A method for predicting the performance of a computing power network and related devices

Abstract

The present invention provides a computing power network performance prediction method and related devices, which obtain historical performance data of the computing power network as a training set; input the training set into a computing power network performance prediction model including a test time training layer, pre-train the computing power network performance prediction model to obtain a pre-trained computing power network performance prediction model, in which the weights of the backbone network module in the pre-trained computing power network performance prediction model are locked, train the test time training layer in the pre-trained computing power network performance prediction model, and adapt to changes in the network performance mode by updating the weights of the test time training layer; input the network performance data of the target computing power network into the trained computing power network performance prediction model for prediction to obtain a performance prediction result of the target computing power network; it can adapt to changes in the computing power network performance mode while ensuring prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of computing power network performance prediction, and particularly to a computing power network performance prediction method and related devices. Background Art

[0002] As a new type of information infrastructure, the computing power network integrates multi-level computing power resources such as cloud platforms, edge servers, and terminal devices, and provides services such as data perception, transmission, storage, and operation. Its core lies in the on-demand allocation and flexible scheduling of computing resources, storage resources, and network resources among the cloud, edge, and terminal according to business requirements, so as to improve the quality of service experience and the optimization and efficient utilization of network and computing resources. The computing power network has four characteristics: resource abstraction, service guarantee, unified management and control, and elastic scheduling. These characteristics make the performance prediction of the computing power network complex and important.

[0003] As a link connecting computing power resources from all parties, the network performance of the computing power network has a decisive impact on the computing power utilization rate. Therefore, problems such as network latency and bandwidth limitations cannot be ignored. In this context, intelligent and automated technologies have become the key means to improve the efficiency of the computing power network, and performance prediction is an important prerequisite for achieving the improvement of the computing power network efficiency. Through accurate performance prediction, the dynamic adjustment and optimization of the computing power network can be realized, thereby improving resource utilization rate and service quality.

[0004] However, there are still some problems in the current computing power network performance prediction:

[0005] 1. The performance change pattern of the computing power network shows extremely strong non-linearity. A single advanced general time series model cannot be directly used for computing power network performance prediction, and good prediction results can only be obtained through adaptation or a combination of multiple models.

[0006] 2. The performance pattern of the computing power network will change greatly over time, which requires the model to continuously learn the changes in the performance pattern, and most models do not have the ability of online learning.

[0007] 3. There is a lack of a good online learning mechanism. Directly introducing the online learning function into the model often causes the parameters of the model to jitter, and the weights are positive and negative, which instead reduces the prediction ability of the model. Summary of the Invention

[0008] The present invention provides a computing power network performance prediction method and related devices, the purpose of which is to adapt to the changes in the computing power network performance pattern while ensuring prediction accuracy.

[0009] To achieve the above object, the present invention provides a computing power network performance prediction method, including:

[0010] Step 1: Obtain the historical performance data of the computing power network as the training set;

[0011] Step 2: Construct a computing power network performance prediction model. The computing power network performance prediction model includes an embedding module for performing embedding operations on input data, a Mamba backbone network module for capturing the non-linear features of the data after the embedding operation, and an output module for converting the captured non-linear features into a prediction result. Among them, the Mamba backbone network module includes a test-time training layer;

[0012] Step 3: Input the training set into the computing power network performance prediction model to pre-train the computing power network performance prediction model. In the pre-trained computing power network performance prediction model, the weights of the backbone network module are locked. Train the test-time training layer in the pre-trained computing power network performance prediction model. By updating the weights of the test-time training layer to adapt to the changes in the network performance mode, obtain the trained computing power network performance prediction model;

[0013] Step 4: Input the network performance data of the target computing power network into the trained computing power network performance prediction model for prediction to obtain the performance prediction result of the target computing power network.

[0014] Furthermore, before inputting the training set into the constructed computing power network performance prediction model, it also includes:

[0015] Select wavelet decomposition to decompose the training set to obtain a low-frequency component and a high-frequency component;

[0016] Use windows of multiple different scales to perform sliding average pooling on the training set to obtain the processed training set;

[0017] Concatenate the processed training set with the low-frequency component and the high-frequency component and perform an embedding operation to obtain a frequency-domain reconstruction target.

