analyzed based on the information of the multiple first parameters and the preset rule, the dependence on artificial experience is reduced, and the analysis efficiency of performance analysis of the HPC platform is improved.

Description

Technical Field

[0001] The present application relates to the field of high-performance computing technology, and in particular to a performance analysis method, electronic device, and storage medium for a high-performance computing platform. Background Art

[0002] In the performance analysis of high-performance computing (HPC) platforms, HPC monitoring data often includes data from thousands of nodes, and the data for each node contains dozens of performance indicators.

[0003] In related technologies, technicians need to have in-depth knowledge in the field of HPC technology. Based on manual experience, they spend a lot of time sorting out the changing trends of performance indicators from monitoring data, resulting in low efficiency in manual performance analysis of HPC platforms. Summary of the Invention

[0004] The present application provides a performance analysis method, electronic device, and storage medium for a high-performance computing platform, to at least solve the problem of low analysis efficiency of performance analysis of HPC platforms in related technologies.

[0005] This application provides a performance analysis method for a high-performance computing platform, including:

[0006] Obtaining information of a plurality of first parameters, the first parameters being associated with performance of the high-performance computing platform;

[0007] Acquire target information of the plurality of first parameters that satisfies preset rules from among information of the plurality of first parameters, the preset rules being used to indicate abnormality judgment conditions of the plurality of first parameters;

[0008] Aggregating the target information according to the type of the first parameter corresponding to the target information to obtain first description information describing a change state of the first parameter;

[0009] The performance of the high-performance computing platform is analyzed according to the plurality of first description information corresponding to the plurality of first parameters.

[0010] The present application also provides a performance analysis device for a high-performance computing platform, comprising:

[0011] A first acquisition module, configured to acquire information of a plurality of first parameters, where the first parameters are associated with the performance of the high-performance computing platform;

[0012] A second acquisition module is configured to acquire target information of multiple first parameters that meet preset rules from information of multiple first parameters, where the preset rules are used to indicate abnormality judgment conditions of the various first parameters;

[0013] A first processing module is configured to aggregate the target information according to the type of the first parameter corresponding to the target information to obtain first description information describing a change state of the first parameter;

[0014] The second processing module is used to analyze the performance of the high-performance computing platform according to the multiple first description information corresponding to the multiple first parameters.

[0015] The present application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for implementing the steps of any of the above-mentioned performance analysis methods for high-performance computing platforms when executing the computer program.

[0016] The present application also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the steps of the performance analysis method of any of the above-mentioned high-performance computing platforms are implemented.

[0017] The present application also provides a computer program product, including a computer program, which implements the steps of any of the above-mentioned performance analysis methods for high-performance computing platforms when executed by a processor.

[0018] Through this application, since it is possible to automatically screen and aggregate abnormal information of the first parameter related to the performance of the high-performance computing platform based on preset rules, automated analysis of the performance of the high-performance computing platform can be achieved, which can solve the technical problem of low analysis efficiency caused by manual analysis relying on manual experience, and achieve the technical effect of improving the analysis efficiency of the performance analysis of the high-performance computing platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0020] Figure 1 A schematic diagram of the architecture of a performance analysis system for a high-performance computing platform provided in an embodiment of the present application;

[0021] Figure 2 A schematic diagram of a flow chart of a performance analysis method for a high-performance computing platform provided in an embodiment of the present application;

[0022] Figure 3 A schematic diagram of information of multiple first parameters provided in an embodiment of the present application;

[0023] Figure 4 A schematic diagram of a first description information process provided in an embodiment of the present application;

[0024] Figure 5 A schematic diagram of the aggregation process provided in an embodiment of the present application;

[0025] Figure 6 A schematic diagram of a process for analyzing the performance of a high-performance computing platform provided in an embodiment of the present application;

[0026] Figure 7 A schematic diagram of the structure of the performance analysis method of the high-performance computing platform provided in an embodiment of the present application;

[0027] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0029] It should be noted that, in the description of this application, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. The terms "first," "second," etc., in this application are used to distinguish similar objects, and are not used to describe a particular order or sequence.

[0030] Glossary:

[0031] Pre-training: Pre-training is a deep learning model training strategy that focuses on using large-scale datasets to initially train the model, enabling it to learn common feature representations. This process is similar to the basic learning stage humans undergo before acquiring new knowledge, accumulating experience through extensive reading and observation.

[0032] A pre-trained language model generally refers to a language model training task designed based on a large-scale corpus (including language training materials such as sentences and paragraphs). A large-scale neural network algorithm structure is trained to learn and implement the model. The resulting large-scale neural network algorithm structure and parameters are the pre-trained language model. Subsequent tasks can use this model to extract features or fine-tune the model to achieve specific task objectives. The idea behind pre-training is to first train a task to obtain a set of model parameters. This set of model parameters is then used to initialize the network model parameters. The initialized network model is then used to train other tasks to obtain a model adapted for the other tasks. By pre-training on a large-scale corpus, neural language representation models can acquire powerful language representation capabilities and extract rich syntactic and semantic information from text. Pre-trained language models can provide token- and sentence-level features containing rich semantic information for use in downstream tasks. Fine-tuning can also be performed directly on the pre-trained model for downstream tasks, quickly and easily obtaining a dedicated downstream model.

[0033] The neural network algorithm structure used to train the pre-trained language model can be a convolutional neural network (CNN), a recurrent neural network (RNN), a long short-term memory network (LSTM), etc., or a model built with an attention network, which is not limited in this application. An attention network refers to a network model that uses the attention mechanism for training. The model assigns different weights to each part of the input sequence, thereby extracting more important feature information from the input sequence, allowing the model to ultimately obtain more accurate output.

[0034] Model fine-tuning: Model fine-tuning involves further training a pre-trained language model on a task-specific dataset to adjust model parameters and better adapt it to the target task. During fine-tuning, most layers of the pre-trained language model are typically frozen, with only newly added layers trained or a small number of key layers adjusted. This approach preserves the useful features learned by the pre-trained language model while enabling rapid adaptation to the specific requirements of the new task. Choosing an appropriate learning rate and number of training rounds is also crucial for successful fine-tuning.

[0035] In the performance analysis of high-performance computing (HPC) platforms, HPC monitoring data often includes data from thousands of nodes, and the data for each node contains dozens of performance indicators.

[0036] In related technologies, technicians need to have in-depth knowledge in the field of HPC technology. Based on manual experience, they spend a lot of time sorting out the changing trends of performance indicators from monitoring data, resulting in low efficiency in manual performance analysis of HPC platforms.

[0037] In response to the above problems, in an embodiment of the present application, information on multiple first parameters is obtained, where the first parameters are associated with the performance of a high-performance computing platform; from the information on multiple first parameters, target information on multiple first parameters that meet preset rules is obtained, where the preset rules are used to indicate abnormal judgment conditions for various first parameters; based on the type of the first parameter corresponding to the target information, the target information is aggregated to obtain first descriptive information describing the change state of the first parameter; and based on the multiple first descriptive information corresponding to the multiple first parameters, the performance of the high-performance computing platform is analyzed. Through the above method, based on the information on multiple first parameters and preset rules, the performance of the high-performance computing platform can be automatically analyzed, reducing dependence on manual experience and improving the analysis efficiency of the performance analysis of the HPC platform.

