Performance index prediction method and device, storage medium and electronic device

Abstract

The embodiment of the present application provides a method and device for predicting performance indicators, a storage medium and an electronic device. The method includes: obtaining a target performance prediction model sent by a server through a target device; obtaining a storage performance feature sequence of the target device within a performance detection cycle; inputting multiple storage performance features into a variable selection network in the target performance prediction model in sequence, obtaining a set of indicator weights that match each of the multiple storage performance features, and determining multiple target storage performance features based on the multiple storage performance features and the respective matching indicator weight sets; inputting a target performance feature sequence composed of multiple target storage performance features into a temporal convolutional network in the target performance prediction model, and obtaining a performance indicator prediction result at the target moment.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computers, and more specifically, to a method and device for predicting a performance indicator, a storage medium, and an electronic device. Background Art

[0002] With the rapid development of information technology, data storage has become a core component of various applications and services. In order to ensure the efficient and stable operation of the storage system, it is particularly important to predict its performance indicators. Storage performance indicator prediction refers to predicting storage performance in the future by analyzing and learning historical data, such as CPU utilization, data throughput, I / O latency, etc. This prediction helps enterprises understand and evaluate the operating status of the storage system in advance, so as to make timely optimization and adjustments to avoid service interruptions or data loss caused by performance bottlenecks.

[0003] In order to effectively monitor and predict storage performance, cloud monitoring systems are widely used. Cloud monitoring systems can collect, analyze and display various performance indicators of storage systems in real time, helping administrators to fully understand the operating status of storage systems. However, as the data generated by the edge is not only large in scale, but also usually generated at a high rate, it needs to be processed and analyzed in a timely and accurate manner. When this data is collected in the cloud, it places extremely high demands on the computing power of the cloud. In addition, in the process of monitoring and predicting storage performance, since there are a large number of performance indicators that can be used for storage performance monitoring and prediction, if the raw data of all performance indicators is processed and analyzed, there will be a technical problem of low prediction efficiency.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present application provide a method and device for predicting a performance indicator, a storage medium, and an electronic device, so as to at least solve the technical problem that the method for predicting storage performance in the related art is inefficient.

[0006] According to one aspect of an embodiment of the present application, a performance indicator prediction method is provided, comprising: obtaining, through a target device, a target performance prediction model sent by a server, wherein a target model parameter set of the target performance prediction model is determined by the server according to a reference model parameter set returned by multiple edge devices; obtaining a storage performance feature sequence of the target device within a performance detection period, wherein the storage performance feature sequence includes storage performance features that match multiple time nodes respectively, the multiple time nodes are time nodes within the performance detection period, the storage performance features include indicator performance sub-features that match multiple storage performance indicators respectively, and the indicator performance sub-features are used to indicate that at the time nodes that match them, the storage performance of the target device The performance of the energy index; inputting the multiple storage performance characteristics into the variable selection network in the target performance prediction model in turn to obtain the indicator weight sets matching the multiple storage performance characteristics respectively, and determining the multiple target storage performance characteristics according to the multiple storage performance characteristics and the respective matching indicator weight sets, wherein the indicator weight set includes the indicator weights corresponding to the multiple indicator performance sub-characteristics respectively; inputting the target performance characteristic sequence composed of the multiple target storage performance characteristics into the time convolution network in the target performance prediction model to obtain the performance indicator prediction result at the target moment, wherein the time convolution network is used to determine the time dependency between at least two of the multiple target storage performance characteristics.

[0007] According to another aspect of an embodiment of the present application, a training method for a performance indicator prediction model is provided, comprising: constructing a first reference prediction model in a training stage, wherein the first reference prediction model includes a reference variable selection network in a training stage and a reference time convolutional network in a training stage, the first reference prediction model is described according to a first model parameter set, the variable selection network is used to determine the indicator weights corresponding to the indicator performance sub-features respectively matching multiple storage performance indicators in the storage performance feature, and the indicator performance sub-features are used to indicate the performance of the storage performance indicators of the edge device; when at least one target device for training the first reference prediction model is determined from multiple edge devices, the first reference prediction model is sent to the target device, wherein the first reference prediction model is described according to the first model parameter set; obtaining a second model parameter set returned by the multiple target devices, wherein the second model parameter set is used to describe a second reference prediction model, and the second reference prediction model is obtained by training the first reference prediction model by the target device based on a training sample library; and determining a target model parameter set for describing a target performance prediction model according to at least one of the second model parameter sets.