[0018] Furthermore, the expression of the frequency-domain reconstruction target is:

[0019] ;

[0020] Where, represents the frequency-domain reconstruction target, represents the embedding operation, represents the concatenation operation, represents the wavelet transform, represents the sliding average pooling, represents the input network performance data.

[0021] Furthermore, the Mamba backbone network module further includes a first linear layer, a second linear layer, a third linear layer, a first convolutional layer, a first activation layer, a second activation layer, and a multiplier;

[0022] The input ends of the first linear layer and the second linear layer are both the input end of the Mamba backbone network module, and are connected to the output end of the embedding layer;

[0023] The output end of the first linear layer is connected to the input end of the first convolutional layer, the output end of the first convolutional layer is connected to the input end of the first activation layer, the input end of the first activation layer is connected to the input end of the test-time training layer, and the output end of the test-time training layer is connected to the first input end of the multiplier;

[0024] The output end of the second linear layer is connected to the input end of the second activation layer, and the output end of the second activation layer is connected to the second input end of the multiplier;

[0025] The output end of the multiplier is connected to the input end of the third linear layer, and the output end of the third linear layer is connected to the input end of the output module.

[0026] Furthermore, the output module is a multi-layer perceptron network composed of 2 linear layers.

[0027] Furthermore, step 3 includes:

[0028] Pre-training the computing power network performance prediction model using the training set to obtain the pre-trained computing power network performance prediction model, and the weights of the backbone network module in the pre-trained computing power network performance prediction model are locked;

[0029] Construct the self-supervised loss function of the test-time training layer in the computing power network performance prediction model in a self-supervised manner, train the test-time training layer based on the self-supervised loss function, and update the weights of the test-time training layer to adapt to the changes in the network performance mode, so as to obtain the trained computing power network performance prediction model.

[0030] Furthermore, the self-supervised loss function is:

[0031] ;

[0032] Among them, represents the self-supervised loss function, represents the weight of the test-time training layer, represents the input of the test-time training layer, 、 、 represent low-rank matrices, represents the model of the test-time training layer, represents the frequency-domain reconstruction target.

[0033] The present invention also provides a computing power network performance prediction device, including:

[0034] An acquisition module, configured to acquire historical performance data of the computing power network as a training set;

[0035] A construction module, configured to construct a computing power network performance prediction model, where the computing power network performance prediction model includes an embedding module connected in sequence for performing an embedding operation on input data, a Mamba backbone network module for capturing non-linear features of the data after the embedding operation, and an output module for converting the captured non-linear features into a prediction result; wherein, the Mamba backbone network module includes a test-time training layer;

[0036] A training module, configured to input the training set into the computing power network performance prediction model, pre-train the computing power network performance prediction model, obtain a pre-trained computing power network performance prediction model, lock the weights of the backbone network module in the pre-trained computing power network performance prediction model, train the test-time training layer in the pre-trained computing power network performance prediction model, and update the weights of the test-time training layer to adapt to changes in the network performance mode, so as to obtain a trained computing power network performance prediction model;

[0037] A prediction module, configured to input the network performance data of the target computing power network into the trained computing power network performance prediction model for prediction, and obtain a performance prediction result of the target computing power network.

[0038] The present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements the computing power network performance prediction method when executing the computer program.

[0039] The present invention also provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and the computer program implements the computing power network performance prediction method when executed by a processor.

[0040] The above solution of the present invention has the following beneficial effects:

[0041] The present invention takes the historical performance data of the computing power network as the training set; inputs the training set into the computing power network performance prediction model including the test-time training layer to pre-train the computing power network performance prediction model, and obtains the pre-trained computing power network performance prediction model. In the pre-trained computing power network performance prediction model, the weights of the backbone network module are locked. Then, the test-time training layer in the pre-trained computing power network performance prediction model is trained. By updating the weights of the test-time training layer to adapt to the changes in the network performance pattern, the trained computing power network performance prediction model is obtained. The test-time training layer is set within the Mamba backbone network module; inputs the network performance data of the target computing power network into the trained computing power network performance prediction model for prediction, and obtains the performance prediction result of the target computing power network. Compared with the prior art, the present invention combines the characteristics of the recurrent neural network and the convolutional neural network through the Mamba backbone network module, solves the problem of computing efficiency when dealing with long time series, and uses the test-time training layer to regard the hidden state as a model, enabling it to better capture the non-linear features of the sequence, and can adapt to the changes in the computing power network performance pattern while ensuring the prediction accuracy.