[0038] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0039] In conjunction with the specific application environment architecture or specific hardware architecture that the performance analysis method of the high-performance computing platform relies on, the specific application environment architecture or specific hardware architecture is described here. Figure 1 , Figure 1 This is a schematic diagram of the architecture of the performance analysis system of the high-performance computing platform provided in the embodiment of this application. Figure 1 The performance analysis system for a high-performance computing platform includes an event capturer, an event aggregator, and a parsing module. After obtaining information about multiple first parameters, the event capturer can obtain target information about multiple first parameters that meet preset rules from the information about the multiple first parameters. The event aggregator aggregates the target information based on the type of the first parameter corresponding to the target information to obtain first description information describing the change state of the first parameter. The parsing module then analyzes the performance of the high-performance computing platform based on the multiple first description information corresponding to the multiple first parameters.

[0040] The modular architecture of event capture, event aggregator, and analysis modules enables independent deployment and flexible combination of functional units. Each module can independently process performance data from different sources without relying on specific collection tools or platforms. This design enables integration with the performance analysis system of any HPC platform with an open data interface in related technologies, thereby obtaining information on a variety of primary parameters. This allows for easy adaptation to various HPC platform application scenarios such as meteorological simulation, molecular dynamics, and fluid dynamics, meeting the diverse analysis needs of HPC platforms and demonstrating strong scalability and compatibility.

[0041] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0042] Figure 2 A flow chart of a performance analysis method for a high-performance computing platform provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the embodiment of the present application provides a performance analysis method for a high-performance computing platform, and the method is described in detail as follows:

[0043] S201. Obtain information of a plurality of first parameters, wherein the first parameters are associated with the performance of a high-performance computing platform.

[0044] In an embodiment of the present application, the information on multiple first parameters may be information on multiple first parameters of a single node or a cluster. The performance analysis method of a high-performance computing platform based on the information on multiple first parameters of multiple nodes or multiple clusters is similar to the performance analysis method of a high-performance computing platform based on the information on multiple first parameters of a single node or a cluster, and will not be repeated here.

[0045] Among them, the node is an independent computing unit in the high-performance computing platform and is the basic component of the cluster. The cluster is a computing cluster formed by multiple nodes interconnected by a network. The nodes work together through a high-speed network (such as an infinite bandwidth technology network) and can jointly undertake large-scale computing tasks. The overall performance not only depends on the performance of a single node, but also on the communication efficiency between nodes, resource scheduling capabilities, etc.

[0046] Optionally, the first parameter includes, but is not limited to, memory bandwidth, cache hit rate, InfiniBand (IB) bandwidth, central processing unit (CPU) utilization, graphics processing unit (GPU) utilization, and message passing interface (MPI) communication waiting time.

[0047] Optionally, the information of the various first parameters is obtained by sampling or instrumentation. Exemplarily, the information of the various first parameters is obtained by periodic sampling (eg, once per second).

[0048] The information about the first parameter may be specific values ​​of the first parameter obtained by downsampling at multiple sampling moments.

[0049] The first parameter is associated with the performance of the high-performance computing platform. It can be understood that the first parameter is a performance indicator of the high-performance computing platform. It can also be understood that the information of the first parameter can directly or indirectly reflect the performance and status of the high-performance computing platform during operation.

[0050] S202. From information on a plurality of first parameters, obtain target information on a plurality of first parameters that meet preset rules, wherein the preset rules are used to indicate abnormality judgment conditions for the various first parameters.

[0051] In an embodiment of the present application, the event capturer obtains target information of multiple first parameters that meet preset rules from information of multiple first parameters.

[0052] In some embodiments, the preset rule includes: an abnormality judgment condition of the first parameter, and an abnormal event corresponding to the abnormality judgment condition. The abnormality judgment condition includes at least one of the following: a judgment condition of the abnormal event corresponding to the first parameter, and an upper limit on the duration that the first parameter meets the judgment condition.

[0053] The judgment condition of the abnormal event corresponding to the first parameter can be understood as that the specific values ​​of the first parameter at multiple sampling moments meet the judgment condition of the abnormal event corresponding to the first parameter.

[0054] Correspondingly, the upper limit of the time duration for which the first parameter meets the judgment condition can be understood as the upper limit of the time duration corresponding to the multiple consecutive sampling moments when the specific values ​​of the first parameter at the multiple consecutive sampling moments meet the judgment condition of the abnormal event corresponding to the first parameter.

[0055] In some embodiments, the abnormality judgment condition of the first parameter includes one or more sub-abnormality judgment conditions, and each sub-abnormality judgment condition corresponds to an abnormal event.

[0056] In some embodiments, the target information of the first parameter includes: multiple target information corresponding to the first parameter, wherein each target information includes an identifier of the target information, an abnormal event corresponding to the target information, a time period corresponding to the target information, a judgment condition corresponding to the abnormal event corresponding to the target information, and third descriptive information corresponding to the target information. The third descriptive information corresponding to the target information includes, but is not limited to, the highest value and the lowest value of a specific value within the time period corresponding to the target information.

[0057] Below, through Figure 3 , an exemplary description is given of obtaining target information of multiple first parameters that meet preset rules from information of multiple first parameters.

[0058] Figure 3 A schematic diagram of information of multiple first parameters provided in an embodiment of the present application, such as Figure 3 As shown, Figure 3 is information of three first parameters obtained at node A, the three first parameters including memory bandwidth, cache hit rate, and IB bandwidth, wherein,

[0059] The memory bandwidth information includes specific values ​​of the memory bandwidth at multiple sampling moments, including 100s, 101s, 102s, 103s, 104s, 105s, 106s, 107s, 108s, 109s, 110s, and 111s. The specific value of the memory bandwidth is, for example, 320 GB / s.

[0060] Preset rules can be obtained from the HPC semantic rule library.

[0061] exist Figure 3 In the example, the preset rules obtained from the HPC semantic rule library include: abnormal judgment conditions of memory bandwidth and abnormal events corresponding to the abnormal judgment conditions of memory bandwidth, abnormal judgment conditions of cache hit rate and abnormal events corresponding to the abnormal judgment conditions of cache hit rate, abnormal judgment conditions of IB bandwidth and abnormal events corresponding to the abnormal judgment conditions of IB bandwidth.

[0062] The abnormality judgment condition of the memory bandwidth includes a first sub-abnormality judgment condition, a second sub-abnormality judgment condition, and a third sub-abnormality judgment condition.

[0063] For example, the first sub-anomaly judgment condition is: the specific value of the memory bandwidth at multiple consecutive sampling moments is greater than or equal to the first anomaly threshold, and the upper limit of the duration corresponding to the multiple consecutive sampling moments is 3 seconds. For example, the first anomaly threshold is 350 GB / s. The abnormal event corresponding to the first sub-anomaly judgment condition is: a memory bandwidth-intensive period.

