System performance prediction model training method, system performance prediction method and device

Abstract

The present application discloses a system performance prediction model training method, a system performance prediction method, and an apparatus, relating to the fields of artificial intelligence technology and computing power cluster system performance prediction technology. The method comprises: obtaining sample feature data for training the system performance prediction model, the sample feature data including cluster feature data of a known computing power cluster and system performance data obtained by benchmarking the known computing power cluster; inputting the cluster feature data into the system performance prediction model to obtain output system performance prediction data, wherein the model structure of the system performance prediction model comprises a residual block stacking layer; determining whether the current model training meets the convergence condition based on the system performance prediction data and the corresponding system performance data; if the convergence condition is not met, adjusting the model parameters of the system performance prediction model and executing the next model training. Using this method, accurate prediction of the system performance of the computing power cluster is achieved.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence technology and computing cluster system performance prediction technology, and in particular to a system performance prediction model training method, a system performance prediction method and a device. Background Art

[0002] With the rapid development of artificial intelligence, especially in the field of large models, the computing power requirements of systems are increasing. Building a computing center typically requires thousands of processors, accelerator cards, and various supporting high-speed buses and other facilities. The overall performance of the system is affected by many factors, including processors, accelerator cards, communication methods, model algorithms, and scheduling algorithms. Failure to fully assess system performance and facility investment costs in the early stages of project construction can easily lead to wasted or insufficient computing power, both of which are obviously undesirable.

[0003] How to recommend a reasonable combination of software and hardware facilities based on actual application needs and combined with historical test data analysis to ensure the optimal cost and performance of the final system has become a topic of concern in the industry. Summary of the Invention

[0004] The embodiments of the present application provide a system performance prediction model training method, a system performance prediction method and an apparatus to solve the problem existing in the prior art of how to accurately predict the system performance of a computing power cluster.

[0005] The present invention provides a method for training a system performance prediction model, including:

[0006] Obtaining sample feature data for training a system performance prediction model, the sample feature data including cluster feature data of a known computing power cluster and system performance data obtained by benchmarking the known computing power cluster;

[0007] Inputting the cluster feature data into the system performance prediction model to obtain output system performance prediction data, wherein the model structure of the system performance prediction model has a residual block stacking layer, and the residual block stacking layer includes residual blocks;

[0008] Determining whether the current model training meets the convergence condition based on the system performance prediction data and the corresponding system performance data;

[0009] If the convergence condition is met, it is determined that the training of the system performance prediction model is completed; if the convergence condition is not met, the model parameters of the system performance prediction model are adjusted and the next model training is performed.

[0010] Furthermore, the cluster feature data includes: cluster quantitative feature data and cluster qualitative feature data;

[0011] The residual block stacking layer includes a plurality of residual blocks connected in series;

[0012] Each of the residual blocks has two inputs and one output;

[0013] The inputs of the plurality of residual blocks each include a qualitative feature vector representing the qualitative feature data of the cluster;

[0014] Another input of the first residual block is a quantitative feature vector representing the quantitative feature data of the cluster, another input of the residual blocks other than the first one is the output of the connected previous residual block, and the output of the last residual block serves as the output of the residual block stacking layer.

[0015] Furthermore, the operations performed in the residual block include the following operations:

[0016] Multiplying the qualitative eigenvector by the correlation matrix to obtain a correlation vector;

[0017] Adding the association vector to the quantitative feature vector or the output of the previous residual block to obtain a merged feature vector;

[0018] Multiplying the combined eigenvector by the residual matrix to obtain a residual vector;

[0019] Adding the residual vector to the quantitative feature vector or the output of the previous residual block to obtain a skip connection feature vector;

[0020] Normalizing the skip connection feature vector to obtain the output of the residual block;

[0021] The correlation matrix and the residual matrix serve as model parameters of the system performance prediction model.

[0022] Furthermore, the model structure of the system performance prediction model has an input layer and a feature preprocessing layer;

[0023] The input layer is used to receive the cluster quantitative feature data and the cluster qualitative feature data;

[0024] The feature preprocessing layer is used to preprocess the cluster quantitative feature data through a multi-layer perceptron MLP network to obtain a quantitative feature vector, and to generate a qualitative feature vector corresponding to the cluster qualitative feature data by using a label code lookup method.