[0042] Other beneficial effects of the present invention will be described in detail in the subsequent specific implementation part. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic flowchart of an embodiment of the present invention;

[0044] Figure 2 It is a schematic structural diagram of the computing power network performance prediction model in an embodiment of the present invention;

[0045] Figure 3 It is a schematic structural diagram of the computing power network performance prediction device in an embodiment of the present invention;

[0046] Figure 4 It is a schematic structural diagram of the terminal device in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0049] In the description of the present invention, it should be noted that unless otherwise clearly specified and defined, the terms "installed", "connected", and "coupled" should be understood in a broad sense. For example, it can be a locking connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0050] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0051] The present invention provides a method and related device for predicting the performance of a computing power network in view of existing problems.

[0052] As Figure 1 、 Figure 2 shown, an embodiment of the present invention provides a method for predicting the performance of a computing power network, including:

[0053] Step 1, obtaining historical performance data of the computing power network as a training set;

[0054] Step 2, constructing a computing power network performance prediction model, where the computing power network performance prediction model includes an embedding module for performing an embedding operation on input data, a Mamba backbone network module for capturing non-linear features of the data after the embedding operation, and an output module for converting the captured non-linear features into a prediction result, which are connected in sequence; among them, the Mamba backbone network module includes a test-time training layer;

[0055] Step 3, inputting the training set into the computing power network performance prediction model, pre-training the computing power network performance prediction model to obtain a pre-trained computing power network performance prediction model, locking the weights of the backbone network module in the pre-trained computing power network performance prediction model, training the test-time training layer in the pre-trained computing power network performance prediction model, and updating the weights of the test-time training layer to adapt to changes in the network performance mode, so as to obtain a trained computing power network performance prediction model;

[0056] Step 4: Input the network performance data of the target computing power network into the trained computing power network performance prediction model for prediction to obtain the performance prediction result of the target computing power network.

[0057] In the embodiment of the present invention, the historical performance data includes but is not limited to network traffic, bandwidth, delay, etc.

[0058] In the embodiment of the present invention, first, obtain the maximum performance value and the minimum performance value from the historical performance data, and perform normalization processing on the historical performance data. The normalization expression is:

[0059] ;

[0060] where represents the normalized historical performance data, represents the historical performance data, represents the maximum performance value, represents the minimum performance value;

[0061] Set the size of each piece of historical performance data as T and the corresponding label size as P to obtain multiple samples and the corresponding labels. Use the first 80% of the samples as the training set and the remaining 20% of the samples as the test set. In the embodiment of the present invention, T is set to 24 and P is set to 10, that is, predict the data of the next 10 time instants through the historical performance data of the first 24 time instants.

[0062] Specifically, before inputting the training set into the constructed computing power network performance prediction model, it further includes:

[0063] Select wavelet decomposition to decompose the training set to obtain the low-frequency component and the high-frequency component;

[0064] Use windows of multiple different scales to perform sliding average pooling on the training set to obtain the processed training set;

[0065] Concatenate the processed training set with the low-frequency component and the high-frequency component and perform an embedding operation to obtain the frequency domain reconstruction target.