[0008] According to another aspect of the embodiment of the present application, a performance indicator prediction device is also provided, including:

[0009] A first acquisition unit is used to acquire, through a target device, a target performance prediction model sent by a server, wherein target model parameters of the target performance prediction model are determined by the server according to reference model parameters returned by multiple edge devices;

[0010] A second acquisition unit is used to acquire a storage performance feature sequence of the target device within a performance detection period, wherein the storage performance feature sequence includes storage performance features that are matched with a plurality of time nodes respectively, the plurality of time nodes are time nodes within the performance detection period, the storage performance features include indicator performance sub-features that are matched with a plurality of storage performance indicators respectively, and the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the target device at the time nodes that are matched therewith;

[0011] A feature processing unit, used to sequentially input the plurality of storage performance features into the variable selection network in the target performance prediction model, obtain an indicator weight set that matches each of the plurality of storage performance features, and determine a plurality of target storage performance features according to the plurality of storage performance features and the respective matching indicator weight sets, wherein the indicator weight set includes indicator weights corresponding to each of the plurality of indicator performance sub-features;

[0012] A prediction unit is used to input a target performance feature sequence consisting of multiple target storage performance features into the temporal convolutional network in the target performance prediction model to obtain a performance indicator prediction result at the target moment, wherein the temporal convolutional network is used to determine the time dependency between at least two of the multiple target storage performance features.

[0013] Optionally, the feature processing unit comprises: an acquisition module for acquiring a current storage performance feature from a plurality of the above-mentioned storage performance features, wherein the current storage performance feature is a current feature matrix composed of a plurality of feature vectors, and the plurality of feature vectors are respectively used to characterize one of the above-mentioned indicator performance sub-features; an input module for inputting the current storage performance feature into an indicator weight module in the above-mentioned variable selection network to obtain a current weight vector matching the current storage performance feature, wherein the current weight vector includes the indicator weights respectively matching the plurality of the above-mentioned feature vectors; and a calculation module for determining the product result between the current weight vector and the current feature matrix as the target storage performance feature matching the current storage performance feature.

[0014] Optionally, the above-mentioned input module is used to: process the above-mentioned current storage performance characteristics through a feature conversion function to obtain a reference storage performance characteristic, wherein the above-mentioned feature conversion function includes one of the following: a linear transformation function, a multivariate perceptron function; process the above-mentioned reference storage performance characteristics through a sparse normalization function to obtain the above-mentioned current weight vector.

[0015] Optionally, the input module is used to: obtain the product result between the Lagrange multiplier and the first reference vector to obtain a second reference vector, wherein the vector elements included in the first reference vector are all 1s; determine the vector difference between the reference storage performance characteristic and the second reference vector as the first intermediate storage performance characteristic; convert each vector element in the intermediate storage performance characteristic into the corresponding reciprocal square root to obtain the second intermediate storage performance characteristic; update the values ​​of the vector elements in the second intermediate storage performance characteristic whose values ​​are less than 0 to 0 to obtain the current weight vector.

[0016] Optionally, the prediction unit is used to: input the current performance feature sequence into the current feature processing module in the temporal convolutional network, wherein the current performance feature sequence is obtained by processing the target performance feature sequence via at least one of the feature processing modules, the temporal convolutional network includes a plurality of stacked feature processing modules, and the feature processing module includes: two dilated causal convolution submodules, two layer regularization submodules, and an activation function submodule located between the two dilated causal convolution submodules; obtain the current output result of the current feature processing module based on the current performance feature sequence; when the current feature processing module is connected to the next feature processing module, input the current output result into the next feature processing module; when the current feature processing module is not connected to the next feature processing module, process the current output result according to the linear transformation layer to obtain the performance indicator prediction result at the target moment.

[0017] Optionally, the prediction unit is used to: obtain a convolution kernel and a dilation factor that match the current feature processing module; perform a convolution operation based on the convolution kernel, the current performance feature sequence, and the dilation factor, wherein the dilation factor is used to skip the input value corresponding to the current moment, and apply the operation area corresponding to the convolution operation to the target feature area, and the step range corresponding to the target feature area is larger than the step range corresponding to the convolution kernel.

