AI large model-based computing power resource regulation and control optimization method and system, and storage medium

Abstract

The invention discloses a computing power resource regulation and control optimization method and system based on an AI large model and a storage medium, and relates to the technical field of computing power resource regulation and control, and the method comprises the following steps: collecting historical operation data of each node of a computing power cluster, and constructing a first data set based on the historical operation data; the method comprises the following steps: performing improvement on the basis of a Transform large model, constructing a first computing power prediction model, setting an evaluation index, and performing optimization training on the first computing power prediction model to obtain a second computing power prediction model; predicting the computing power resource demand based on a second computing power prediction model to obtain a prediction result; a regulation and control optimization instruction is generated based on the prediction result, and calculation power resource regulation and control optimization is completed based on the regulation and control optimization instruction; the computing power resource regulation and control method is used for solving the problem that the existing computing power resource regulation and control technology cannot predict future computing power resource data under the condition of considering the influence of other dimension resources and cannot utilize the current and future computing power resource states to regulate and control and optimize the computing power resources at the same time.

Description

Technical Field

[0001] The present invention relates to the field of computing power resource regulation technology, and specifically to a computing power resource regulation optimization method, system and storage medium based on an AI big model. Background Art

[0002] Computing resource control technology refers to dynamically adjusting and optimizing the configuration and use of computing resources through a series of methods and strategies to meet real-time changing computing power requirements and improve resource utilization and computing performance. The core purpose of computing resource control technology is to provide appropriate computing resources at the right time and place, thereby ensuring that various computing tasks can run efficiently and stably.

[0003] The existing computing resource control technology is based on the traditional cluster scheduling method, and optimizes the computing task characteristics and computing resources in real time through the optimization algorithm to improve the computing efficiency or computing resource utilization. The optimization algorithm can only obtain the current and historical computing resource status, and many computing resource scheduling strategies only make decisions based on the current status. However, due to factors such as algorithm calculation delay and network transmission delay, the computing resource status input by the algorithm may be very different from the actual status when the computing node provides services. This leads to limited optimization effects on computing resource control. For example, in the patent application with publication number CN115277598A, a A computing power resource scheduling method, device and computer-readable storage medium. This solution optimizes computing power resource regulation by obtaining the characteristics of computing tasks and their own computing power resources. It does not consider the impact of computing delay, network transmission delay and other factors. The degree of optimization of computing power resource regulation is limited, and it is not smart enough and lacks robustness. At the same time, existing computing power resource regulation technologies regard GPU and CPU hardware resources as the main factors affecting computing power resources, ignoring the impact of other dimensional resources such as memory and network. It is impossible to predict future computing power resource data while considering the impact of other dimensional resources, and at the same time use the current and future computing power resource status to optimize computing power resource regulation. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems in the prior art to a certain extent. It improves the Transformer large model and performs optimization training to obtain a second computing power prediction model. The second computing power prediction model predicts future computing power resources by integrating other dimensional resources to obtain prediction results, and optimizes the computing power resource regulation by integrating the current and future computing power resource states. This solves the problem that it is impossible to predict future computing power resource data while considering the influence of other dimensional resources, and at the same time use the current and future computing power resource states to optimize the computing power resource regulation.

[0005] To achieve the above objectives, in a first aspect, the present application provides a computing resource control optimization method based on an AI big model, comprising the following steps:

[0006] Collect historical operation data of each node in the computing power cluster, and build a first data set based on the historical operation data;

[0007] Improvements were made based on the Transformer large model, the first computing power prediction model was built, and evaluation indicators were set;

[0008] Optimize and train the first computing power prediction model based on the first data set and the evaluation index to obtain a second computing power prediction model;

[0009] Predicting the computing power resource demand based on the second computing power prediction model to obtain a prediction result;

[0010] Generate control and optimization instructions based on the prediction results, and complete the control and optimization of computing power resources based on the control and optimization instructions.

[0011] Furthermore, collecting the historical operation data of each node in the computing power cluster includes the following sub-steps:

[0012] The historical operation data of each node in the computing power cluster is obtained once every first time interval within a first time length; the first time length is set to T, the first time interval is set to a, and the historical operation data includes GPU usage, CPU usage, memory usage and network bandwidth IO occupancy of each node in the computing power cluster.

[0013] Furthermore, constructing the first data set based on the historical operation data includes the following sub-steps:

[0014] The historical operation data are normalized according to different data types, and a time label is bound to each historical operation data according to the time when each historical operation data is obtained. Each time label includes the GPU usage, CPU usage, memory usage and network bandwidth IO occupancy at the corresponding time. They are stored in the time order on the time label and marked as the first data set.