[0025] Furthermore, the model structure of the system performance prediction model has a summary layer and an output layer;

[0026] The summary layer is used to process the output of the residual block stacking layer through the MLP network to obtain the system performance prediction data;

[0027] The output layer is used to output the system performance prediction data.

[0028] The present application also provides a method for predicting the performance of a computing cluster system, including:

[0029] Obtain cluster feature data of the computing power cluster to be predicted;

[0030] Based on the cluster feature data, a system performance prediction model trained by any of the above-mentioned system performance prediction model training methods is used to predict the system performance of the computing power cluster to be predicted to obtain system performance prediction data.

[0031] The present application also provides a system performance prediction model training device, including:

[0032] A sample data acquisition module is used to acquire sample feature data for training a system performance prediction model, wherein the sample feature data includes cluster feature data of a known computing power cluster and system performance data obtained by benchmarking the known computing power cluster;

[0033] a system performance prediction module, configured to input the cluster feature data into the system performance prediction model to obtain output system performance prediction data, wherein the model structure of the system performance prediction model comprises a residual block stacking layer, and the residual block stacking layer comprises residual blocks;

[0034] A convergence judgment module is used to determine whether the current model training meets the convergence condition based on the system performance prediction data and the corresponding system performance data;

[0035] The model training module is used to determine that the training of the system performance prediction model is completed if the convergence condition is met, and if the convergence condition is not met, adjust the model parameters of the system performance prediction model and perform the next model training.

[0036] Furthermore, the cluster feature data includes: cluster quantitative feature data and cluster qualitative feature data;

[0037] The residual block stacking layer includes a plurality of residual blocks connected in series;

[0038] Each of the residual blocks has two inputs and one output;

[0039] The inputs of the plurality of residual blocks each include a qualitative feature vector representing the qualitative feature data of the cluster;

[0040] Another input of the first residual block is a quantitative feature vector representing the quantitative feature data of the cluster, another input of the residual blocks other than the first one is the output of the connected previous residual block, and the output of the last residual block serves as the output of the residual block stacking layer.

[0041] Furthermore, the operations performed in the residual block include the following operations:

[0042] Multiplying the qualitative eigenvector by the correlation matrix to obtain a correlation vector;

[0043] Adding the association vector to the quantitative feature vector or the output of the previous residual block to obtain a merged feature vector;

[0044] Multiplying the combined eigenvector by the residual matrix to obtain a residual vector;

[0045] Adding the residual vector to the quantitative feature vector or the output of the previous residual block to obtain a skip connection feature vector;

[0046] Normalizing the skip connection feature vector to obtain the output of the residual block;

[0047] The correlation matrix and the residual matrix serve as model parameters of the system performance prediction model.

[0048] Furthermore, the model structure of the system performance prediction model has an input layer and a feature preprocessing layer;

[0049] The input layer is used to receive the cluster quantitative feature data and the cluster qualitative feature data;

[0050] The feature preprocessing layer is used to preprocess the cluster quantitative feature data through a multi-layer perceptron MLP network to obtain a quantitative feature vector, and to generate a qualitative feature vector corresponding to the cluster qualitative feature data by using a label code lookup method.

[0051] Furthermore, the model structure of the system performance prediction model has a summary layer and an output layer;

[0052] The summary layer is used to process the output of the residual block stacking layer through the MLP network to obtain the system performance prediction data;

[0053] The output layer is used to output the system performance prediction data.

[0054] The present application also provides a computing power cluster system performance prediction device, including:

[0055] Cluster data acquisition module, used to obtain cluster feature data of the computing power cluster to be predicted;

[0056] The system performance prediction module is used to predict the system performance of the computing power cluster to be predicted based on the cluster feature data and the system performance prediction model trained by the above-mentioned system performance prediction model training device to obtain system performance prediction data.

[0057] An embodiment of the present application also provides an electronic device, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions that can be executed by the processor, and the processor is prompted by the machine-executable instructions to: implement any of the above-mentioned system performance prediction model training methods, or implement a computing power cluster system performance prediction method.