[0018] Optionally, the second acquisition unit comprises: a feature extraction module, used to acquire the multiple time nodes within the performance detection cycle, and multiple indicator parameters respectively matching the multiple time nodes, wherein the multiple indicator parameters respectively correspond to one storage performance indicator, and the multiple storage performance indicators include at least one of the following: a read / write rate indicator, a bandwidth indicator, a latency indicator, a CPU utilization indicator, and a data throughput indicator, wherein the read / write rate indicator is used to indicate the number of read / write operations performed by the target device within a unit time; according to each of the time nodes, normalizing the multiple indicator parameters corresponding to the time nodes to obtain indicator characteristic values ​​respectively matching the multiple indicator parameters; determining the storage performance characteristics matching the time nodes according to the multiple indicator characteristic values; and determining the storage performance characteristics corresponding to the multiple time nodes as the storage performance characteristic sequence.

[0019] Optionally, the feature extraction module is used to: obtain a current time parameter from a plurality of time parameters corresponding to the time node, wherein the plurality of time parameters are used to describe the time node; convert the current time parameter into a current time feature value according to a time period corresponding to the current time parameter; and determine a time feature vector indicating the time node according to the current time feature values ​​corresponding to each of the plurality of time parameters.

[0020] Optionally, the feature extraction module is used to: obtain the vector dimension of the time feature vector used to characterize the time node; convert the multiple indicator feature values ​​into indicator feature vectors matching the vector dimension, wherein the indicator feature vectors are used to indicate the indicator performance sub-features; and determine the feature matrix composed of the multiple indicator feature vectors as the storage performance feature matching the time node.

[0021] Optionally, the prediction device for the above-mentioned performance indicators also includes: a training unit, used to obtain a first reference prediction model in a training stage from the above-mentioned server, wherein the above-mentioned first reference prediction model includes a reference variable selection network in a training stage and a reference time convolutional network in a training stage, and the above-mentioned first reference prediction model is described according to a first model parameter set; obtaining a training sample library determined according to the historical operating status of the above-mentioned target device, wherein the above-mentioned training sample library includes training samples respectively matched with multiple historical time nodes, and the above-mentioned training samples include a historical storage performance feature sequence matching the above-mentioned historical time node, and a historical label storage performance feature matching the above-mentioned historical time node; training the above-mentioned first reference prediction model according to the above-mentioned training sample library to obtain a second reference prediction model that meets local training conditions; and sending a second model parameter set used to indicate the above-mentioned second reference prediction model to the above-mentioned server.

[0022] Optionally, the training unit is used to: in the current training round, obtain a reference training loss gradient set corresponding to a training sample subset matching the current training round, wherein the training sample library includes multiple training sample subsets, and the reference training loss gradient set includes first training loss gradients corresponding to multiple training samples in the training sample subset; perform gradient clipping on the first training loss gradients in the reference training loss gradient set to obtain second training loss gradients corresponding to the first training loss gradients; determine a third training loss gradient based on reference noise information and multiple second training loss gradients; and adjust the first model parameter set of the first reference prediction model based on the third training loss gradient to obtain the updated first model parameter set corresponding to the current training round.

[0023] Optionally, the training unit is used to: obtain a first ratio between the second-order norm corresponding to the first training loss gradient and a target constant; when the first ratio is greater than or equal to 1, determine a second ratio between the first training loss gradient and the first ratio as the second training loss gradient; when the first ratio is less than 1, determine the first training loss gradient as the second training loss gradient.

[0024] Optionally, the above-mentioned training unit is used to: obtain the above-mentioned reference noise information determined according to the normally distributed random variable; determine the weighted summation result between multiple second training loss gradients as the third reference noise gradient; determine the weighted summation result between the above-mentioned third reference noise gradient and the above-mentioned reference noise information as the above-mentioned third training loss gradient.

[0025] Optionally, the training unit is used to: determine the product of the target learning rate and the third training loss gradient as the parameter adjustment degree corresponding to the current training round; adjust the model parameters in the first model parameter set according to the parameter adjustment degree to obtain the updated first model parameter set.

[0026] Optionally, the training unit is used to: when the third reference prediction model determined according to the second reference prediction model meets the target training conditions, obtain the target performance prediction model sent by the server, wherein the model parameters of the third reference prediction model are obtained by the server after processing the reference model parameters sent by multiple edge devices based on the federal averaging algorithm, and the ratio of the first number of edge devices sending the reference model parameters to the second number of devices in the edge device set is a target value, and the edge device set includes all edge devices connected to the server; when the third reference prediction model determined according to the second reference prediction model does not meet the target training conditions, obtain the third reference prediction model sent by the server; and train the third prediction model according to the training sample library.

[0027] Optionally, the above-mentioned training unit is used to: obtain first encryption parameter information from the target encryption channel; decrypt the above-mentioned first encryption parameter information to obtain the above-mentioned first model parameter set used to indicate the above-mentioned first reference prediction model; encrypt the above-mentioned second model parameter set to obtain second encryption parameter information; and send the above-mentioned second encryption parameter information to the above-mentioned server through the above-mentioned target encryption channel.