[0015] Furthermore, based on the Transformer large model, improvements are made to construct the first computing power prediction model, which includes the following sub-steps:

[0016] Based on the Transformer large model, the Transformer large model is combined with the convolutional neural network, the convolutional self-attention mechanism calculation of the convolutional neural network is used to replace the self-attention mechanism calculation used by the Transformer large model, and the logarithmic sparsity strategy is added to the convolutional self-attention mechanism calculation. The improved Transformer large model is marked as the first computing power prediction model.

[0017] Furthermore, setting the evaluation index includes the following sub-steps:

[0018] The mean absolute percentage error and root mean square error are set as evaluation indicators. The mean absolute percentage error formula is: The root mean square error formula is:

[0019]

[0020] Where MAPE stands for mean absolute percentage error; RMSE stands for root mean square error; Y t Represents the actual usage at time t; Represents the usage predicted by the model at time t, including GPU usage, CPU usage, memory usage, and network bandwidth IO usage; N represents the number of data of the same type;

[0021] The qualified parameters m and r are set based on the evaluation index.

[0022] Further, optimizing and training the first computing power prediction model based on the first data set and the evaluation index to obtain the second computing power prediction model includes the following sub-steps:

[0023] The first data set is divided into a training set and a test set in a ratio of 8:2;

[0024] Set the training hyperparameters of the first computing power prediction model, set the learning rate to E, the batch size to B, and the number of model training rounds to P;

[0025] Set mean square error (MSE) as the loss function and use Adam optimizer as the model training optimizer;

[0026] The first computing power prediction model is trained using the training set according to the set training hyperparameters. After each P round of training, the first computing power prediction model is tested using the test set based on the evaluation indicators, and the evaluation results M and R are output. When M and R satisfy M≤m and R≤r, the first computing power prediction model is saved and marked as the second computing power prediction model; if M and R do not satisfy M≤m and R≤r, the first computing power prediction model is trained again using the training set according to the set training hyperparameters until M and R satisfy M≤m and R≤r to obtain the second computing power prediction model.

[0027] Further, predicting the computing power resource demand based on the second computing power prediction model to obtain the prediction result includes the following sub-steps:

[0028] Get a section of historical operation data of each node in the computing cluster, including GPU usage, CPU usage, memory usage, and network bandwidth IO usage of each node in the computing cluster;

[0029] Normalize the historical operation data and bind a time label to each data;

[0030] The historical operation data with the bound time label is input into the second computing power prediction model to obtain the operation data of each node in the computing power cluster within the second time length. The second time length is T2, which is marked as predicted operation data. The predicted operation data includes the predicted GPU usage, CPU usage, memory usage and network bandwidth IO usage of each node in the computing power cluster.

[0031] Furthermore, a control optimization instruction is generated based on the prediction result, and the computing power resource control optimization is completed based on the control optimization instruction, including the following sub-steps:

[0032] Get the maximum computing power resource F of all computing power nodes in the computing power cluster max The maximum computing resources include the maximum GPU usage, the maximum CPU usage, the maximum memory usage and the maximum network bandwidth IO.

[0033] Get the current operation data of the computing cluster, including the GPU usage, CPU usage, memory usage, and network bandwidth IO usage of each node in the computing cluster, and get the computing resources F currently used by the computing cluster o ; o represents the current moment;

[0034] Get future computing power resources F based on predicted operation data i , i represents the i-th moment, F i Including the predicted GPU usage, CPU usage, memory usage and network bandwidth IO usage of each node in the computing cluster at time i;

[0035] Generate control optimization instructions based on the maximum computing power resources Fmax, used computing power resources Fo and future computing power resources Fi of the computing power cluster: subtract the maximum computing power resources from the used computing power resources to obtain the current available computing power resources Fx. When the future computing power resources Fi exceed the current available computing power resources Fx, automatically increase computing power resources and expand the maximum computing power resources, including increasing computing power nodes, expanding memory capacity and increasing network bandwidth; if the current available computing power resources Fx are greater than the future computing power resources Fi, automatically reduce computing power resources, including reducing computing power nodes, reducing memory capacity and reducing network bandwidth;

[0036] The computing power cluster is controlled based on the control and optimization instructions to complete the control and optimization of computing power resources.