[0058] An embodiment of the present application also provides a computer-readable storage medium, characterized in that a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, it implements any of the above-mentioned system performance prediction model training methods, or implements a computing power cluster system performance prediction method.

[0059] An embodiment of the present application also provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute any of the above-mentioned system performance prediction model training methods, or to execute a computing power cluster system performance prediction method.

[0060] The beneficial effects of this application include:

[0061] In the method provided in the embodiment of the present application, cluster feature data of a known computing power cluster is used as the input of a system performance prediction model, and the system performance prediction data output by the system performance prediction model is compared with the system performance data obtained by benchmarking the known computing power cluster to determine whether the current model training meets the convergence conditions. Through repeated iterative training until the convergence conditions are met, a trained system performance prediction model is obtained. In addition, the model structure of the system performance prediction model has a residual block stacking layer, and the residual block stacking layer includes residual blocks, so that the system performance prediction model can be used to accurately predict the system performance of the computing power cluster to be predicted based on the cluster feature data of the computing power cluster to be predicted.

[0062] Other features and advantages of the present application will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present application. The purposes and other advantages of the present application can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings are used to provide a further understanding of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings:

[0064] Figure 1 A flowchart of a system performance prediction model training method provided in an embodiment of the present application;

[0065] Figure 2 A flowchart of a computing cluster system performance prediction method provided in an embodiment of the present application;

[0066] Figure 3 This is a schematic diagram of the structure of the system performance prediction model in the embodiment of the present application;

[0067] Figure 4 Schematic diagram of the structure of the residual block stacking layer of the system performance prediction model in an embodiment of the present application;

[0068] Figure 5 Schematic diagram of the structure of the residual block of the residual block stacking layer in an embodiment of the present application;

[0069] Figure 6 A schematic diagram of the structure of a system performance prediction model training device provided in an embodiment of the present application;

[0070] Figure 7 A schematic diagram of the structure of a computing power cluster system performance prediction device provided in an embodiment of the present application;

[0071] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0072] To provide an implementation solution for accurately predicting the system performance of a computing cluster, the present application provides a system performance prediction model training method, a system performance prediction method, and an apparatus. The preferred embodiments of the present application are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are intended only to illustrate and explain the present application and are not intended to limit the present application. Furthermore, the embodiments and features within the embodiments of the present application may be combined with each other unless there is a conflict.

[0073] The present application embodiment provides a system performance prediction model training method, such as Figure 1 As shown, including:

[0074] Step 11: Obtain sample feature data for training a system performance prediction model. The sample feature data includes cluster feature data of a known computing power cluster and system performance data obtained by benchmarking the known computing power cluster.

[0075] Step 12: input the cluster feature data into a system performance prediction model to obtain output system performance prediction data, wherein the model structure of the system performance prediction model has a residual block stacking layer, and the residual block stacking layer includes residual blocks;

[0076] Step 13: Based on the system performance prediction data and the corresponding system performance data, determine whether the current model training meets the convergence conditions;

[0077] Step 14: If the convergence condition is met, determine that the training of the system performance prediction model is completed; if the convergence condition is not met, adjust the model parameters of the system performance prediction model and perform the next model training.

[0078] Accordingly, the present application also provides a method for predicting the performance of a computing cluster system. Figure 2 As shown, including:

[0079] 21. Obtain cluster feature data of the computing power cluster to be predicted;

[0080] 22. Based on the cluster feature data, the system performance prediction model trained by the above-mentioned system performance prediction model training method is used to predict the system performance of the computing power cluster to be predicted, and obtain system performance prediction data.

[0081] The above method provided in the embodiment of the present application is adopted, and the cluster feature data of the known computing power cluster is used as the input of the system performance prediction model. The system performance prediction data output by the system performance prediction model is compared with the system performance data obtained by benchmarking the known computing power cluster to determine whether the current model training meets the convergence conditions. Through repeated iterative training until the convergence conditions are met, a trained system performance prediction model is obtained. In addition, the model structure of the system performance prediction model has a residual block stacking layer, and the residual block stacking layer includes residual blocks, so that the system performance prediction model can be used to accurately predict the system performance of the computing power cluster to be predicted based on the cluster feature data of the computing power cluster to be predicted.