[0028] According to another aspect of an embodiment of the present application, a training device for a performance indicator prediction model is also provided, comprising: a construction unit, configured to construct a first reference prediction model in a training stage, wherein the first reference prediction model includes a reference variable selection network in a training stage and a reference time convolutional network in a training stage, the first reference prediction model is described according to a first model parameter set, the variable selection network is used to determine the indicator weights corresponding to the indicator performance sub-features respectively matched by multiple storage performance indicators in the storage performance feature, and the indicator performance sub-features are used to indicate the performance performance of the storage performance indicator of the edge device; a sending unit, configured to send the first reference prediction model to the target device when at least one target device for training the first reference prediction model is determined from multiple edge devices, wherein the first reference prediction model is described according to the first model parameter set; an acquisition unit, configured to acquire a second model parameter set returned by the multiple target devices, wherein the second model parameter set is used to describe the second reference prediction model, and the second reference prediction model is obtained by the target device training the first reference prediction model based on a training sample library; a training unit, configured to determine a target model parameter set for describing a target performance prediction model according to at least one of the second model parameter sets.

[0029] According to another aspect of an embodiment of the present application, a computer-readable storage medium is further provided, in which a computer program is stored, wherein the computer program is configured to execute the above-mentioned performance indicator prediction method or the above-mentioned performance indicator prediction model training method at runtime.

[0030] According to another aspect of the embodiments of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the above performance indicator prediction method or the above performance indicator prediction model training method.

[0031] According to another aspect of an embodiment of the present application, there is also provided an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the above-mentioned performance indicator prediction method or the above-mentioned performance indicator prediction model training method through the computer program.

[0032] Through the above-mentioned implementation mode of the present application, the target performance prediction model sent by the server can be obtained through the target device; the storage performance feature sequence of the target device within the performance detection cycle is obtained; multiple storage performance features are sequentially input into the variable selection network in the target performance prediction model to obtain a set of indicator weights that match each of the multiple storage performance features, and multiple target storage performance features are determined based on the multiple storage performance features and the respective matching indicator weight sets; the target performance feature sequence composed of multiple target storage performance features is input into the temporal convolutional network in the target performance prediction model to obtain the performance indicator prediction result at the target moment.

[0033] In the above-mentioned implementation mode of the present application, the server sends the prediction model for predicting the storage performance of the target device to the target terminal, thereby avoiding the need for the server to centrally receive the operating data of multiple devices and perform analysis based on a large number of devices, thereby reducing the operating pressure of the server; in addition, the performance prediction of each target device is only based on local data, thereby avoiding the data privacy risk brought by sending local data to the server; finally, the target performance prediction model sent by the server includes a variable selection network, thereby realizing the indicator weight set determined by the variable selection network, and determining the performance indicators that are more critical to the prediction task from multiple indicator performance sub-features, thereby improving the efficiency of storage performance prediction, and solving the technical problem of low efficiency in predicting storage performance in related technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 It is a hardware structure block diagram of a server device of a performance indicator prediction method according to an embodiment of the present application;

[0036] Figure 2 is a flow chart of a method for predicting a performance indicator according to an embodiment of the present application;

[0037] Figure 3 is a schematic diagram of a method for predicting a performance indicator according to an embodiment of the present application;

[0038] Figure 4 is a schematic diagram of another method for predicting performance indicators according to an embodiment of the present application;

[0039] Figure 5 is a schematic diagram of another method for predicting a performance indicator according to an embodiment of the present application;

[0040] Figure 6 is a schematic diagram of another method for predicting a performance indicator according to an embodiment of the present application;

[0041] Figure 7 is a flow chart of a training method for a performance indicator prediction method according to an embodiment of the present application;

[0042] Figure 8 is a structural schematic diagram of a performance indicator prediction device according to an embodiment of the present application;

[0043] Fig. 9 is a schematic diagram of the structure of an electronic device for predicting performance indicators according to an embodiment of the present application;

[0044] Fig.10 is a structural schematic diagram of a training device for a performance indicator prediction method according to an embodiment of the present application;

[0045] Fig.11 It is a structural schematic diagram of a training electronic device for a performance indicator prediction method according to an embodiment of the present application. DETAILED DESCRIPTION