[0037] In the second aspect, the present application provides a computing resource control and optimization system based on an AI big model, including a data management module, a model building module, a computing power prediction module, and a computing power control module;

[0038] The data management module includes a collection unit and a processing unit, the collection unit is used to collect historical operation data of each node in the computing power cluster, and the processing unit is used to process the historical operation data and construct a first data set based on the historical operation data;

[0039] The model construction module includes a construction unit and a training unit. The construction unit is used to improve the Transformer large model, construct a first computing power prediction model, and set evaluation indicators; the training unit optimizes and trains the first computing power prediction model based on the first data set and the evaluation indicators to obtain a second computing power prediction model;

[0040] The computing power prediction module includes a second computing power prediction model, which predicts the computing power resource demand based on the first data set to obtain a prediction result;

[0041] The computing power control module includes a generation unit and an execution unit. The generation unit generates a control optimization instruction based on the prediction result, and the execution unit completes the computing power resource control optimization based on the control optimization instruction.

[0042] In a third aspect, the present application provides a storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps in the above method are performed.

[0043] Beneficial effects of the present invention: The present invention collects historical operation data of each node in the computing power cluster, and constructs a first data set based on the historical operation data; then improves based on the Transformer large model, constructs a first computing power prediction model, and sets evaluation indicators, and then optimizes and trains the first computing power prediction model based on the first data set and the evaluation indicators to obtain a second computing power prediction model; then predicts computing power resource demand based on the second computing power prediction model to obtain a prediction result; finally, generates a control optimization instruction based on the prediction result, and completes computing power resource control optimization based on the control optimization instruction; the present invention uses the second computing power prediction model to predict computing power resource demand, and can simultaneously use the current and future computing power resource states to optimize computing power resource control while considering the influence of other dimensional resources;

[0044] The present invention is based on the improvement and optimization training of the Transformer large model. The Transformer algorithm can not only predict complex data changes, but also is good at processing long sequence problems. It is suitable for predicting computing resources with the characteristics of fast change frequency and unclear change rules, thereby improving the accuracy of regulation and optimization. In addition, in view of the problems existing in the prediction of Transformer, the logarithmic sparsity strategy and convolutional self-attention mechanism calculation are used for improvement, thereby improving the spatial efficiency and prediction effect of regulation and optimization. At the same time, in the prediction of future computing power demand and the formulation of regulation and optimization strategies, the GPU usage, CPU usage, memory usage and network bandwidth IO occupancy are comprehensively considered, thereby improving the accuracy of the prediction of future computing power demand, thereby improving the accuracy of regulation and optimization, and enhancing robustness and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a functional block diagram of the system of the present invention;

[0046] Figure 2 is a flow chart of the steps of the method of the present invention;

[0047] Figure 3 It is a structural framework diagram of the Transformer model of the present invention;

[0048] Figure 4 Schematic diagram of the shortcomings of the self-attention mechanism of the present invention; DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] Example 1, please refer to Figure 1 As shown, in the first aspect, the present application provides a computing resource control and optimization system based on an AI big model, including a data management module, a model building module, a computing power prediction module, and a computing power control module;

[0051] The data management module includes a collection unit and a processing unit, the collection unit is used to collect historical operation data of each node of the computing power cluster, and the processing unit is used to process the historical operation data and construct a first data set based on the historical operation data;

[0052] The collection unit is configured with a collection strategy, which includes: acquiring historical operation data of each node in the computing power cluster once at a first time interval within a first time length; the first time length is set to T, the first time interval is set to a, and the historical operation data includes GPU usage, CPU usage, memory usage, and network bandwidth IO occupancy of each node in the computing power cluster;

[0053] In this embodiment, the first time length T for obtaining historical operation data is 20 days, and the first time interval a is 10 minutes, that is, the PU usage, CPU usage, memory usage, and network bandwidth IO usage of each node in the computing power cluster are obtained every 10 minutes;

[0054] The processing unit is configured with a processing strategy, which includes: normalizing the historical operation data according to different data types, binding a time tag to each historical operation data according to the time when each historical operation data is acquired, each time tag includes the GPU usage, CPU usage, memory usage and network bandwidth IO usage at the corresponding time, storing them in the time order on the time tag, and marking them as the first data set; normalization processing includes: using min-max normalization for the same type of data, scaling all data sizes to [0, 1] to eliminate the impact of different dimensions on subsequent processing, and the min-max normalization formula is: Where x max is the maximum value in the data, x min is the minimum value in the data, x i is the original value of the data, y i is the normalized value;

[0055] In the specific implementation process, the demand for computing power resources has certain regular characteristics on the time scale. The prediction of computing power resource demand belongs to the field of time series data prediction. Time series data is data arranged in chronological order. The operation data of each computing power node in the computing power cluster is collected at a fixed period, and a time label is bound to each operation data. It is stored in the database in chronological order. Therefore, the operation data is a typical time series data. Time series data prediction can help make decisions in advance, optimize resource allocation, reduce risks, and improve efficiency; comprehensively considering GPU usage, CPU usage, memory usage, and network bandwidth IO occupancy can more accurately predict future computing power needs, and then more accurately regulate and optimize computing power resources to improve the utilization efficiency of computing power resources.