[0064] For example, according to Figure 3As shown, the specific values ​​of the memory bandwidth at 102s, 103s, and 104s are 360 ​​GB / s, 380 GB / s, and 380 GB / s, respectively, which are all greater than 350 GB / s, and the duration corresponding to 102s, 103s, and 104s is 3 seconds. Therefore, target information 1 for determining the memory bandwidth includes: event E1, memory bandwidth-intensive segment, 102s-104s, memory bandwidth ≥ 350 GB / s and lasting for 3 seconds, maximum 380 GB / s, and minimum 360 GB / s. Among them, event E1 represents the identifier of target information 1, memory bandwidth intensive segment represents the abnormal event corresponding to target information 1, 102s-104s represents the time period corresponding to target information 1, including 102s, 103s, and 104s, memory bandwidth ≥350 GB / s and lasting for 3 seconds represents the judgment condition of the memory bandwidth intensive segment, maximum 380 GB / s and minimum 360 GB / s represent the third descriptive information corresponding to target information 1, maximum 380 GB / s represents the maximum memory bandwidth of 380 GB / s within 102s-104s, and minimum 360 GB / s represents the minimum memory bandwidth of 360 GB / s within 102s-104s.

[0065] Memory bandwidth target information 2 includes: event E7, memory bandwidth-intensive period, duration of 105s-107s, memory bandwidth ≥ 350 GB / s for 3 seconds, maximum 384 GB / s, and minimum 382 GB / s. Memory bandwidth target information 3 includes: event E8, memory bandwidth-intensive period, duration of 108s-110s, memory bandwidth ≥ 350 GB / s for 3 seconds, maximum 390 GB / s, and minimum 388 GB / s. The determination of target information 2 and 3 is similar to that of target information 1 and is not repeated here.

[0066] For example, the second sub-anomaly judgment condition is: the specific value of the memory bandwidth at the current sampling time is greater than the moving average (MA) 5 corresponding to the memory bandwidth at the current sampling time. The MA5 corresponding to the memory bandwidth at the current sampling time is specifically: the specific value of the memory bandwidth at the current sampling time and the average of the specific values ​​of the memory bandwidth at the four sampling times before the current sampling time. The abnormal event corresponding to the second sub-anomaly judgment condition is: a short-term upward trend in memory bandwidth.

[0067] When the number of sampling moments before the current sampling moment is less than 4, the MA5 corresponding to the memory bandwidth at the current sampling moment is determined according to the specific value of the memory bandwidth at the current sampling moment and the specific values ​​of the memory bandwidth at the existing sampling moments.

[0068] For example, according to Figure 3As shown, the specific values ​​of the memory bandwidth of 100s, 101s, 102s, 103s, and 104s are 320GB / s, 340GB / s, 360GB / s, 380GB / s, and 380GB / s respectively. The MA5 corresponding to the memory bandwidth of 101s is , and 340GB / s is greater than The method for determining the MA5 corresponding to the memory bandwidth of 102s and 103s is similar to the method for determining the MA5 corresponding to the memory bandwidth of 101s, and will not be repeated here. The MA5 corresponding to the memory bandwidth of 104s is , and 380GB / s is greater than The method for determining the MA5 corresponding to the memory bandwidth of 105s-11s is similar to that for 104s and is not repeated here. Target information 4 for memory bandwidth can be determined as follows: event E4, short-term upward trend in memory bandwidth, 100s-110s, memory bandwidth above MA5, maximum of 390 GB / s, and minimum of 320 GB / s.

[0069] The third sub-abnormality judgment condition includes: the specific value of the memory bandwidth at the current sampling time is less than the MA5 corresponding to the memory bandwidth at the current sampling time. The abnormal event corresponding to the third sub-abnormality judgment condition is: a short-term downward trend in memory bandwidth.

[0070] The abnormality judgment condition of the cache hit rate includes a fourth sub-abnormality judgment condition, a fifth sub-abnormality judgment condition, and a sixth sub-abnormality judgment condition.

[0071] The fourth sub-anomaly judgment condition can be, for example, that the cache hit ratio at multiple consecutive sampling moments is less than or equal to the second anomaly threshold, and the duration corresponding to the multiple consecutive sampling moments is capped at 3 seconds. The second anomaly threshold can be, for example, 75%. The abnormal event corresponding to the fourth sub-anomaly judgment condition is a cache pressure event.

[0072] For example, according to Figure 3 As shown, based on the fourth sub-anomaly judgment condition, target cache hit rate information 5 can be confirmed to include: event E2, cache pressure event, 110s-111s, cache hit rate ≤ 75% and lasting for 3 seconds, maximum 74%, and minimum 72%. The specific method for determining target cache hit rate information 5 is similar to the specific method for determining target memory bandwidth information 1 and is not repeated here.

[0073] The fifth abnormality judgment condition is, for example, that the specific value of the cache hit rate at the current sampling moment is greater than the MA5 corresponding to the cache hit rate at the current sampling moment. The abnormal event corresponding to the fifth abnormality judgment condition is that the cache hit rate has a short-term upward trend.

[0074] The sixth abnormality judgment condition includes: the specific value of the cache hit rate at the current sampling moment is less than the MA5 corresponding to the cache hit rate at the current sampling moment. The abnormal event corresponding to the third abnormality judgment condition is: the cache hit rate has a short-term downward trend.

[0075] For example, according to Figure 3 As shown, based on the sixth sub-anomaly judgment condition, target cache hit rate information 6 can be confirmed to include: event E5, a short-term downward trend in cache hit rate, 100s-111s, a cache hit rate below MA5, a maximum of 74%, and a minimum of 72%. The specific method for determining target cache hit rate information 6 is similar to the specific method for determining target memory bandwidth information 4 and will not be repeated here.

[0076] The abnormality judgment condition of the IB bandwidth includes a seventh sub-abnormality judgment condition, an eighth sub-abnormality judgment condition, and a ninth sub-abnormality judgment condition.

[0077] For example, the seventh sub-abnormality judgment condition is: the specific value of the IB bandwidth at the current sampling time is less than or equal to the third abnormality threshold, the specific value of the memory bandwidth at the current sampling time is greater than or equal to the first abnormality threshold, and the specific value of the cache hit rate at the current sampling time is less than or equal to the second abnormality threshold. The third abnormality threshold is, for example, 10 GB / s. The abnormal event corresponding to the seventh sub-abnormality judgment condition is: a communication bottleneck accompanied by memory access degradation.

[0078] For example, according to Figure 3 As shown, the specific values ​​of the IB bandwidth of 110s and 111s are 12 GB / s and 10 GB / s, respectively, both less than or equal to 10 GB / s; the specific values ​​of the memory bandwidth of 110s and 111s are 390 GB / s and 388 GB / s, respectively, both greater than 350 GB / s; the specific values ​​of the cache hit rate of 110s and 111s are 74% and 72%, respectively, both less than 75%. It can be determined that the target information 7 of the IB bandwidth includes: event E3, communication bottleneck accompanied by memory access degradation, 110s-111s, IB bandwidth ≤ 15 GB / s, maximum 12 GB / s, and minimum 10 GB / s.