[0046] The embodiments of the present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0047] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0048] The method embodiments provided in the embodiments of the present application can be executed in a server device or a similar computing device. Taking running on a server device as an example, Figure 1 1 is a hardware structure block diagram of a server device of a performance indicator prediction method according to an embodiment of the present application. Figure 1 As shown, the server device may include one or more ( Figure 1 Only one is shown) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data, wherein the above-mentioned server device may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above server device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0049] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the data processing method of the memory in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, the above method is implemented. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the server device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0050] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the server device. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0051] As an optional implementation, Figure 2 As shown, the above performance indicator prediction method can be applied to an edge device, including:

[0052] S202, obtaining, through the target device, a target performance prediction model sent by the server, wherein a target model parameter set of the target performance prediction model is determined by the server according to a reference model parameter set returned by multiple edge devices;

[0053] S204, obtaining a storage performance feature sequence of the target device within a performance detection period, wherein the storage performance feature sequence includes storage performance features that are matched with a plurality of time nodes respectively, the plurality of time nodes are time nodes within the performance detection period, the storage performance features include indicator performance sub-features that are matched with a plurality of storage performance indicators respectively, the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the target device at the time nodes that are matched therewith;

[0054] S206, sequentially inputting the plurality of storage performance characteristics into the variable selection network in the target performance prediction model to obtain a set of indicator weights that match each of the plurality of storage performance characteristics, and determining a plurality of target storage performance characteristics according to the plurality of storage performance characteristics and the respective matching indicator weight sets, wherein the indicator weight set includes indicator weights corresponding to each of the plurality of indicator performance sub-characteristics;

[0055] S208, inputting a target performance feature sequence consisting of multiple target storage performance features into a temporal convolutional network in a target performance prediction model to obtain a performance indicator prediction result at a target moment, wherein the temporal convolutional network is used to determine a temporal dependency relationship between at least two of the multiple target storage performance features.

[0056] It should be noted that, in this embodiment, the target device may be one of the multiple edge devices that establish a communication connection with the server. In this embodiment, the multiple edge devices may be device nodes in the target distributed storage system for storing and processing data, and each edge device stores and processes the data received locally, wherein the server may be a cloud server in the target distributed storage system for managing and controlling multiple edge devices. For example Figure 3 As shown, the cloud server 300 can communicate with multiple edge devices including edge device 301, edge device 302 and edge device 303, and receive storage performance monitoring and prediction results returned by each edge device. It can be understood that the above Figure 3 Multiple edge devices in the network can predict local storage performance indicators through the above steps S202 to S208 and return the prediction results to the cloud server 300.

[0057] The target performance prediction model can be the above Figure 3The cloud server 300 in the example 1 determines the prediction model based on the training results of each edge device. The cloud server 300 can also be as follows based on the interaction process between each edge device. Figure 3 As shown, the cloud server 300 performs model parameter distribution and scheduling tasks, receives parameters from multiple edge devices (including edge device 301, edge device 302, and edge device 303), updates the global model according to the parameter aggregation algorithm, and then determines the trained model as the target performance prediction model.

[0058] Furthermore, in the above step S204, before predicting the performance indicators at the target moment according to the above target performance prediction model, the storage performance feature sequence of the performance monitoring period can be obtained first. Optionally, the above performance monitoring period can be a period of time before the current moment, and the target moment can be a time point after the current moment. Then, the storage performance indicators corresponding to the time nodes after the current time node are predicted by using multiple storage performance features monitored in the historical time period and respectively matching multiple historical time nodes.

[0059] In a storage performance feature, an indicator performance sub-feature that matches multiple storage performance indicators is further included. In this embodiment, the above storage performance indicator may include but is not limited to at least one of the following performance parameter indicators:

[0060] IOPS (Input / Output Operations Per Second): The number of input / output operations per second, which is used to measure the number of read and write requests that a storage system can handle per second;

[0061] Throughput: It is measured in megabytes per second (MB / s) or gigabytes per second (GB / s) and is used to measure the amount of data that a storage system can transfer per unit time.

[0062] Latency: It indicates the time from sending a request to receiving a response, usually in milliseconds (ms) or microseconds (µs);

[0063] Bandwidth: It is used to indicate the maximum data transmission rate that the storage system can handle per unit time.

[0064] Response Time: It indicates the time from when a request is sent to when it is fully processed, including service time and delay;

[0065] Data Transfer Rate: The speed at which data is actually transferred in a storage system.

[0066] Cache Hit Rate: An indicator used to measure the effectiveness of cache in a storage system, that is, the ratio of requested data being found in the cache;

[0067] Availability: It indicates the ratio of the available time of the storage system to the total time. High availability means that the storage system rarely fails.