[0066] In the embodiments of the present invention, the low-frequency component corresponds to a signal with slow changes, and the high-frequency component corresponds to a signal with fast changes; since the signals are mixed together, they lose their characteristics, and the frequency-domain information can often better display the variation characteristics of the time series. The performance series has strong non-linear characteristics, indicating that the performance series is non-stationary. Therefore, it is necessary to decompose the training set to obtain multiple low-frequency components and high-frequency components with more characteristics; the low-frequency component has a low scale and a large window, and the result is smoother. The high-frequency component has a high scale and a small window, retaining more details. The processed training set is concatenated with the low-frequency component and the high-frequency component, and then mapped to a specified dimension through an embedding operation to obtain a frequency-domain reconstruction target that highlights details. The expression of the frequency-domain reconstruction target is:

[0067] ;

[0068] Wherein, represents the frequency-domain reconstruction target, represents the embedding operation, which maps the input data to a specified dimension, represents the concatenation operation, represents the wavelet transform, represents the moving average pooling, represents the input network performance data.

[0069] Most preferably, in the embodiments of the present invention, the embedding module consists of a linear layer. The linear layer maps the input data to a specified dimension, converting the discrete data into a continuous vector with a fixed dimension. The calculation expression is:

[0070] ;

[0071] Wherein, represents the embedding vector, represents the input network performance data, represents the embedding operation;

[0072] This conversion increases the dimension of the data and also increases the expressive power of the numerical data, enabling the model to more effectively process various types of data.

[0073] It should be noted that the Mamba backbone network module is the Mamba backbone network. In the embodiments of the present invention, the state space model in the Mamba backbone network is replaced by a test-time training layer.

[0074] Specifically, the Mamba backbone network module further includes a first linear layer, a second linear layer, a third linear layer, a first convolutional layer, a first activation layer, a second activation layer, and a multiplier;

[0075] The input ends of the first linear layer and the second linear layer are both the input ends of the Mamba backbone network module and are connected to the output end of the embedding layer;

[0076] The output end of the first linear layer is connected to the input end of the first convolutional layer, the output end of the first convolutional layer is connected to the input end of the first activation layer, the input end of the first activation layer is connected to the input end of the test-time training layer, and the output end of the test-time training layer is connected to the first input end of the multiplier;

[0077] The output end of the second linear layer is connected to the input end of the second activation layer, and the output end of the second activation layer is connected to the second input end of the multiplier;

[0078] The output end of the multiplier is connected to the input end of the third linear layer, and the output end of the third linear layer is connected to the input end of the output module.

[0079] In the embodiment of the present invention, after the historical performance data is processed by the embedding module, the obtained embedding vector is input into the Mamba backbone network module and divided into two branches. The first branch first passes through the first linear layer, then through the first convolutional layer, and then is activated by the first activation layer and input into the test-time training layer for prediction to obtain the first feature vector; the second branch first passes through the second linear layer and then is activated by the second activation layer to obtain the second feature vector, and then the first feature vector and the second feature vector respectively output by the two branches are multiplied by the multiplier and processed by the third linear layer and used as the output of the Mamba backbone network module to be input into the output module.

[0080] In the embodiment of the present invention, both the first activation layer and the second activation layer are silu functions.

[0081] In the embodiment of the present invention, the key of the test-time training layer is that instead of compressing the context into a fixed-size hidden state , but regarding the hidden state as a model, the hidden state is equivalent to the weights of a model, and this model can be a linear model, a small neural network model or any other form of model. The output rule of the test-time training layer can be expressed as:

[0082] ;

[0083] Among them, represents the input of the test-time training layer, represents the model of the test-time training layer, represents the output of the test-time training layer, that is, the prediction made by the updated weights of the model on the input ;

[0084] In an embodiment of the present invention, in order to enable the model to capture the mutability of performance data, the test-time training layer is improved so that the improved test-time training layer can enhance the non-linear ability of the model and be more sensitive to performance changes. That is, the reconstruction objective of the test-time training layer should further highlight the detail features and trend features.

[0085] Most preferably, the output module is a multi-layer perceptron network composed of 2 linear layers.

[0086] In an embodiment of the present invention, the output result of the backbone network is converted into a performance prediction result for a future period of time through a multi-layer perceptron network.

[0087] Specifically, step 3 includes:

[0088] Pre-train the computing power network performance prediction model using the training set to obtain the pre-trained computing power network performance prediction model. The weights of the backbone network module in the pre-trained computing power network performance prediction model are locked.