[0082] The method and device provided in this application are described in detail below with reference to specific embodiments in conjunction with the accompanying drawings.

[0083] In the embodiment of the present application, the model structure of the system performance prediction model has a residual block stacking layer, such as Figure 3 As shown, it can also have an input layer, a feature preprocessing layer, a summary layer, and an output layer. The connection relationship between the layers is as follows Figure 3 As shown in Figure 1, the input layer, feature preprocessing layer, residual block stacking layer, summary layer, and output layer are connected in sequence.

[0084] In the embodiment of the present application, the cluster feature data of the computing power cluster used as the model input may include cluster quantitative feature data and cluster qualitative feature data;

[0085] The cluster quantitative feature data may include: number of processors, number of accelerator cards, memory size, main frequency, etc., which are generally numerical data.

[0086] Cluster qualitative feature data is feature data that cannot be quantified numerically. Specifically, it may include: processor model, accelerator card model, chip architecture, process technology, instruction set type, etc., and is generally an enumeration type.

[0087] In one embodiment of the present application, the input layer is used to receive cluster quantitative feature data and cluster qualitative feature data;

[0088] The feature preprocessing layer is used to preprocess the cluster quantitative feature data output by the input layer through the MLP network to obtain a quantitative feature vector. Specifically, all cluster quantitative feature data can be merged into an initial numerical vector. The vector length of the numerical vector is determined by the number of quantitative indicators actually selected. Then, the numerical vector is preprocessed through the MLP network to obtain a 1*n quantitative feature vector.

[0089] The feature preprocessing layer is also used to generate qualitative feature vectors corresponding to the cluster qualitative feature data by using the label code lookup method. Specifically, the label code lookup method is used to directly assign corresponding values ​​to the types represented by each cluster qualitative feature data to obtain a 1*n label vector. It can be randomly assigned at the beginning, as long as the values ​​assigned to the same type subsequently remain consistent. This label vector is the qualitative feature vector.

[0090] Through the input layer and the feature preprocessing layer, the input of the residual block stacking layer, namely the cluster quantitative feature vector and the cluster qualitative feature vector, is obtained.

[0091] In one embodiment of the present application, the residual block stacking layer may include a residual block such as Figure 4 As shown, the residual block stacking layer may also include multiple residual blocks connected in series, and each residual block has two inputs and one output, the inputs of the multiple residual blocks include qualitative feature vectors representing qualitative feature data of the cluster, another input of the first residual block is a quantitative feature vector representing quantitative feature data of the cluster, another input of other residual blocks except the first one is the output of the previous residual block connected, and the output of the last residual block is used as the output of the residual block stacking layer.

[0092] Further, such as Figure 5 As shown, the operations performed in the residual block include the following operations:

[0093] Multiplying the qualitative eigenvector by the correlation matrix to obtain a correlation vector, wherein the correlation matrix is ​​an n*n matrix and the obtained correlation vector is a 1*n vector;

[0094] Add the correlation vector to the quantitative feature vector or the output of the previous residual block to obtain a merged feature vector. In this operation, if the residual block is the first residual block, it is added to the quantitative feature vector. If it is a residual block other than the first residual block, it is added to the output of the previous residual block. The resulting merged feature vector is a 1*n vector, thereby achieving the merging of quantitative features and qualitative features.

[0095] Multiply the merged eigenvector by the residual matrix to obtain a residual vector, where the residual matrix is ​​an n*n matrix and the obtained residual vector is a 1*n vector;

[0096] The residual vector is added to the quantitative feature vector or the output of the previous residual block to obtain a skip connection feature vector. In this operation, if the residual block is the first residual block, it is added to the quantitative feature vector. If it is a residual block other than the first residual block, it is added to the output of the previous residual block. The resulting skip connection feature vector is a 1*n vector, thereby realizing a skip connection between the vector residual vector and the quantitative feature vector or the output of the previous residual block;

[0097] Normalize the skip connection feature vector to obtain the output of the residual block;

[0098] Among them, the correlation matrix and the residual matrix are model parameters of the system performance prediction model, which will be adjusted through continuous iteration during the model training process.