[0068] Reliability: It is used to indicate the probability that a storage system can work normally under specified conditions and within a specified time.

[0069] It should be noted that in the above-mentioned implementation mode of the present application, an indicator performance sub-feature may be a feature vector determined according to a parameter value corresponding to the above-mentioned performance indicator, and further, feature vectors determined respectively according to the parameter values ​​corresponding to multiple performance indicators may constitute the above-mentioned storage performance feature.

[0070] In an optional implementation, the above step S204, obtaining the storage performance feature sequence of the target device within the performance detection period, includes:

[0071] S1, obtaining multiple time nodes within a performance detection cycle, and multiple indicator parameters respectively matching the multiple time nodes, wherein the multiple indicator parameters respectively correspond to a storage performance indicator, and the multiple storage performance indicators include at least one of the following: a read / write rate indicator, a bandwidth indicator, a latency indicator, a CPU utilization indicator, and a data throughput indicator, wherein the read / write rate indicator is used to indicate the number of read / write operations performed by the target device within a unit time;

[0072] S2, according to each time node, normalizing multiple indicator parameters corresponding to the time node to obtain indicator characteristic values ​​that match the multiple indicator parameters respectively;

[0073] S3, determining storage performance characteristics matching the time node according to multiple indicator characteristic values;

[0074] S4, determining the storage performance characteristics corresponding to the multiple time nodes as a storage performance characteristic sequence.

[0075] It is understandable that in the above implementation of the present application, the indicator parameter sets that match the multiple time nodes can be obtained first, and the multiple indicator parameters included in the above indicator parameter set are respectively mapped to corresponding indicator feature values ​​through normalization processing operations, and then the storage performance characteristics that match the time nodes are determined based on the multiple indicator feature values. It is understandable that after obtaining the above indicator feature values, the indicator feature values ​​can be further vectorized to obtain feature vectors for indicating indicator performance sub-features, and then the multiple indicator performance sub-features are spliced ​​to obtain the above storage performance characteristics.

[0076] The following combination Figure 4 The relationship between the above storage performance feature sequence, storage performance features and index performance sub-features is specifically described. Figure 4 In the embodiment, the storage performance feature sequence 400 may be a matrix set, including multiple feature matrices for indicating storage performance features. It is understandable that the multiple storage performance features in the above storage performance feature sequence 400 may correspond to multiple historical time nodes respectively, that is, each storage performance feature is used to indicate the storage performance of the above target device at the corresponding historical time node. Furthermore, in the feature matrix corresponding to the storage performance feature 401, each column vector may correspond to an indicator performance sub-feature respectively, and the multiple indicator performance sub-features include the indicator performance sub-feature 402. Assuming that the storage performance feature 401 corresponds to the first historical time node, and the indicator performance sub-feature 402 corresponds to the IOPS indicator, it is understandable that the indicator performance sub-feature 402 may specifically be a feature vector obtained by vectorizing the IOPS measurement value of the above target device at the above first historical time node.

[0077] In an optional implementation, after acquiring multiple time nodes in the performance detection period and multiple indicator parameters respectively matching the multiple time nodes, the method further includes:

[0078] S1, obtaining a current time parameter from among a plurality of time parameters corresponding to a time node, wherein the plurality of time parameters are used to describe the time node;

[0079] S2, converting the current time parameter into a current time characteristic value according to the time period corresponding to the current time parameter;

[0080] S3, determining a time feature vector for indicating a time node according to current time feature values ​​corresponding to each of the multiple time parameters.

[0081] It should be noted that, in this embodiment, multiple time nodes can be processed to obtain a time feature vector for characterizing the time node. Specifically, multiple time parameters corresponding to the time node and the time period corresponding to the time parameter can be mapped to corresponding time feature values, and then the multiple time feature values ​​can be spliced ​​to obtain a time feature vector for characterizing the time node.