[0089] Construct the self-supervised loss function of the test-time training layer in the computing power network performance prediction model through the self-supervised method. Train the test-time training layer based on the self-supervised loss function. By updating the weights of the test-time training layer to adapt to the changes in the network performance pattern, obtain the trained computing power network performance prediction model.

[0090] Specifically, in an embodiment of the present invention, the computing power network performance prediction model is pre-trained using the training set, where the parameters of the Mamba backbone network and the test-time training layer are trained through supervised and self-supervised methods respectively.

[0091] The Mamba backbone network is trained in a supervised manner, and the loss is calculated after the output module. The loss function is:

[0092] ;

[0093] Among them, is the true performance result, is the prediction result of the entire model.

[0094] The specific training process is as follows:

[0095] (1): Set the batch size and divide the training set into several batches;

[0096] (2): Feed the data of each batch into the computing power network performance prediction model and calculate the LOSS loss function;

[0097] (3): Perform iterative optimization based on the Adam algorithm until the set number of iterations.

[0098] (4): The iteration ends, and the trained model parameters are saved.

[0099] To adapt to unknown changes in network performance patterns, the weights of the test-time training layer are always updated independently. A self-supervised loss function for the test-time training layer in the computing power network performance prediction model is constructed through a self-supervised approach. The test-time training layer is trained based on the self-supervised loss function to obtain the weights of the test-time training layer.

[0100] In the embodiment of the present invention, the update rule of the test-time training layer is to perform further gradient descent on a certain self-supervised loss The calculation expression is:

[0101] ;

[0102] where represents the learning rate;

[0103] loss function is selected as reconstructing itself. To make the learning problem non-trivial, in the embodiment of the present invention, the input is first processed into a corrupted input , and then the loss function is optimized. The expression is:

[0104] ;

[0105] Through the process of reconstructing the input the model can discover the correlations between dimensions in order to reconstruct the original information from the corrupted information . The process of mapping the input sequence to the sequence is programmed into the forward propagation of the test-time training layer, using the above-mentioned hidden states, update rules, and output rules. Even during testing, the test-time training layer still trains a different weight sequence for each input sequence, that is, online learning is achieved.

[0106] The most important part of the test-time training layer is the self-supervised task because it determines the type of features that the weights learn from the test set. In the design of this task, the original test-time training layer adopted a more end-to-end approach, directly optimizing the self-supervised task to achieve the ultimate goal of the next token prediction. Starting from a simple reconstruction task, some outer loop parameters were added to make this task learnable. Therefore, the self-supervised loss is optimized as:

[0107] ;

[0108] The output becomes:

[0109] ;

[0110] , , are a series of low-rank matrices that can disrupt ; , , are outer loop parameters learned in the Mamba backbone network module, while the weights of the test-time training layer are inner loop parameters and do not participate in outer loop updates. After the training of the computing power network performance prediction model is completed, the outer loop parameters are no longer updated, but new performance data can update the inner loop parameters, enabling the model to learn subsequent performance change patterns, but causing large jitters in the overall model.

[0111] The improved TTT layer calculates the gradient through an improved self-supervised loss function and further updates the weight W and bias term b; during the subsequent use of the model, the backbone network parameters do not change, but when data is input into the TTT layer, the TTT layer parameters can be updated, which can avoid large jitters in the entire model and enable the model to adapt to new data.

[0112] Specifically, the self-supervised loss function is:

[0113] ;

[0114] Among them, represents the self-supervised loss function, represents the weights of the test-time training layer, represents the input of the test-time training layer, , , represent low-rank matrices, represents the model of the test-time training layer, represents the frequency domain reconstruction target;

[0115] Since there are a large number of redundant features in the high-frequency and low-frequency components, the low-rank matrix can be exactly used for filtering.