[0079] The eighth sub-abnormality judgment condition is, for example, that the specific value of the IB bandwidth at the current sampling moment is greater than the MA5 corresponding to the IB bandwidth at the current sampling moment. The abnormal event corresponding to the eighth sub-abnormality judgment condition is a short-term upward trend in the IB bandwidth.

[0080] For example, according to Figure 3As shown, based on the eighth sub-abnormality judgment condition, target IB bandwidth information 8 can be confirmed to include: event E9, short-term upward trend in IB bandwidth, 102s-107s, IB bandwidth exceeding MA5, maximum of 46 GB / s, and minimum of 35 GB / s. The specific method for determining target IB bandwidth information 8 is similar to the specific method for determining target memory bandwidth information 4 and is not further described here.

[0081] The ninth sub-abnormality judgment condition includes: the specific value of the IB bandwidth at the current sampling time is less than the MA5 corresponding to the IB bandwidth at the current sampling time. The abnormal event corresponding to the ninth sub-abnormality judgment condition is: a short-term downward trend in the IB bandwidth.

[0082] For example, according to Figure 3 As shown, based on the ninth sub-anomaly judgment condition, target IB bandwidth information 9 can be determined to include: event E6, short-term downward trend in IB bandwidth, 108s-111s, IB bandwidth below MA5, maximum 30 GB / s, and minimum 10 GB / s. The specific method for determining target IB bandwidth information 9 is similar to the specific method for determining target memory bandwidth information 4 and is not further described here.

[0083] Optionally, the preset rule may further include: an abnormality judgment condition is that the specific value of the first parameter at the current sampling moment is greater than the MA20 corresponding to the first parameter at the current sampling moment; the abnormal event corresponding to this abnormality judgment condition is: a medium-term upward trend of the first parameter. The MA20 corresponding to the memory bandwidth at the current sampling moment is specifically: the specific value of the memory bandwidth at the current sampling moment and the average of the specific values ​​of the memory bandwidth at the 19 sampling moments before the current sampling moment.

[0084] Optionally, the preset rules may further include: the abnormal judgment condition is that the specific value of the first parameter at the current sampling moment is less than the MA20 corresponding to the first parameter at the current sampling moment, and the abnormal event corresponding to the abnormal judgment condition is: the medium-term downward trend of the first parameter.

[0085] For details on how to handle MA20, please refer to the MA20 Handling Method, which will not be repeated here.

[0086] Optionally, the third description information of the target information may further include: an average value of specific values ​​within a time period corresponding to the target information.

[0087] Optionally, the preset rule may further include: an abnormality determination condition is that the specific value of the first parameter at the current sampling moment is greater than a first preset value times a preset normal threshold value of the first parameter; the abnormal event corresponding to this abnormality determination condition is: a sudden spike in memory bandwidth. For example, the specific value of the memory bandwidth at the current sampling moment is greater than 5 times the preset normal threshold value of the memory bandwidth.

[0088] The preset rules in the HPC semantic rule library are classified according to the numerical analysis methods of the HPC platform. The numerical analysis methods include but are not limited to: explicit finite element method, implicit finite element method, finite volume method, boundary element method, and particle method.

[0089] In some embodiments, obtaining target information of multiple first parameters that meet preset rules from multiple first parameter information includes:

[0090] Obtain numerical analysis methods for HPC platforms;

[0091] According to the numerical analysis method of the HPC platform, the preset rules are obtained in the HPC semantic rule library;

[0092] From the information of the multiple first parameters, target information of the multiple first parameters that meet the preset rules is obtained.

[0093] Specifically, the HPC semantic rule base includes the numerical analysis method and preset rules of the HPC platform, wherein the numerical analysis method of the HPC platform corresponds one-to-one to the preset rules. According to the numerical analysis method of the HPC platform, the corresponding preset rules can be obtained in the HPC semantic rule base.

[0094] For example, when the numerical analysis method of the HPC platform is explicit finite element, explicit finite element has high requirements for memory bandwidth, and the preset rules obtained include: the abnormal judgment condition is: the specific value of the memory bandwidth at multiple consecutive sampling moments is greater than or equal to the first abnormal threshold, and the upper limit of the time length corresponding to the multiple consecutive sampling moments is 3 seconds.

[0095] For example, when the numerical analysis method used on the HPC platform is implicit finite element analysis, which has high requirements for memory bandwidth, the preset rules for obtaining the information include: the specific values ​​of the memory bandwidth at multiple consecutive sampling moments are greater than or equal to a first abnormality threshold, and the upper limit of the duration corresponding to the multiple consecutive sampling moments is 3 seconds; and the abnormal event corresponding to the abnormality judgment condition is: a memory bandwidth-intensive period. In this case, the first abnormality threshold is, for example, 200 GB / s.

[0096] In the examples of this application, different numerical analysis methods have significantly different requirements for the performance indicators of the HPC platform. For example, explicit finite element analysis requires higher memory bandwidth than implicit finite element analysis. Based on the preset rules for classifying numerical analysis methods, differentiated anomaly thresholds can be customized for different scenarios, avoiding misjudgments or missed judgments caused by using a unified threshold. This makes the screening of target information more relevant to actual business scenarios and improves the accuracy of anomaly identification.

[0097] S203: Aggregate the target information according to the type of the first parameter corresponding to the target information to obtain first description information describing a change state of the first parameter.

[0098] In an embodiment of the present application, the event aggregator aggregates the target information according to the type of the first parameter corresponding to the target information to obtain first description information describing the change state of the first parameter.

[0099] In some embodiments, the target information is aggregated according to the type of the first parameter corresponding to the target information to obtain first description information describing the change state of the first parameter, which can be understood as:

[0100] Aggregation processing is performed on target information corresponding to the same first parameter to obtain first description information describing a change state of the first parameter.

[0101] Aggregation processing specifically involves analyzing and merging target information with the same abnormal events based on time.

[0102] The first description information includes: an identifier of the first description information, a time period corresponding to the first description information, and a description of the first description information.

[0103] The following is an exemplary description of performing aggregation processing on the target information according to the type of the first parameter corresponding to the target information to obtain the first description information describing the change state of the first parameter:

[0104] The type of the first parameter corresponding to target information 1, target information 2, target information 3, and target information 4 is memory bandwidth. Target information 1, target information 2, target information 3, and target information 4 are aggregated to obtain first description information 1 that describes the change state of the memory bandwidth.

[0105] The type of the first parameter corresponding to the target information 5 and the target information 6 is the cache hit rate. The target information 5 and the target information 6 are aggregated to obtain the first description information 2 describing the change state of the cache hit rate.

[0106] The type of the first parameter corresponding to target information 7, target information 8, and target information 9 is IB bandwidth. Target information 7, target information 8, and target information 9 are aggregated to obtain first description information 3 describing the change state of the IB bandwidth.

[0107] S204: Analyze the performance of the high-performance computing platform according to the plurality of first description information corresponding to the plurality of first parameters.

[0108] In an embodiment of the present application, the parsing module analyzes the performance of the high-performance computing platform based on multiple first description information corresponding to multiple first parameters.