[0099] In one embodiment of the present application, Figure 3 As shown, the model structure of the system performance prediction model has a summary layer and an output layer;

[0100] The aggregation layer is used to process the output of the residual block stacking layer through the MLP network to obtain system performance prediction data. Specifically, the n-dimensional feature vector can be processed into an m-dimensional performance indicator vector, where m is the number of system performance indicators. The performance indicators can be temperature, power consumption, throughput, inference latency, and inference accuracy.

[0101] The output layer is used to output system performance prediction data. The output layer can be one or more neurons, corresponding to the system performance indicators that need to be output.

[0102] In the embodiment of the present application, during the model training process, after obtaining the output system performance prediction data, it is determined whether the current model training meets the convergence condition based on the system performance prediction data and the corresponding system performance data obtained by the benchmark test;

[0103] In one embodiment of the present application, specifically, the mean square error (MSE) can be used as the loss function to train the model, that is, to determine whether the current model training has converged. The formula is as follows:

[0104] L=1 / N*Σ(Y_pred - Y_true)^2, where Y_pred is the system performance prediction data output by the model, Y_true is the actual system performance data obtained from the benchmark test, and N is the number of samples;

[0105] Accordingly, if the convergence condition is met, the training of the system performance prediction model is determined to be completed; if the convergence condition is not met, the model parameters of the system performance prediction model are adjusted, and the next model training is performed;

[0106] Specifically, gradient descent or its variants (such as Adam optimizer) can be used to minimize the loss function and update the model parameters, mainly including the parameters of the MLP network of the feature preprocessing layer, the residual matrix and correlation matrix of the residual block stacking layer, and the parameters of the MLP network of the summary layer.

[0107] Based on the same inventive concept, according to the system performance prediction model training method provided in the above embodiment of the present application, correspondingly, another embodiment of the present application also provides a system performance prediction model training device, the structural diagram of which is shown in FIG. Figure 6 As shown, specifically including:

[0108] A sample data acquisition module 61 is configured to acquire sample feature data for training a system performance prediction model, wherein the sample feature data includes cluster feature data of a known computing power cluster and system performance data obtained by performing a benchmark test on the known computing power cluster;

[0109] a system performance prediction module 62 configured to input the cluster feature data into the system performance prediction model to obtain output system performance prediction data, wherein the system performance prediction model has a model structure comprising a residual block stacking layer, wherein the residual block stacking layer comprises residual blocks;

[0110] A convergence judgment module 63 is used to determine whether the current model training meets the convergence condition based on the system performance prediction data and the corresponding system performance data;

[0111] The model training module 64 is configured to determine that the training of the system performance prediction model is completed if the convergence condition is met, and to adjust the model parameters of the system performance prediction model and perform the next model training if the convergence condition is not met.

[0112] Furthermore, the cluster feature data includes: cluster quantitative feature data and cluster qualitative feature data;

[0113] The residual block stacking layer includes a plurality of residual blocks connected in series;

[0114] Each of the residual blocks has two inputs and one output;

[0115] The inputs of the plurality of residual blocks each include a qualitative feature vector representing the qualitative feature data of the cluster;

[0116] Another input of the first residual block is a quantitative feature vector representing the quantitative feature data of the cluster, another input of the residual blocks other than the first one is the output of the connected previous residual block, and the output of the last residual block serves as the output of the residual block stacking layer.

[0117] Furthermore, the operations performed in the residual block include the following operations:

[0118] Multiplying the qualitative eigenvector by the correlation matrix to obtain a correlation vector;

[0119] Adding the association vector to the quantitative feature vector or the output of the previous residual block to obtain a merged feature vector;

[0120] Multiplying the combined eigenvector by the residual matrix to obtain a residual vector;

[0121] Adding the residual vector to the quantitative feature vector or the output of the previous residual block to obtain a skip connection feature vector;

[0122] Normalizing the skip connection feature vector to obtain the output of the residual block;

[0123] The correlation matrix and the residual matrix serve as model parameters of the system performance prediction model.