[0056] The model construction module includes a construction unit and a training unit. The construction unit is used to improve the Transformer large model, construct a first computing power prediction model, and set evaluation indicators. The training unit optimizes and trains the first computing power prediction model based on the first data set and the evaluation indicators to obtain a second computing power prediction model.

[0057] The construction unit is configured with a construction strategy, which includes: based on the Transformer large model, combining the Transformer large model with the convolutional neural network, replacing the self-attention mechanism calculation used by the Transformer large model with the convolutional self-attention mechanism calculation of the convolutional neural network, adding a logarithmic sparsity strategy to the convolutional self-attention mechanism calculation, and marking the improved Transformer large model as the first computing power prediction model;

[0058] See also Figure 3 As shown in the figure, Transformer is an encoder-decoder model based on the calculation of the self-attention mechanism. It consists of two parts: the encoder and the decoder. The encoder is divided into two sublayers, namely the multi-head self-attention sublayer and the fully connected feedforward network sublayer; the decoder has three sublayers, namely the masked multi-head self-attention sublayer, the multi-head self-attention sublayer and the fully connected feedforward network sublayer;

[0059] The self-attention mechanism used by the original Transformer only calculates the attention between two data points without considering the changing trend of the context; see Figure 4 As shown, Figure 4 The change trends of position 1 and position 2 are the same, but the values ​​are different; while the values ​​of position 1 and position 2 are the same, but the data change trends are different. When predicting the position 1 part, the original Transformer is more likely to focus on position 2 with a similar value, resulting in a large deviation between the prediction result and the actual result;

[0060] In order to improve the algorithm's ability to learn local change trends, the Transformer is combined with a convolutional neural network, and the convolutional self-attention mechanism is used to calculate not only the attention between two data points, but also the context of the data points to learn nearby change trends; after using the convolutional self-attention mechanism, Figure 4 The changing trend of position 1 can be learned from position 3, and the prediction accuracy is greatly improved;

[0061] The logarithmic sparse strategy can reduce the spatial complexity of the model algorithm and improve the prediction efficiency without significantly affecting the model performance;

[0062] The mean absolute percentage error and root mean square error are set as evaluation indicators. The mean absolute percentage error formula is: The root mean square error formula is:

[0063]

[0064] Where MAPE stands for mean absolute percentage error; RMSE stands for root mean square error; Yt Represents the actual usage at time t; Represents the usage predicted by the model at time t, including GPU usage, CPU usage, memory usage, and network bandwidth IO usage; N represents the number of data of the same type;

[0065] The qualified parameters m and r are set based on the evaluation index; in this embodiment, m is set to 0.08 and r is set to 8.00;

[0066] In the specific implementation process, when predicting time series resources, the Transformer model will have problems with high computational space complexity and insensitivity to local context information. In order to solve the problem of low prediction efficiency caused by high computational space complexity, a logarithmic sparsity strategy is added; in order to solve the problem of low prediction accuracy caused by insensitivity to local context information, the Transformer large model is combined with the convolutional neural network, and the convolutional self-attention mechanism calculation of the convolutional neural network is used to replace the self-attention mechanism calculation used by the Transformer large model; the Transformer calculated using the convolutional self-attention mechanism is fully compatible with the original Transformer. When the convolution kernel size calculated by the convolutional self-attention mechanism is set to 1, it returns to the original Transformer model; the qualified parameters m and r can be set by yourself.

[0067] The training unit is configured with a training strategy, and the training strategy includes: dividing the first data set into a training set and a test set in a ratio of 8:2;

[0068] Set the training hyperparameters of the first computing power prediction model, set the learning rate to E, set the batch size to B, and set the number of model training rounds to P. In this embodiment, E=0.00, B=64, and P=2000.