[0109] In some embodiments, based on a Retrieval-Augmented Generation (RAG) system, the performance of the high-performance computing platform is analyzed according to a plurality of first description information corresponding to a plurality of first parameters, and a performance analysis report of the high-performance computing platform is output.

[0110] The RAG system includes a preset knowledge base and a RAG analysis module. Specifically, the RAG system analyzes the performance of the high-performance computing platform based on the preset knowledge base and multiple first description information corresponding to multiple first parameters, and outputs a performance analysis report of the high-performance computing platform.

[0111] The performance analysis report of the high-performance computing platform includes but is not limited to: a natural language description text, a root cause analysis of abnormal performance phenomena of the high-performance computing platform, and optimization suggestions for the abnormal phenomena.

[0112] Optionally, the performance analysis report of the high-performance computing platform supports output in the form of a portable document format (PDF), a World Wide Web (WEB) interface, or an application programming interface (API).

[0113] In an embodiment of the present application, information on multiple first parameters is obtained, where the first parameters are associated with the performance of a high-performance computing platform; from the information on the multiple first parameters, target information on multiple first parameters that meet preset rules is obtained, where the preset rules are used to indicate abnormal judgment conditions for the various first parameters; based on the type of the first parameter corresponding to the target information, the target information is aggregated to obtain first descriptive information describing the change state of the first parameter; and based on the multiple first descriptive information corresponding to the multiple first parameters, the performance of the high-performance computing platform is analyzed. Through the above method, based on the information on the multiple first parameters and the preset rules, the performance of the high-performance computing platform can be automatically analyzed, reducing dependence on manual experience and improving the analysis efficiency of the performance analysis of the HPC platform.

[0114] Secondly, by filtering target information through preset rules and generating structured descriptions based on the type of first parameters, the system can quickly focus on abnormal events while systematically integrating the changing states of multiple first parameters. This achieves the efficient transformation from scattered information to structured descriptions, making the analysis process more targeted and comprehensive, and effectively enhancing the depth of HPC platform performance analysis. Because preset rules can be dynamically updated, HPC platform performance analysis has good scalability and better adapts to the ever-changing performance monitoring needs of HPC platforms.

[0115] In addition, the structured first description information provides standardized input for the RAG system pipeline, ensuring full compatibility with the RAG system, enabling performance data to seamlessly connect with knowledge retrieval and analysis logic, and providing stable data support for the automated processing of performance analysis on the HPC platform, avoiding the high-cost model fine-tuning caused by the excessive complexity of the HPC platform's performance data.

[0116] In the above Figure 2 Based on the examples below, combined Figure 4 , a detailed description is given of a process of aggregating the target information according to the type of the first parameter corresponding to the target information to obtain the first description information describing the change state of the first parameter.

[0117] Figure 4 This is a schematic diagram of the process of first describing information provided in the embodiment of this application. Figure 4 , the method may include:

[0118] S401: Determine an abnormal event corresponding to target information.

[0119] Exemplarily, the abnormal events corresponding to target information 1, target information 2, and target information 3 are determined to be memory bandwidth intensive segments, the abnormal event corresponding to target information 4 is determined to be a short-term upward trend in memory bandwidth, the abnormal event corresponding to target information 5 is determined to be a cache pressure event, the abnormal event corresponding to target information 6 is determined to be a short-term downward trend in cache hit rate, the abnormal event corresponding to target information 7 is determined to be a communication bottleneck accompanied by memory access degradation, the abnormal event corresponding to target information 8 is determined to be a short-term upward trend in IB bandwidth, and the abnormal event corresponding to target information 9 is determined to be a short-term downward trend in IB bandwidth.

[0120] S402: Merge the same target information of the abnormal events to obtain second description information of the abnormal events.

[0121] In some embodiments, the same target information of the abnormal events is merged to obtain the second description information of the abnormal events, including:

[0122] When there are at least two pieces of target information of the abnormal event, merging the same target information of the abnormal event to obtain second description information of the abnormal event;

[0123] When there is only one target information of the abnormal event, the target information is determined as the second description information of the abnormal event.

[0124] For example, if target information 1, target information 2, and target information 3 correspond to the same abnormal event, target information 1, target information 2, and target information 3 are merged to obtain second description information 1 for the memory bandwidth-intensive segment. Second description information 1 includes: event G1, memory bandwidth-intensive segment, 102s-110s, memory bandwidth ≥ 350 GB / s and lasting for 3 seconds, maximum 390 GB / s, and minimum 360 GB / s. Event G1 indicates the identifier of second description information 1, memory bandwidth-intensive segment indicates the abnormal event corresponding to second description information 1, 102s-111s indicates the time period corresponding to second description information 1, memory bandwidth ≥ 350 GB / s and lasting for 3 seconds indicates the judgment condition for the memory bandwidth-intensive segment, maximum 390 GB / s indicates that the memory bandwidth is at most 380 GB / s within 102s-111s, and minimum 360 GB / s indicates that the memory bandwidth is at least 360 GB / s within 102s-111s.

[0125] Target information 4 is determined as second description information 2 of a short-term upward trend in memory bandwidth. Target information 5 is determined as second description information 3 of a cache pressure event. Target information 6 is determined as second description information 4 of a short-term downward trend in cache hit rate. Target information 7 is determined as second description information 5 of a communication bottleneck accompanied by memory access degradation. Target information 8 is determined as second description information 6 of a short-term upward trend in IB bandwidth. Target information 9 is determined as second description information 7 of a short-term downward trend in IB bandwidth.

[0126] S403: Merge the second description information according to the type of the first parameter corresponding to the abnormal event to obtain the first description information.

[0127] In some embodiments, the second description information is merged according to the type of the first parameter corresponding to the abnormal event to obtain the first description information, which can be understood as:

[0128] The second description information of abnormal events corresponding to the same type of first parameter is merged to obtain first description information describing the change state of the first parameter.

[0129] The following is an exemplary description of how the second description information is merged according to the type of the first parameter corresponding to the abnormal event to obtain the first description information:

[0130] The type of the first parameter corresponding to the abnormal events of the second description information 1 and the second description information 2 is memory bandwidth. The second description information 1 and the second description information 2 are merged to obtain the first description information 1 describing the change state of the memory bandwidth.

[0131] The type of the first parameter corresponding to the abnormal events of the second description information 3 and the second description information 4 is the cache hit rate. The second description information 3 and the second description information 4 are merged to obtain the first description information 2 describing the change state of the cache hit rate.

[0132] The type of the first parameter corresponding to the abnormal events of the second description information 5, the second description information 6, and the second description information 7 is IB bandwidth. The second description information 5, the second description information 6, and the second description information 7 are merged to obtain the first description information 3 describing the change state of the IB bandwidth.

[0133] In an embodiment of the present application, the abnormal events corresponding to the target information are determined; the target information with the same abnormal events is merged to obtain second description information of the abnormal events; and the second description information is merged according to the type of the first parameter corresponding to the abnormal event to obtain first description information. Through the above method, the abnormal events are merged and then merged according to the type of the first parameter, thereby constructing a hierarchical system for determining the first description information, which not only retains the key features of the original data but also significantly reduces information redundancy.