[0124] Furthermore, the model structure of the system performance prediction model has an input layer and a feature preprocessing layer;

[0125] The input layer is used to receive the cluster quantitative feature data and the cluster qualitative feature data;

[0126] The feature preprocessing layer is used to preprocess the cluster quantitative feature data through a multi-layer perceptron MLP network to obtain a quantitative feature vector, and to generate a qualitative feature vector corresponding to the cluster qualitative feature data by using a label code lookup method.

[0127] Furthermore, the model structure of the system performance prediction model has a summary layer and an output layer;

[0128] The summary layer is used to process the output of the residual block stacking layer through the MLP network to obtain the system performance prediction data;

[0129] The output layer is used to output the system performance prediction data.

[0130] Based on the same inventive concept, according to the computing power cluster system performance prediction method provided in the above embodiment of the present application, correspondingly, another embodiment of the present application also provides a computing power cluster system performance prediction device, whose structural diagram is shown as follows: Figure 7 As shown, specifically including:

[0131] Cluster data acquisition module 71, used to obtain cluster feature data of the computing power cluster to be predicted;

[0132] The system performance prediction module 72 is used to predict the system performance of the computing power cluster to be predicted based on the cluster feature data and the system performance prediction model trained by the above-mentioned system performance prediction model training device to obtain system performance prediction data.

[0133] The functions of the above modules can correspond to Figure 1 and Figure 2 The corresponding processing steps in the shown process will not be repeated here.

[0134] The system performance prediction model training device and computing power cluster system performance prediction device provided in the embodiments of this application can be implemented via computer programs. Those skilled in the art should understand that the aforementioned module division method is only one of many module division methods. If the system performance prediction model training device and computing power cluster system performance prediction device are divided into other modules or not divided into modules, as long as they have the aforementioned functions, they should be within the scope of protection of this application.

[0135] The present application also provides an electronic device, such as Figure 8 As shown, it includes a processor 81 and a machine-readable storage medium 82, and the machine-readable storage medium 82 stores machine-executable instructions that can be executed by the processor 81. The processor 81 is prompted by the machine-executable instructions to: implement any of the above-mentioned system performance prediction model training methods, or implement a computing power cluster system performance prediction method.

[0136] An embodiment of the present application also provides a computer-readable storage medium, characterized in that a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, it implements any of the above-mentioned system performance prediction model training methods, or implements a computing power cluster system performance prediction method.

[0137] An embodiment of the present application also provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute any of the above-mentioned system performance prediction model training methods, or to execute a computing power cluster system performance prediction method.

[0138] The machine-readable storage medium in the electronic device may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Alternatively, the memory may be at least one storage device located remotely from the processor.

[0139] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0140] Each embodiment in this specification is described in a related manner. Similar portions between the various embodiments can be referenced to each other. Each embodiment focuses on the differences between the other embodiments. In particular, the device, electronic device, computer-readable storage medium, and computer program product embodiments are generally similar to the method embodiments, so their descriptions are relatively simple. For related portions, reference can be made to the descriptions of the method embodiments.

[0141] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so 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. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0142] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0143] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0144] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0145] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A system performance prediction model training method, characterized in that: include: Obtaining sample feature data for training a system performance prediction model, the sample feature data including cluster feature data of a known computing power cluster and system performance data obtained by benchmarking the known computing power cluster, the cluster feature data including: cluster quantitative feature data and cluster qualitative feature data; Inputting the cluster feature data into the system performance prediction model to obtain output system performance prediction data, wherein the model structure of the system performance prediction model has a residual block stacking layer, and the residual block stacking layer includes multiple residual blocks connected in series; each of the residual blocks has two inputs and one output; the inputs of the multiple residual blocks all include qualitative feature vectors representing the qualitative feature data of the cluster; another input of the first residual block is a quantitative feature vector representing the quantitative feature data of the cluster, and another input of the other residual blocks except the first one is the output of the previous residual block connected to it, and the output of the last residual block is used as the output of the residual block stacking layer; Determining whether the current model training meets the convergence condition based on the system performance prediction data and the corresponding system performance data; If the convergence condition is met, determining that the training of the system performance prediction model is completed; if the convergence condition is not met, adjusting the model parameters of the system performance prediction model and performing the next model training; The operations performed in the residual block include the following operations: Multiplying the qualitative eigenvector by the correlation matrix to obtain a correlation vector; Adding the association vector to the quantitative feature vector or the output of the previous residual block to obtain a merged feature vector; Multiplying the combined eigenvector by the residual matrix to obtain a residual vector; Adding the residual vector to the quantitative feature vector or the output of the previous residual block to obtain a skip connection feature vector; Normalizing the skip connection feature vector to obtain the output of the residual block; The correlation matrix and the residual matrix serve as model parameters of the system performance prediction model.