[0069] Set the mean square error MSE as the loss function and use the Adam optimizer as the model training optimizer; the mean square error MSE formula is as follows: Where MSE stands for mean square error; Y t Represents the actual usage at time t; Represents the usage predicted by the model at time t, including GPU usage, CPU usage, memory usage, and network bandwidth IO usage; N represents the number of data of the same type;

[0070] The first computing power prediction model is trained using the training set according to the set training hyperparameters. Every time P rounds of training are completed, the first computing power prediction model is tested using the test set based on evaluation metrics, and the evaluation results M and R are output. When M and R satisfy M ≤ m and R ≤ r, the first computing power prediction model is saved and marked as the second computing power prediction model; if M and R do not satisfy M ≤ m and R ≤ r, the first computing power prediction model is retrained using the training set according to the set training hyperparameters until M and R satisfy M ≤ m and R ≤ r, and the second computing power prediction model is obtained. For example, in this embodiment, the qualified parameters m = 0.080 and r = 8.00 are set, and the model evaluation unit outputs MAPE = 0.063% and RMSE = 5.24; MAPE < m and RMSE < r, so the first computing power prediction model is saved and marked as the second computing power prediction model;

[0071] In the specific implementation process, the smaller the values of the output mean absolute percentage error and root mean square error, the better the performance of the second computing power prediction model, that is, the higher the prediction accuracy.

[0072] The computing power prediction module includes a second computing power prediction model, which predicts the computing power resource requirements based on the first data set to obtain a prediction result;

[0073] The computing power prediction module is configured with a computing power prediction strategy, which includes: obtaining the historical operation data of each node of a computing power cluster; including the GPU usage, CPU usage, memory usage, and network bandwidth IO occupancy of each node of the computing power cluster;

[0074] Perform normalization processing on the historical operation data and bind a time tag to each data;

[0075] Input the historical operation data with the time tag bound into the second computing power prediction model to obtain the operation data of each node of the computing power cluster within the second time length. The second time length is T2, which is marked as the predicted operation data. The predicted operation data includes the predicted GPU usage, predicted CPU usage, predicted memory usage, and predicted network bandwidth IO occupancy of each node of the computing power cluster;

[0076] In the specific implementation process, the first time length T and the first time interval a for obtaining the historical operation data are set according to the prediction requirements. For example, if it is necessary to predict the computing power resource requirements for the next 24 hours, the first time length for obtaining the historical operation data can be 20 days, that is, the data for the previous 30 days is obtained, and the first time interval a of the data can be 10 minutes, 15 minutes, or 30 minutes, etc.; to ensure the prediction accuracy, the longer the prediction time length, the more operation data of each node of the computing power cluster is required.

[0077] The computing power control module includes a generation unit and an execution unit. The generation unit generates control optimization instructions based on the prediction results, and the execution unit completes the control optimization of computing power resources based on the control optimization instructions.

[0078] The generation unit is configured with a control generation strategy, which includes: obtaining the maximum computing power resource F of all computing power nodes in the computing power cluster max The maximum computing resources include the maximum GPU usage, the maximum CPU usage, the maximum memory usage and the maximum network bandwidth IO.

[0079] Get the current operation data of the computing cluster, including the GPU usage, CPU usage, memory usage, and network bandwidth IO usage of each node in the computing cluster, and get the computing resources F currently used by the computing cluster o ; o represents the current moment;

[0080] Get future computing power resources F based on predicted operation data i , i represents the i-th moment, F i Including the predicted GPU usage, CPU usage, memory usage and network bandwidth IO usage of each node in the computing cluster at time i;

[0081] The maximum computing power resource F based on the computing power cluster max 、Used computing resources F o and future computing resources F i Generate control optimization instructions: subtract the maximum computing power resources from the used computing power resources to obtain the current available computing power resources F x , that is, F x =F max -F o When future computing resources F i Exceeds the current available computing resources F x , automatically increase computing resources and expand the maximum computing resources, including adding computing nodes, expanding memory capacity, and increasing network bandwidth; if the current available computing resources F x Greater than future computing resources F i , it will automatically reduce computing resources, including reducing computing nodes, reducing memory capacity, and reducing network bandwidth;

[0082] The execution unit is configured with an execution strategy, which includes: controlling the computing power cluster based on the control optimization instructions to complete the control optimization of computing power resources;

[0083] In the specific implementation process, the maximum computing power resource F of the computing power cluster can be calculated by computer equipment or fixed algorithms. max 、Used computing resources F o and future computing resources F iGenerate control optimization instructions to improve the control optimization efficiency, and transmit the optimization instructions to the control center of the computing power cluster to complete the control optimization of computing power resources.