[0082] In a specific implementation, an entity embedding method can be used to map discrete time nodes into time feature vectors. For example, taking the performance test time node "09 / 20 / 21 / 30" as an example, the time parameter corresponding to "09" is "September", and the time period corresponding to the month parameter is "12", and then the time feature vector corresponding to "month" can be determined based on 9 / 12 (i.e., the ratio between 4 and 12);

[0083] The time parameter corresponding to "20" is "20 days", and the time period corresponding to the number of days in a month is "30". Then, the time feature vector corresponding to "date" can be determined based on 20 / 30 (i.e., the ratio between 20 and 30);

[0084] In another optional manner, it can also be determined that the week number corresponding to "20th" is the 5th day of the 4th week of September, and then the time feature vector corresponding to "week" can be determined according to 4 (the ratio between 4 and 5), and the time feature vector corresponding to "week N" can be determined according to 5 / 7 (the ratio between 5 and 7);

[0085] Furthermore, in the process of determining the time feature vector corresponding to 4 / 7, the corresponding time feature vector can also be directly obtained according to the 4 / 7 mapping. For example, a preset mapping relationship can be set: 1 / 7 corresponds to the first feature vector (0000001), 2 / 7 corresponds to the second feature vector (0000020), 3 / 7 corresponds to the third feature vector (0000100), 4 / 7 corresponds to the fourth feature vector (0010000), 5 / 7 corresponds to the fifth feature vector (0010000), 6 / 7 corresponds to the fifth feature vector (0100000), 7 / 7 corresponds to the seventh feature vector (1000000), and then the time feature vector corresponding to 5 / 7 is determined to be the fifth feature vector (0010000).

[0086] In another optional implementation, the time parameters in the login time can be encoded as a week feature vector week embedding, an hour feature vector hour embedding, and a second feature vector second embedding, that is, the hours corresponding to each week and each day and the seconds corresponding to each minute of the time corresponding to the performance detection time node are encoded into separate embedding vectors (7 days in a week can correspond to 7 embeddings, 24 hours in a day can correspond to 24 embeddings, and 60 seconds in a minute can correspond to 60 embeddings).

[0087] In this implementation, a sine function may be specifically used to encode the time feature, and the specific encoding method is as follows:

[0088] ,

[0089] here The maximum possible value of the time parameter t. For example, the maximum value of seconds is 60, the maximum value of days is 24, and the maximum value of weeks is 7.

[0090] Continue to take "09 / 20 / 21 / 30" as an example for description. As can be seen in the above embodiment, the date indicated by "09 / 20" is the 5th day of the 4th week of September.

[0091] Furthermore, for the week feature vector week embedding In terms of , and then determine , and then determine the week embedding as a vector ;

[0092] Furthermore, for the hour feature vector hour embedding In terms of , and then determine , and then determine the hour embedding as a vector [ ];

[0093] Furthermore, for the second feature vector second embedding In terms of , and then determine , and then determine the second embedding as vector [ ].

[0094] By concatenating the above week feature vector week embedding, hour feature vector hour embedding and second feature vector second embedding in sequence, we can get the feature vector [ ].

[0095] Through the above-mentioned implementation mode of the present application, the periodic distribution characteristics of the detection time nodes can be accurately characterized, thereby realizing the accurate identification of the periodic performance changes of the device storage performance, thereby improving the prediction accuracy of the performance indicator prediction based on the above-mentioned time feature vector.

[0096] In an optional implementation manner, determining the storage performance characteristics matching the time node according to the multiple indicator characteristic values ​​includes:

[0097] S1, obtaining the vector dimension of the time feature vector used to characterize the time node;

[0098] S2, converting the multiple indicator feature values ​​into indicator feature vectors matching the vector dimension, wherein the indicator feature vectors are used to indicate the indicator performance sub-features;

[0099] S3, determines the feature matrix composed of multiple indicator feature vectors as the storage performance feature matching the time node.

[0100] In this embodiment, a linear transformation can be used to convert each indicator characteristic value into a vector dimension corresponding to the time characteristic vector. Matching indicator performance sub-characteristics.

[0101] A specific example of a method for obtaining the above storage performance characteristics is described below.

[0102] Before performing the performance prediction operation, the storage performance parameter set obtained during the performance detection cycle may be preprocessed. The preprocessing may include but is not limited to performing a data cleaning operation on the storage performance parameter set. For example, 3 The abnormal data were removed according to the criterion and the missing values ​​were filled by linear interpolation to ensure the validity and completeness of the data.

[0103] Next, normalization is performed on each performance indicator parameter in the data set obtained after preprocessing to eliminate the dimensional differences between different performance indicators. In this implementation, the performance indicator data can be converted to the [0-1] range by Min-max normalization. The conversion formula is as follows:

[0104] ,

[0105] in, The indicator parameters for the current conversion process. is the minimum value in the parameter set corresponding to the current indicator parameter. is the maximum value in the parameter set corresponding to the current indicator parameter. is the normalized characteristic value of the indicator.