2. The method according to claim 1, wherein The model structure of the system performance prediction model comprises an input layer and a feature preprocessing layer; The input layer is used to receive the cluster quantitative feature data and the cluster qualitative feature data; The feature preprocessing layer is used to preprocess the cluster quantitative feature data through a multi-layer perceptron MLP network to obtain a quantitative feature vector, and to generate a qualitative feature vector corresponding to the cluster qualitative feature data by using a label code lookup method.

3. The method according to claim 1, wherein The model structure of the system performance prediction model has a summary layer and an output layer; The summary layer is used to process the output of the residual block stacking layer through the MLP network to obtain the system performance prediction data; The output layer is used to output the system performance prediction data.

4. A system performance prediction method, characterized in that: include: Obtain cluster feature data of the computing power cluster to be predicted; Based on the cluster feature data, a system performance prediction model obtained by training using the method described in any one of claims 1 to 3 is used to predict the system performance of the computing power cluster to be predicted, thereby obtaining system performance prediction data.

5. A system performance prediction model training device, characterized in that: include: A sample data acquisition module is configured to acquire sample feature data for training a system performance prediction model, wherein the sample feature data includes cluster feature data of a known computing power cluster and system performance data obtained by benchmarking the known computing power cluster. The cluster feature data includes: cluster quantitative feature data and cluster qualitative feature data; A system performance prediction module is configured to input the cluster feature data into the system performance prediction model to obtain output system performance prediction data, wherein the model structure of the system performance prediction model comprises a residual block stacking layer, the residual block stacking layer comprises a plurality of residual blocks connected in series; each residual block has two inputs and one output; the inputs of the plurality of residual blocks each comprise a qualitative feature vector representing the qualitative feature data of the cluster; another input of the first residual block is a quantitative feature vector representing the quantitative feature data of the cluster; another input of the residual blocks other than the first one is the output of the connected previous residual block; the output of the last residual block serves as the output of the residual block stacking layer; The operations performed in the residual block include the following operations: Multiplying the qualitative eigenvector by the correlation matrix to obtain a correlation vector; Adding the association vector to the quantitative feature vector or the output of the previous residual block to obtain a merged feature vector; Multiplying the combined eigenvector by the residual matrix to obtain a residual vector; Adding the residual vector to the quantitative feature vector or the output of the previous residual block to obtain a skip connection feature vector; Normalizing the skip connection feature vector to obtain the output of the residual block; Wherein, the correlation matrix and the residual matrix serve as model parameters of the system performance prediction model; A convergence judgment module is used to determine whether the current model training meets the convergence condition based on the system performance prediction data and the corresponding system performance data; The model training module is used to determine that the training of the system performance prediction model is completed if the convergence condition is met, and if the convergence condition is not met, adjust the model parameters of the system performance prediction model and perform the next model training.

6. A system performance prediction device, characterized in that: include: Cluster data acquisition module, used to obtain cluster feature data of the computing power cluster to be predicted; A system performance prediction module is used to predict the system performance of the computing power cluster to be predicted based on the cluster feature data and the system performance prediction model obtained by training the device according to claim 5 to obtain system performance prediction data.

7. An electronic device, characterized in that: The method comprises a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions that can be executed by the processor, and the processor is prompted by the machine-executable instructions to implement the method according to any one of claims 1 to 3, or to implement the method according to claim 4.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 3 is implemented, or the method according to claim 4 is implemented.

Images