[0116] In the embodiments of the present invention, the model adopted by the test-time training layer

[0117] ;

[0118] Among them, is the input of the model , is the weight of the model is the bias term of the function

[0119] In order to test and evaluate the prediction accuracy of the model, the embodiments of the present invention input the test set into the trained computing power network performance prediction model for prediction, obtain the single-step prediction result, and evaluate the model with two indicators of root mean square error and mean absolute percentage error. The calculation expressions are as follows:

[0120] ;

[0121] ;

[0122] Among them, represents the root mean square error represents the mean absolute percentage error represents the prediction result of the model represents the label in the training set represents the number of prediction results or labels

[0123] The embodiments of the present invention collect the network performance data of a certain computing power node in the target computing power network. Taking one hour as the interval, there are 24 groups per day, and there are 204 days of network performance data in total. The data of the first 194 days are used as the training set, and the data of the last 10 days are used as the test set. The size of the retrospective window is set to 24, and samples are continuously generated in the form of a sliding window. When the size of the prediction window is set to 10, multi-step prediction is performed to predict the performance values of the next 10 hours. The relationship between the prediction error and the number of layers of the Mamba backbone network module is shown in Table 1:

[0124] Table 1 Error magnitudes corresponding to different numbers of layers of the Mamba backbone network module

[0125]

[0126] It can be seen from Table 1 above that when the number of layers of the Mamba backbone network module is set to 3, the performance of the model is the best

[0127] Set the number of layers of the Mamba backbone network module to 3, and perform single-step prediction and multi-step prediction respectively. The results are shown in Table 2:

[0128] Table 2 Single-step prediction and multi-step prediction results

[0129]

[0130] Set the prediction step to 5 and compare it with other time series prediction models. The results are shown in Table 3:

[0131] Table 3 Comparison experiment results when the prediction step size is 5

[0132]

[0133] As can be seen from Table 3 above, the method provided by the present invention outperforms some benchmark time series models and advanced time series models.

[0134] In the embodiment of the present invention, the original test-time training layer is compared with the improved test-time training layer, and the comparison results are shown in Table 4 below:

[0135] Table 4 Comparison results between the original test-time training layer and the improved test-time training layer

[0136]

[0137] As can be seen from Table 4 above, the result error of the improved test-time training layer is lower and it has stronger non-linear ability, which proves the effectiveness of the embodiment of the present invention.

[0138] In the embodiment of the present invention, the historical performance data of the computing power network is obtained as the training set; the training set is input into the computing power network performance prediction model including the test-time training layer to pre-train the computing power network performance prediction model, and the pre-trained computing power network performance prediction model is obtained. The weights of the backbone network module in the pre-trained computing power network performance prediction model are locked, and the test-time training layer in the pre-trained computing power network performance prediction model is trained. By updating the weights of the test-time training layer to adapt to the change of the network performance mode, the trained computing power network performance prediction model is obtained. The test-time training layer is set in the Mamba backbone network module; the network performance data of the target computing power network is input into the trained computing power network performance prediction model for prediction, and the performance prediction result of the target computing power network is obtained; compared with the prior art, in the embodiment of the present invention, by integrating the characteristics of the recurrent neural network and the convolutional neural network through the Mamba backbone network module, the problem of computing efficiency when processing long time series is solved. The test-time training layer treats the hidden state as a model, enabling it to better capture the non-linear features of the sequence, and can adapt to the change of the computing power network performance mode while ensuring the prediction accuracy.

[0139] Corresponding to the computing power network performance prediction method described in the above embodiments, as Figure 3 shown, the embodiment of the present invention further provides a computing power network performance prediction device 100, and the computing power network performance prediction device 100 includes:

[0140] An acquisition module 101, configured to obtain the historical performance data of the computing power network as the training set;

[0141] A building module 102 for building a computing power network performance prediction model, the computing power network performance prediction model including an embedding module for performing an embedding operation on input data, a Mamba backbone network module for capturing non-linear features of the data after the embedding operation, and an output module for converting the captured non-linear features into a prediction result; wherein, the Mamba backbone network module includes a test-time training layer;

[0142] A training module 103 for inputting a training set into the computing power network performance prediction model to pre-train the computing power network performance prediction model, obtaining a pre-trained computing power network performance prediction model, locking the weights of the backbone network module in the pre-trained computing power network performance prediction model, training the test-time training layer in the pre-trained computing power network performance prediction model, and adapting to changes in the network performance mode by updating the weights of the test-time training layer, thereby obtaining a trained computing power network performance prediction model;

[0143] A prediction module 104 for inputting the network performance data of a target computing power network into the trained computing power network performance prediction model for prediction, obtaining a performance prediction result of the target computing power network.