[0134] In the above Figure 4 Based on the examples below, combined Figure 5 , for any first parameter, a process of merging the second description information according to the type of the first parameter corresponding to the abnormal event to obtain the first description information is described in detail.

[0135] Figure 5 This is a schematic diagram of the aggregation process provided in the embodiment of this application. Figure 5 , the method may include:

[0136] S501: Determine a plurality of first time periods corresponding to second description information of an abnormal event of the same type as the first parameter, and a second time period corresponding to information of the first parameter.

[0137] Hereinafter, an exemplary description is given of determining a plurality of first time periods corresponding to the second description information of an abnormal event of the same type as the first parameter.

[0138] For example, the type of the first parameter corresponding to the abnormal events of the second description information 1 and the second description information 2 is memory bandwidth, the time period corresponding to the second description information 1 is 102s-111s, and the time period corresponding to the second description information 2 is 100s-110s. Then, it is determined that the multiple first time periods corresponding to the second description information of the abnormal event corresponding to the memory bandwidth include: 102s-111s, and 100s-110s.

[0139] Similarly, the multiple first time periods corresponding to the second description information of the abnormal event corresponding to the cache hit rate include: 110s-111s, and 100s-111s. The multiple first time periods corresponding to the second description information of the abnormal event corresponding to the IB bandwidth include: 110s-111s, 102s-107s, and 108s-111s.

[0140] according to Figure 3 The information of the first parameter shown can determine that the second time period corresponding to the information of the first parameter is 100s-111s.

[0141] S502: Determine first description information based on multiple first time periods, second time periods, and second description information of abnormal events of the same type as the first parameter.

[0142] In some embodiments, determining the first description information based on the plurality of first time periods, the second time periods, and the second description information of abnormal events of the same type as the first parameter includes:

[0143] determining whether the plurality of first time periods overlap the second time period;

[0144] When multiple first time periods overlap the second time period, based on the second time period, the second description information of abnormal events of the same type as the first parameter is merged to obtain the first description information;

[0145] When the plurality of first time periods do not overlap the second time period, the first description information is determined based on the plurality of first time periods, the second time period, and second description information of the abnormal event of the same type as the first parameter.

[0146] In some embodiments, determining whether the plurality of first time periods overlaps the second time period includes:

[0147] determining a total time period of the plurality of first time periods;

[0148] Determine whether the total period covers the second period.

[0149] For example, for memory bandwidth, multiple first time periods are 102s-111s and 100s-110s, and the second time period is 100s-111s. If the total time period of the multiple first time periods is determined to be 100s-111s, then it is determined that the total time period covers the second time period.

[0150] Exemplarily, for cache hit rate, multiple first time periods are 110s-111s and 100s-111s, the second time period is 100s-111s, and the total time period of the multiple first time periods is determined to be 100s-111s, then it is determined that the total time period covers the second time period.

[0151] In the following, regarding memory bandwidth, when multiple first time periods overlap a second time period, second description information of abnormal events of the same type as the first parameter is merged based on the second time period to obtain first description information is exemplarily described.

[0152] The second description information 1 includes: event G1, memory bandwidth intensive segment, 102s-110s, memory bandwidth ≥350 GB / s and lasting 3 seconds, maximum 390 GB / s, and minimum 360 GB / s. The second description information 2 includes: event E4, short-term upward trend of memory bandwidth, 100s-110s, memory bandwidth higher than MA5, maximum 390 GB / s, and minimum 320 GB / s. Based on the second time period, the second description information 1 and the second description information 2 are merged to obtain the first description information 1.

[0153] The first description information 1 includes: event Q1, 100s-111s, and description 1. Description 1 states: Memory bandwidth experienced a short-term upward trend from 320 GB / s to a maximum of 390 GB / s between 100s and 110s, and remained above 350 GB / s between 102s and 111s. Event Q1 identifies first description information 1, and 100s-111s represents the time period corresponding to first description information 1 (i.e., the second time period).

[0154] Exemplarily, for IB bandwidth, multiple first time periods are 110s-111s, 102s-107s, and 108s-111s, and the second time period is 101s-111s. If the total time period of the multiple first time periods is determined to be 100s-111s, then it is determined that the total time period does not cover the second time period.

[0155] Similarly, it can be determined that first description information 2 includes: event Q2, 100s-111s, and description 2. Description 2 states: the cache hit rate shows a short-term downward trend from a maximum of 94% to 72% between 10s and 111s, and the cache hit rate is below 75% between 110s and 111s. The process for determining first description information 2 is similar to that for first description information 1 and is not repeated here.

[0156] In some embodiments, determining the first description information based on the plurality of first time periods, the second time periods, and the second description information of the abnormal event of the same type as the first parameter includes:

[0157] Based on the plurality of first time periods, determining, in the second time period, a third time period that is not covered by the plurality of first time periods;

[0158] Obtain information of the first parameter corresponding to the third time period;

[0159] The second description information of the abnormal event of the same type as the first parameter and the information of the first parameter corresponding to the third time period are combined to obtain the first description information.

[0160] For example, for IB bandwidth, multiple first time periods are 110s-111s, 102s-107s, and 108s-111s, and the second time period is 101s-111s. A third time period not covered by the multiple first time periods can be determined to be 100s-101s. Then, information about the first parameter corresponding to the third time period can be obtained, including: a specific value of the IB bandwidth for 100s is 35 GB / s, and a specific value of the IB bandwidth for 101s is 35 GB / s. Second description information 5, second description information 6, second description information 7, and the information about the first parameter corresponding to the third time period are combined to obtain first description information 2.

[0161] The first description information 3 includes: event Q3, 100s-111s, and description 3. Description 3 is: from 100s to 102s, the IB bandwidth is stable at 35 GB / s. From 102s to 107s, the IB bandwidth shows a short-term upward trend, from 35 GB / s to a maximum of 46 GB / s. From 108s to 111s, the IB bandwidth shows a short-term downward trend, from a maximum of 30 GB / s to 10 GB / s. From 110s to 111s, the IB bandwidth is less than 15 GB / s.

[0162] In an embodiment of the present application, multiple first time periods corresponding to the second description information of an abnormal event of the same type as the first parameter, and the second time period corresponding to the information of the first parameter are determined; based on the multiple first time periods, the second time period, and the second description information of the abnormal event of the same type as the first parameter, the first description information is determined. Through the above method, the first description information can be made more comprehensive. Specifically, when multiple first time periods do not cover the second time period, based on the multiple first time periods, a third time period not covered by the multiple first time periods is determined in the second time period; the information of the first parameter corresponding to the third time period is obtained; the second description information of the abnormal event of the same type as the first parameter and the information of the first parameter corresponding to the third time period are merged to obtain the first description information. By comparing the multiple first time periods with the second time periods, the third time period not covered by the abnormal event is accurately identified, and the information of the time period is supplemented (such as the IB bandwidth is stable at 35GB / s within 100s-101s). This coverage check, combined with the missing time period completion mechanism, avoids information gaps caused by abnormal event screening. It enables the first description information to fully present the parameter change trajectory within the entire monitoring cycle (such as 100s-111s), ensuring the coherence of the timing logic (such as the complete stage of IB bandwidth from "stable to rising and then falling"), and improves the comprehensiveness of the performance analysis of the HPC platform.