[0084] Example 2, please refer to Figure 2 As shown, in the second aspect, the present application provides a computing resource control optimization method based on an AI big model, comprising the following steps:

[0085] Step S1, collecting historical operation data of each node in the computing power cluster, and constructing a first data set based on the historical operation data; Step S1 includes the following sub-steps:

[0086] Step S101, obtaining historical operation data of each node in the computing power cluster once at a first time interval within a first time length; the first time length is set to T, the first time interval is set to a, and the historical operation data includes GPU usage, CPU usage, memory usage, and network bandwidth IO usage of each node in the computing power cluster;

[0087] Step S102, normalizing the historical operation data according to different data types;

[0088] Step S103, binding a time tag to each historical operation data according to the time when each historical operation data is obtained, each time tag includes the GPU usage, CPU usage, memory usage and network bandwidth IO usage at the corresponding time;

[0089] Step S104, storing according to the time sequence on the time tag and marking it as the first data set.

[0090] Step S2, based on the Transformer large model, improvements are made to build a first computing power prediction model and set evaluation indicators; Step S2 includes the following sub-steps:

[0091] Step S201, based on the Transformer large model, combining the Transformer large model with the convolutional neural network, and replacing the self-attention mechanism calculation used by the Transformer large model with the convolutional self-attention mechanism calculation of the convolutional neural network;

[0092] Step S202, adding a logarithmic sparse strategy to the convolutional self-attention mechanism calculation, marking the improved Transformer large model as the first computing power prediction model;

[0093] Step S203, setting the mean absolute percentage error and the root mean square error as evaluation indicators, the mean absolute percentage error formula is: The root mean square error formula is:

[0094]

[0095] Where MAPE stands for mean absolute percentage error; RMSE stands for root mean square error; Y t Represents the actual usage at time t; Represents the usage predicted by the model at time t, including GPU usage, CPU usage, memory usage, and network bandwidth IO usage; N represents the number of data of the same type;

[0096] Step S204, setting qualified parameters m and r based on the evaluation index.

[0097] Step S3, optimizing and training the first computing power prediction model based on the first data set and the evaluation index to obtain a second computing power prediction model; Step S3 includes the following sub-steps:

[0098] Step S301, dividing the first data set into a training set and a test set in a ratio of 8:2;

[0099] Step S302, setting the training hyperparameters of the first computing power prediction model, setting the learning rate to E; setting the batch size to B; and setting the number of model training rounds to P rounds;

[0100] Step S303, setting the mean square error MSE as the loss function, and using the Adam optimizer as the model training optimizer;

[0101] Step S304, training the first computing power prediction model using the training set according to the set training hyperparameters;

[0102] Step S305: Test the first computing power prediction model using the test set based on the evaluation index, output evaluation results M and R, and when M and R satisfy M≤m and R≤r, save the first computing power prediction model and mark it as the second computing power prediction model;

[0103] Step S306: If M and R do not satisfy M≤m and R≤r, the first computing power prediction model is retrained using the training set according to the set training hyperparameters until M and R satisfy M≤m and R≤r, thereby obtaining a second computing power prediction model.

[0104] Step S4, predicting the computing power resource demand based on the second computing power prediction model to obtain a prediction result; Step S4 includes the following sub-steps:

[0105] Step S401, obtaining a section of historical operation data of each node in the computing power cluster, including GPU usage, CPU usage, memory usage, and network bandwidth IO usage of each node in the computing power cluster;

[0106] Step S402, normalize the historical operation data and bind a time tag to each data;

[0107] Step S403, input the historical operation data with the bound time tag into the second computing power prediction model to obtain the operation data of each node in the computing power cluster within the second time length, the second time length is T2, marked as predicted operation data, the predicted operation data includes the GPU predicted usage, CPU predicted usage, memory predicted usage and network bandwidth IO predicted occupancy of each node in the computing power cluster.

[0108] Step S5, generating a control optimization instruction based on the prediction result, and completing the control optimization of computing power resources based on the control optimization instruction; Step S5 includes the following sub-steps:

[0109] Step S501: Obtain the maximum computing power resource F of all computing power nodes in the computing power cluster max ,The maximum computing resources include the maximum GPU usage, the maximum CPU usage, the maximum memory usage and the maximum network bandwidth IO;

[0110] Step S502, obtain the current operation data of the computing power cluster, including the GPU usage, CPU usage, memory usage and network bandwidth IO usage of each node in the computing power cluster, and obtain the computing power resources F currently used by the computing power cluster o ; o represents the current moment;

[0111] Step S503: Obtain future computing power resources F based on the predicted operation data i , i represents the i-th moment, F i Including the predicted GPU usage, CPU usage, memory usage and network bandwidth IO usage of each node in the computing cluster at time i;

[0112] Step S504, based on the maximum computing power resource Fmax, the used computing power resource Fo and the future computing power resource Fi of the computing power cluster, a control optimization instruction is generated: the maximum computing power resource is subtracted from the used computing power resource to obtain the current available computing power resource Fx;

[0113] Step S505: When the future computing power resources Fi exceed the currently available computing power resources Fx, the computing power resources are automatically increased and the maximum computing power resources are expanded, including increasing computing power nodes, expanding memory capacity, and increasing network bandwidth; if the currently available computing power resources Fx are greater than the future computing power resources Fi, the computing power resources are automatically reduced, including reducing computing power nodes, reducing memory capacity, and reducing network bandwidth.