[0144] It should be noted that, for the information interaction, execution process, etc. between the above-mentioned devices / units, since they are based on the same concept as the method embodiment of the present application, their specific functions and the technical effects brought about can be specifically referred to the method embodiment part, and will not be elaborated here.

[0145] Those skilled in the art can clearly understand that, for the sake of convenience and brevity of description, only the above-mentioned division of each functional unit and module is used as an example for illustration. In actual applications, the above-mentioned functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction and do not limit the protection scope of the present application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the foregoing method embodiment and will not be elaborated here.

[0146] The embodiment of the present invention also provides a terminal device, as Figure 4 shown, the terminal device D10 in this embodiment includes: at least one processor D100 ( Figure 4only shows one processor), a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100. When the processor D100 executes the computer program D102, the above-mentioned computing power network performance prediction method is implemented.

[0147] The terminal device D10 may be a computing device such as a desktop computer, a notebook, a palm computer, a server, a server cluster, and a cloud server. The terminal device may include, but is not limited to, a processor D100 and a memory D101. Those skilled in the art can understand that Figure 4 This is only an example of the terminal device D10 and does not constitute a limitation on the terminal device D10. It may include more or fewer components than shown in the figure, or combine some components, or different components. For example, it may also include input / output devices, network access devices, etc.

[0148] The so-called processor D100 may be a central processing unit (CPU, Central Processing Unit), and the processor D100 may also be other general-purpose processors, digital signal processors (DSP, Digital Signal Processor), application-specific integrated circuits (ASIC, Application Specific Integrated Circuit), off-the-shelf programmable gate arrays (FPGA, Field-Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0149] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as the hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may also be an external storage device of the terminal device D10, such as a plug-in hard disk, a smart media card (SMC, SmartMedia Card), a secure digital (SD, Secure Digital) card, a flash card (Flash Card), etc. equipped on the terminal device D10. Further, the memory D101 may also include both the internal storage unit and the external storage device of the terminal device D10. The memory D101 is used to store an operating system, application programs, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program, etc. The memory D101 may also be used to temporarily store data that has been output or will be output.

[0150] It should be noted that, regarding the information interaction, execution process, etc. between the above-mentioned devices / units, since they are based on the same concept as the method embodiments of this application, for their specific functions and the technical effects brought about, reference can be specifically made to the method embodiment part, and details will not be elaborated here.

[0151] Those skilled in the art can clearly understand that, for the convenience and conciseness of description, only the above-mentioned division of each functional unit and module is used as an example for illustration. In practical applications, the above-mentioned functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiments can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction and do not limit the protection scope of this application. The specific working processes of the units and modules in the above system can refer to the corresponding processes in the foregoing method embodiments, and details will not be elaborated here.

[0152] The embodiments of the present invention also provide a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the method for predicting the performance of a computing power network.

[0153] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-mentioned method embodiments of this application, a computer program can be used to instruct the relevant hardware to complete. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium can at least include: any entity or device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium that can carry the computer program code to the construction device / terminal device. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disc, etc.

[0154] The above are the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A computing power network performance prediction method, characterized in that: include: Step 1: Obtain historical performance data of the computing network as a training set; Step 2, constructing a computing power network performance prediction model, the computing power network performance prediction model includes an embedding module for embedding input data, a Mamba backbone network module for capturing nonlinear features of the data after the embedding operation, and an output module for converting the captured nonlinear features into prediction results, wherein the Mamba backbone network module includes a test time training layer; Step 3, inputting the training set into the computing power network performance prediction model, pre-training the computing power network performance prediction model, obtaining a pre-trained computing power network performance prediction model, wherein the weight of the backbone network module in the pre-trained computing power network performance prediction model is locked, and training the test time training layer in the pre-trained computing power network performance prediction model, and updating the weight of the test time training layer to adapt to changes in the network performance mode, thereby obtaining a trained computing power network performance prediction model; Step 4: Input the network performance data of the target computing power network into the trained computing power network performance prediction model for prediction to obtain the performance prediction result of the target computing power network.