[0163] In the above Figure 2 Based on the examples below, combined Figure 6 , a detailed description is given of the process of analyzing the performance of the high-performance computing platform based on multiple description information corresponding to multiple first parameters.

[0164] Figure 6 This is a schematic diagram of the process of analyzing the performance of a high-performance computing platform provided in an embodiment of the present application. Figure 6 , the method may include:

[0165] S601: Convert multiple description information corresponding to multiple first parameters into description texts in natural language.

[0166] In some embodiments, through a pre-trained language model, multiple description information corresponding to multiple first parameters are converted into description texts in natural language based on standard sentence patterns, adjustment of word order, merging of expressions, etc.

[0167] In some embodiments, the parsing module includes a structured natural language cleaning module, and a RAG system.

[0168] The structured natural language cleaning module converts multiple descriptive information corresponding to multiple first parameters into natural language description text through a pre-trained language model based on standardized sentence structure, adjusted word order, and merged expressions.

[0169] In an embodiment of the present application, multiple description information corresponding to multiple first parameters are converted into description texts in natural language to improve the semantic clarity and expression consistency of the performance description of the HPC platform.

[0170] For example, the natural language descriptions of first description information 1, first description information 2, and first description information 3 are as follows: Between seconds 100 and 111, memory bandwidth utilization continued to rise, from an initial 320 GB / s to a peak of 390 GB / s, and remained stable at over 350 GB / s between seconds 102 and 111. Simultaneously, the cache hit rate gradually declined, from an initial 94% to 72%, and remained below 75% between seconds 110 and 111. During the same time period, IB bandwidth demand exhibited phased fluctuations: from seconds 100 to 101, it remained stable at 35 GB / s; from seconds 102 to 107, it continued to rise, reaching a peak of 46 GB / s; then, from seconds 108 to 111, it rapidly fell back to 10 GB / s, and remained below 15 GB / s between seconds 110 and 111.

[0171] S602: Based on the description text in the natural language, identify knowledge fragments associated with the description text in the natural language in a preset knowledge base.

[0172] The RAG system identifies knowledge fragments associated with the natural language description text in a preset knowledge base based on the natural language description text.

[0173] The preset knowledge base includes a plurality of description texts in natural language, a plurality of knowledge fragments associated with the description texts in natural language, and a correspondence between the description texts in natural language and one or more knowledge fragments associated with the description texts in natural language.

[0174] The following is an exemplary description of identifying knowledge segments associated with the natural language description text in a preset knowledge base based on the natural language description text given in S601:

[0175] Based on the natural language description text, the preset knowledge base identifies the following knowledge fragments associated with the natural language description text: non-local data access will cause cache invalidation, and when switching between the communication and computing phases, the IB bandwidth will first increase and then decrease. Moreover, if the local computing phase fails to effectively reuse synchronized data, the memory bandwidth pressure will continue to increase.

[0176] S603: Analyze the performance of the high-performance computing platform based on the natural language description text and the knowledge fragments to obtain a performance analysis report of the high-performance computing platform.

[0177] Through the RAG system, the performance of the high-performance computing platform is analyzed based on the natural language description text and knowledge fragments, and a performance analysis report of the high-performance computing platform is obtained.

[0178] The following is an exemplary description of analyzing the performance of a high-performance computing platform based on the natural language description text and the knowledge fragments, and obtaining a performance analysis report of the high-performance computing platform based on the analysis report associated with the natural language description text identified in S602:

[0179] Based on the natural language description text and the knowledge fragments associated with natural language description text 1, the root cause analysis of the abnormal performance of the high-performance computing platform is determined as follows: it is speculated that the high-performance computing platform may have loaded a new non-local data block between 100 and 111 seconds, causing the CPU to frequently access memory rather than the cache, which in turn drove the memory bandwidth utilization rate to a continuous increase to 390GB / s (stable above 350GB / s between 102 and 111 seconds), creating a chain reaction of "cache failure leading to a surge in memory accesses." It is speculated that 102 to 107 seconds is the cross-node data synchronization phase (high communication demand), and after 108 seconds, the local computing phase (communication demand drops sharply), but the cache did not effectively reuse the synchronized data, resulting in continued memory pressure.

[0180] Based on the root cause analysis of the abnormal performance of the high-performance computing platform, the optimization suggestions are: adjust the data partitioning strategy to improve cache locality, and add a data prefetching mechanism after the communication is completed to reduce memory bandwidth pressure;

[0181] Combine root cause analysis and optimization suggestions to obtain a performance analysis report for the high-performance computing platform.

[0182] In an embodiment of the present application, multiple descriptive information corresponding to multiple first parameters is converted into natural language description text; based on the natural language description text, knowledge fragments associated with the natural language description text are identified in a preset knowledge base; based on the natural language description text and the knowledge fragments, the performance of the high-performance computing platform is analyzed to obtain a performance analysis report of the high-performance computing platform. In the above method, the performance analysis report of the high-performance computing platform is automatically generated, which effectively improves the efficiency of generating the performance analysis report. At the same time, relying on the preset knowledge base, the comprehensiveness and professionalism of the analysis are enhanced, providing more efficient and reliable support for the performance analysis of the high-performance computing platform.

[0183] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method.

[0184] Figure 7 This is a schematic diagram of the performance analysis method for a high-performance computing platform provided in an embodiment of the present application. Figure 7 As shown, the embodiment of the present application further provides a performance analysis device 70 for a high-performance computing platform, comprising: a first acquisition module 701, a second acquisition module 702, a first processing module 703 and a second processing module 704, wherein:

[0185] A first acquisition module 701 is configured to acquire information of a plurality of first parameters, where the first parameters are associated with the performance of the high-performance computing platform;

[0186] The second acquisition module 702 is configured to acquire target information of multiple first parameters that meet preset rules from information of multiple first parameters, where the preset rules are used to indicate abnormality judgment conditions of the various first parameters;

[0187] A first processing module 703 is configured to aggregate the target information according to the type of the first parameter corresponding to the target information to obtain first description information describing a change state of the first parameter;

[0188] The second processing module 704 is configured to analyze the performance of the high-performance computing platform according to the plurality of first description information corresponding to the plurality of first parameters.

[0189] For the description of the features in the embodiment corresponding to the performance analysis device 70 of the high-performance computing platform, please refer to the relevant description of the embodiment corresponding to the performance analysis method of the high-performance computing platform, which will not be repeated here.

[0190] In some embodiments, the preset rules include:

[0191] Abnormality judgment conditions of the first parameter, the abnormality judgment conditions including at least one of the following: a judgment condition of an abnormal event corresponding to the first parameter, and an upper limit of the time during which the first parameter meets the judgment condition;

[0192] Abnormal events corresponding to abnormal judgment conditions.