[0114] Step S506: Control the computing power cluster based on the control optimization instruction to complete the control optimization of computing power resources.

[0115] Embodiment 3. The present application also provides a computer program product, which includes a computer program stored on a computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute a computing power resource regulation and optimization method based on an AI big model provided by the above methods. The method includes: collecting historical operation data of each node in the computing power cluster, and constructing a first data set based on the historical operation data; improving based on the Transformer big model, constructing a first computing power prediction model, and setting evaluation indicators, optimizing and training the first computing power prediction model based on the first data set and the evaluation indicators to obtain a second computing power prediction model; predicting computing power resource demand based on the second computing power prediction model to obtain a prediction result; generating regulation and optimization instructions based on the prediction result, and completing computing power resource regulation and optimization based on the regulation and optimization instructions.

[0116] Embodiment 4, the present application also provides a computer-readable storage medium, the present application provides a storage medium, on which a computer program is stored. When the computer program is executed by the processor, the steps in the above-mentioned computing power resource regulation and optimization method based on the AI ​​big model are executed to achieve the following functions: collect historical operation data of each node in the computing power cluster, and construct a first data set based on the historical operation data; improve based on the Transformer big model, construct a first computing power prediction model, and set evaluation indicators, optimize and train the first computing power prediction model based on the first data set and the evaluation indicators to obtain a second computing power prediction model; predict the computing power resource demand based on the second computing power prediction model to obtain a prediction result; generate regulation and optimization instructions based on the prediction result, and complete the computing power resource regulation and optimization based on the regulation and optimization instructions.

[0117] Through the description of the above implementation methods, the embodiments of the present invention can be provided as methods, systems or computer program products. Based on such an understanding, the above technical solutions can be essentially or partly contributed to the prior art in the form of software products, which can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and include several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0118] In the embodiments provided in the present application, it should be understood that the disclosed system or method can be implemented in other ways. The embodiments described above are merely schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or units can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of systems, modules and units can be electrical, mechanical or other forms.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A computing resource control and optimization method based on AI big model, characterized in that: The steps include: Collect historical operation data of each node in the computing power cluster, and build a first data set based on the historical operation data; Improvements were made based on the Transformer large model, the first computing power prediction model was built, and evaluation indicators were set; Optimize and train the first computing power prediction model based on the first data set and the evaluation index to obtain a second computing power prediction model; Predicting the computing power resource demand based on the second computing power prediction model to obtain a prediction result; Generate control and optimization instructions based on the prediction results, and complete the control and optimization of computing power resources based on the control and optimization instructions.

2. According to claim 1, a computing resource control and optimization method based on AI big model is characterized in that: Collecting historical operation data of each node in the computing power cluster includes the following sub-steps: The historical operation data of each node in the computing power cluster is obtained once every first time interval within a first time length; the first time length is set to T, the first time interval is set to a, and the historical operation data includes GPU usage, CPU usage, memory usage and network bandwidth IO occupancy of each node in the computing power cluster.

3. According to claim 2, a computing resource control optimization method based on AI big model is characterized in that: Constructing the first data set based on historical operation data includes the following sub-steps: The historical operation data are normalized according to different data types, and a time label is bound to each historical operation data according to the time when each historical operation data is obtained. Each time label includes the GPU usage, CPU usage, memory usage and network bandwidth IO occupancy at the corresponding time. They are stored in the time order on the time label and marked as the first data set.

4. According to claim 3, a computing resource control and optimization method based on an AI large model is characterized in that: Based on the Transformer large model, the construction of the first computing power prediction model includes the following sub-steps: Based on the Transformer large model, the Transformer large model is combined with the convolutional neural network, the convolutional self-attention mechanism calculation of the convolutional neural network is used to replace the self-attention mechanism calculation used by the Transformer large model, and the logarithmic sparsity strategy is added to the convolutional self-attention mechanism calculation. The improved Transformer large model is marked as the first computing power prediction model.