2. The computing power network performance prediction method according to claim 1 is characterized in that: Before inputting the training set into the constructed computing power network performance prediction model, it also includes: Select wavelet decomposition to decompose the training set to obtain low-frequency components and high-frequency components; Using windows of multiple different scales to perform sliding average pooling processing on the training set to obtain a processed training set; The processed training set is concatenated with the low-frequency component and the high-frequency component and an embedding operation is performed to obtain a frequency domain reconstruction target.

3. The computing power network performance prediction method according to claim 2 is characterized in that: The expression of the frequency domain reconstruction objective is: in, represents the frequency domain reconstruction target, represents an embedding operation, Represents a splicing operation, represents wavelet transform, represents sliding average pooling, Represents input network performance data.

4. The computing power network performance prediction method according to claim 3 is characterized in that: The Mamba backbone network module also includes a first linear layer, a second linear layer, a third linear layer, a first convolutional layer, a first activation layer, a second activation layer, and a multiplier; The input end of the first linear layer and the input end of the second linear layer are both input ends of the Mamba backbone network module, and are connected to the output end of the embedded module; An output of the first linear layer is connected to an input of the first convolutional layer, an output of the first convolutional layer is connected to an input of the first activation layer, an input of the first activation layer is connected to an input of the test-time training layer, and an output of the test-time training layer is connected to a first input of the multiplier; The output end of the second linear layer is connected to the input end of the second activation layer, and the output end of the second activation layer is connected to the second input end of the multiplier; The output end of the multiplier is connected to the input end of the third linear layer, and the output end of the third linear layer is connected to the input end of the output module.

5. The computing power network performance prediction method according to claim 4 is characterized in that: The output module is a multi-layer perception network composed of 2 linear layers.

6. The computing power network performance prediction method according to claim 5 is characterized in that: The step 3 comprises: Pre-training the computing power network performance prediction model using the training set to obtain a pre-trained computing power network performance prediction model, wherein the weight of the backbone network module in the pre-trained computing power network performance prediction model is locked; A self-supervised loss function of the test time training layer in the computing power network performance prediction model is constructed in a self-supervised manner, the test time training layer is trained based on the self-supervised loss function, and the weight of the test time training layer is updated to adapt to changes in the network performance pattern, so as to obtain a trained computing power network performance prediction model.

7. The computing power network performance prediction method according to claim 6, characterized in that: The self-supervised loss function is: in, represents the self-supervised loss function, represents the weights of the training layer at test time, represents the input of the test-time training layer, , represents a low-rank matrix, represents the model of the test-time trained layers, Represents the frequency domain reconstruction target.

8. A computing power network performance prediction device, characterized in that: include: The acquisition module is used to obtain the historical performance data of the computing network as a training set; A construction module is used to construct a computing power network performance prediction model, wherein the computing power network performance prediction model includes an embedding module for embedding input data, a Mamba backbone network module for capturing nonlinear features of the data after the embedding operation, and an output module for converting the captured nonlinear features into prediction results, wherein the Mamba backbone network module includes a test time training layer; A training module, used for inputting the training set into the computing power network performance prediction model, pre-training the computing power network performance prediction model, obtaining a pre-trained computing power network performance prediction model, wherein the weight of the backbone network module in the pre-trained computing power network performance prediction model is locked, and training the test time training layer in the pre-trained computing power network performance prediction model, and updating the weight of the test time training layer to adapt to changes in the network performance mode, thereby obtaining a trained computing power network performance prediction model; The prediction module is used to input the network performance data of the target computing power network into the trained computing power network performance prediction model for prediction to obtain the performance prediction result of the target computing power network.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the computing power network performance prediction method as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the computing power network performance prediction method as described in any one of claims 1 to 7 is implemented.

Images