[0193] In some embodiments, the first processing module 703 is specifically configured to:

[0194] Determine the abnormal events corresponding to the target information;

[0195] Merge the same target information of the abnormal events to obtain the second description information of the abnormal events;

[0196] The second description information is merged according to the type of the first parameter corresponding to the abnormal event to obtain the first description information.

[0197] In some embodiments, for any first parameter, the first processing module 703 is specifically configured to:

[0198] Determine a plurality of first time periods corresponding to second description information of abnormal events of the same type as the first parameter, and a second time period corresponding to the information of the first parameter;

[0199] The first description information is determined based on the plurality of first time periods, the second time period, and the second description information of abnormal events of the same type as the first parameter.

[0200] In some embodiments, the first processing module 703 is specifically configured to:

[0201] When multiple first time periods overlap the second time period, based on the second time period, the second description information of abnormal events of the same type as the first parameter is merged to obtain the first description information;

[0202] When the plurality of first time periods do not overlap the second time period, the first description information is determined based on the plurality of first time periods, the second time period, and second description information of the abnormal event of the same type as the first parameter.

[0203] In some embodiments, the first processing module 703 is specifically configured to:

[0204] Based on the plurality of first time periods, determining, in the second time period, a third time period that is not covered by the plurality of first time periods;

[0205] Obtain information of the first parameter corresponding to the third time period;

[0206] The second description information of the abnormal event of the same type as the first parameter and the information of the first parameter corresponding to the third time period are combined to obtain the first description information.

[0207] In some embodiments, the second processing module 704 is specifically configured to:

[0208] Converting a plurality of description information corresponding to the plurality of first parameters into description texts in a natural language;

[0209] According to the description text in natural language, identifying the knowledge fragments associated with the description text in natural language in a preset knowledge base;

[0210] The performance of the high-performance computing platform is analyzed based on the natural language description text and knowledge fragments to obtain a performance analysis report of the high-performance computing platform.

[0211] For the description of the features in the embodiment corresponding to the performance analysis device 70 of the high-performance computing platform, please refer to the relevant description of the embodiment corresponding to the performance analysis method of the high-performance computing platform, which will not be repeated here.

[0212] Figure 8 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present application. Figure 8 As shown, the electronic device 80 provided in this embodiment includes: at least one processor 801 and a memory 802. Optionally, the electronic device 80 further includes a communication component 803. The processor 801, the memory 802 and the communication component 803 are connected via a bus.

[0213] During the specific implementation process, at least one processor 801 executes the computer-executable instructions stored in the memory 802, so that the at least one processor 801 executes the above-mentioned embodiment of the performance analysis method for the high-performance computing platform.

[0214] The specific implementation process of the processor 801 can be found in the above method embodiment. Its implementation principle and technical effects are similar and will not be repeated here in this embodiment.

[0215] In the above embodiments, it should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the application may be directly executed by a hardware processor or by a combination of hardware and software modules within the processor.

[0216] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage.

[0217] A bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. Buses can be categorized as address buses, data buses, and control buses. For ease of illustration, the buses in the drawings of this application are not limited to just one bus or just one type of bus.

[0218] An embodiment of the present application further provides a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps of any of the above-mentioned performance analysis method embodiments for high-performance computing platforms when running.

[0219] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0220] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any of the above-mentioned performance analysis method embodiments for high-performance computing platforms are implemented.

[0221] An embodiment of the present application also provides another computer program product, including a non-volatile computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps in any of the above-mentioned performance analysis method embodiments of the high-performance computing platform.

[0222] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0223] The above is a detailed introduction to the performance analysis method, electronic device and storage medium of a high-performance computing platform provided by this application. Specific examples are used herein to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core idea of ​​this application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of this application, several improvements and modifications can be made to this application, and these improvements and modifications also fall within the scope of protection of the claims of this application.

Claims

1. A performance analysis method for a high-performance computing platform, characterized in that: include: Obtaining information of a plurality of first parameters, wherein the first parameters are associated with performance of a high-performance computing platform; Obtaining target information of the plurality of first parameters that satisfies a preset rule from among the plurality of first parameter information, wherein the preset rule is used to indicate abnormality judgment conditions of the various first parameters; performing aggregation processing on the target information according to the type of the first parameter corresponding to the target information to obtain first description information describing a change state of the first parameter; The performance of the high-performance computing platform is analyzed according to the multiple pieces of first description information corresponding to the multiple first parameters.

2. The method according to claim 1, characterized in that The preset rules include: An abnormality judgment condition of the first parameter, the abnormality judgment condition including at least one of the following: a judgment condition of an abnormal event corresponding to the first parameter, and an upper limit of the time during which the first parameter meets the judgment condition; The abnormal event corresponding to the abnormal judgment condition.

3. The method according to claim 2, characterized in that The aggregating the target information according to the type of the first parameter corresponding to the target information to obtain first description information describing a change state of the first parameter includes: Determining an abnormal event corresponding to the target information; Merging the same target information of the abnormal events to obtain second description information of the abnormal events; The second description information is merged according to the type of the first parameter corresponding to the abnormal event to obtain the first description information.

4. The method according to claim 3, characterized in that For any first parameter; the second description information is merged according to the type of the first parameter corresponding to the abnormal event to obtain the first description information, including: Determine a plurality of first time periods corresponding to second description information of abnormal events of the same type as the first parameter, and a second time period corresponding to information of the first parameter; The first description information is determined based on the multiple first time periods, the second time period, and the second description information of the abnormal event of the same type as the first parameter.

5. The method according to claim 4, characterized in that The determining the first description information based on the multiple first time periods, the second time period, and the second description information of the abnormal event of the same type as the first parameter includes: When the multiple first time periods overlap the second time period, merging the second description information of the abnormal events of the same type as the first parameter based on the second time period to obtain the first description information; When the plurality of first time periods do not cover the second time period, the first description information is determined based on the plurality of first time periods, the second time period, and the second description information of the abnormal event of the same type as the first parameter.

6. The method according to claim 5, characterized in that The determining the first description information based on the plurality of first time periods, the second time period, and the second description information of the abnormal event of the same type as the first parameter includes: Based on the multiple first time periods, determining a third time period in the second time period that is not covered by the multiple first time periods; Obtaining information of the first parameter corresponding to the third time period; The second description information of the abnormal event of the same type as the first parameter and the information of the first parameter corresponding to the third time period are combined to obtain the first description information.

7. The method according to claim 1, characterized in that Analyzing the performance of the high-performance computing platform according to the plurality of first description information corresponding to the plurality of first parameters includes: Converting a plurality of description information corresponding to the plurality of first parameters into description texts in a natural language; According to the description text in the natural language, identifying knowledge fragments associated with the description text in the natural language in a preset knowledge base; The performance of the high-performance computing platform is analyzed according to the natural language description text and the knowledge fragment to obtain a performance analysis report of the high-performance computing platform.

8. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the performance analysis method for a high-performance computing platform as claimed in any one of claims 1 to 7 when executing the computer program.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the steps of the performance analysis method of the high-performance computing platform according to any one of claims 1 to 7 are implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the performance analysis method of the high-performance computing platform according to any one of claims 1 to 7 are implemented.

Images