5. According to claim 4, a computing resource control optimization method based on AI big model is characterized in that: Setting evaluation indicators includes the following sub-steps: The mean absolute percentage error and root mean square error are set as evaluation indicators. The mean absolute percentage error formula is: The root mean square error formula is: Where MAPE stands for mean absolute percentage error; RMSE stands for root mean square error; Y t Represents the actual usage at time t; Represents the usage predicted by the model at time t, which includes GPU usage, CPU usage, memory usage, and network bandwidth IO usage; N represents the number of data of the same type; The qualified parameters m and r are set based on the evaluation index.

6. According to claim 5, a computing resource control and optimization method based on AI big model is characterized in that: Optimizing and training the first computing power prediction model based on the first data set and the evaluation index to obtain the second computing power prediction model includes the following sub-steps: The first data set is divided into a training set and a test set in a ratio of 8:2; Set the training hyperparameters of the first computing power prediction model, set the learning rate to E, the batch size to B, and the number of model training rounds to P; Set mean square error (MSE) as the loss function and use Adam optimizer as the model training optimizer; The first computing power prediction model is trained using the training set according to the set training hyperparameters. After each P round of training, the first computing power prediction model is tested using the test set based on the evaluation indicators, and the evaluation results M and R are output. When M and R satisfy M≤m and R≤r, the first computing power prediction model is saved and marked as the second computing power prediction model; if M and R do not satisfy M≤m and R≤r, the first computing power prediction model is trained again using the training set according to the set training hyperparameters until M and R satisfy M≤m and R≤r to obtain the second computing power prediction model.

7. The method for optimizing computing resource control based on AI big model according to claim 6 is characterized in that: Predicting the computing power resource demand based on the second computing power prediction model to obtain the prediction result includes the following sub-steps: Get a section of historical operation data of each node in the computing cluster, including GPU usage, CPU usage, memory usage, and network bandwidth IO usage of each node in the computing cluster; Normalize the historical operation data and bind a time label to each data; The historical operation data with the bound time label is input into the second computing power prediction model to obtain the operation data of each node in the computing power cluster within the second time length. The second time length is T2, which is marked as predicted operation data. The predicted operation data includes the predicted GPU usage, CPU usage, memory usage and network bandwidth IO usage of each node in the computing power cluster.

8. The method for optimizing computing resource control based on AI big model according to claim 7 is characterized in that: Generate control optimization instructions based on the prediction results. Computing resource control optimization based on the control optimization instructions includes the following sub-steps: Get the maximum computing power resource F of all computing power nodes in the computing power cluster max ,The maximum computing resources include the maximum GPU usage, the maximum CPU usage, the maximum memory usage and the maximum network bandwidth IO; Get the current operation data of the computing cluster, including the GPU usage, CPU usage, memory usage, and network bandwidth IO usage of each node in the computing cluster, and get the computing resources F currently used by the computing cluster o ; o represents the current moment; Get future computing power resources F based on predicted operation data i , i represents the i-th moment, F i Including the predicted GPU usage, CPU usage, memory usage and network bandwidth IO usage of each node in the computing cluster at time i; Generate control optimization instructions based on the maximum computing power resources Fmax, used computing power resources Fo and future computing power resources Fi of the computing power cluster: subtract the maximum computing power resources from the used computing power resources to obtain the current available computing power resources Fx. When the future computing power resources Fi exceed the current available computing power resources Fx, automatically increase computing power resources and expand the maximum computing power resources, including increasing computing power nodes, expanding memory capacity and increasing network bandwidth; if the current available computing power resources Fx are greater than the future computing power resources Fi, automatically reduce computing power resources, including reducing computing power nodes, reducing memory capacity and reducing network bandwidth; The computing power cluster is controlled based on the control and optimization instructions to complete the control and optimization of computing power resources.

9. A computing resource control and optimization system based on an AI big model, used to implement a computing resource control and optimization method based on an AI big model as described in any one of claims 1-8, characterized in that: It includes data management module, model building module, computing power prediction module and computing power control module; The data management module includes a collection unit and a processing unit, the collection unit is used to collect historical operation data of each node in the computing power cluster, and the processing unit is used to process the historical operation data and construct a first data set based on the historical operation data; The model building module includes a building unit and a training unit. The building unit is used to improve the Transformer large model, build a first computing power prediction model, and set evaluation indicators. The training unit optimizes and trains the first computing power prediction model based on the first data set and the evaluation index to obtain a second computing power prediction model; The computing power prediction module includes a second computing power prediction model, which predicts the computing power resource demand based on the first data set to obtain a prediction result; The computing power control module includes a generation unit and an execution unit. The generation unit generates a control optimization instruction based on the prediction result, and the execution unit completes the computing power resource control optimization based on the control optimization instruction.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are executed.

Images