[0106] Next, for multiple indicator feature values ​​corresponding to the same time node, the multiple indicator feature values ​​can be combined into an indicator feature vector, and the target linear transformation matrix is ​​obtained, and then the product result between the above indicator feature vector and the target linear transformation matrix is ​​obtained to determine the storage performance characteristics matching the time node. Assume that the vector dimension of the above time feature vector is , the number of indicators is , and the storage performance characteristics determined by the above implementation can be expressed as .

[0107] Through the above-mentioned implementation mode of the present application, multiple time nodes within a performance detection cycle and multiple indicator parameters respectively matching the multiple time nodes are obtained; according to each time node, the multiple indicator parameters corresponding to the time node are normalized to obtain indicator characteristic values ​​respectively matching the multiple indicator parameters; storage performance characteristics matching the time node are determined according to the multiple indicator characteristic values; the storage performance characteristics corresponding to each of the multiple time nodes are determined as a storage performance characteristic sequence, thereby achieving accurate extraction of the storage performance characteristic sequence detected within the performance detection cycle, and improving the prediction accuracy of storage performance prediction for the target device.

[0108] The model structure of the target performance prediction model in the above-mentioned embodiment of the present application is specifically described below in conjunction with FIG5. A target performance prediction model as shown in FIG5 includes a variable selection network 501 and a temporal convolution network 502. The variable selection network 501 further includes an indicator weight module 501-1 and a feature adjustment network 501-2; the temporal convolution network 502 includes multiple stacked feature processing modules, including a feature processing module 502-1 and a feature processing module 502-N.

[0109] The following further describes the above step S206, which is a method for processing the storage performance characteristics based on the variable selection network. The above S206 further includes:

[0110] S1, obtaining a current storage performance characteristic from a plurality of storage performance characteristics, wherein the current storage performance characteristic is a current characteristic matrix composed of a plurality of characteristic vectors, and the plurality of characteristic vectors are respectively used to represent an indicator performance sub-characteristic;

[0111] S2, inputting the current storage performance characteristics into the indicator weight module in the variable selection network to obtain a current weight vector matching the current storage performance characteristics, wherein the current weight vector includes indicator weights respectively matching the multiple feature vectors;

[0112] S3, determining the product result between the current weight vector and the current feature matrix as the target storage performance feature that matches the current storage performance feature.

[0113] It can be understood that in the above embodiment, multiple storage performance characteristics in the storage performance characteristic sequence can be processed in sequence through the variable selection network to obtain the target storage performance characteristics after processing, and the target storage performance characteristic sequence determined by the multiple target storage performance characteristics can be input into the above-mentioned temporal convolutional network to further perform feature analysis, thereby obtaining the performance indicator prediction result.

[0114] Furthermore, the above-mentioned inputting the current storage performance characteristics into the indicator weight module in the variable selection network to obtain the current weight vector matching the current storage performance characteristics includes:

[0115] S1, processing the current storage performance characteristics through a feature conversion function to obtain a reference storage performance characteristic, wherein the feature conversion function includes one of the following: a linear transformation function, a multivariate perceptron function;

[0116] S2, processes the reference storage performance characteristics through a sparse normalization function to obtain the current weight vector.

[0117] Furthermore, the above sparse normalization function processes the reference storage performance characteristics as follows:

[0118] S1, processes the reference storage performance characteristics through the sparse normalization function to obtain the current weight vector, including:

[0119] S2, obtaining a product result between the Lagrange multiplier and the first reference vector to obtain a second reference vector, wherein all vector elements included in the first reference vector are all 1;

[0120] S3, determining a vector difference between the reference storage performance characteristic and the second reference vector as a first intermediate storage performance characteristic;

[0121] S4, converting each vector element in the intermediate storage performance characteristic into a corresponding reciprocal square root to obtain a second intermediate storage performance characteristic;

[0122] S5, updating the values ​​of the vector elements whose values ​​in the second intermediate storage performance characteristic are less than 0 to 0, and obtaining a current weight vector.

[0123] The following is a specific formula for Figure 5 The weight processing process performed in the indicator weight module 501-1 is explained.

[0124] ,

[0125] In the above formula, is the feature matrix used to characterize storage performance characteristics, A pre-specified variable conversion function, such as a linear transformation or a multivariate perceptron network; It can be a sparse normalization function, which is calculated as follows:

[0126] ,

[0127] middle, can be a Lagrange multiplier, is a vector of all 1s. The function is used to set the negative values ​​in the vector elements to 0. It should be noted that is a scalar value, and is a vector of all 1s, all its elements are 1, and then Indicates that the vector x Subtract each element in Furthermore, in the application Under the function, All elements less than 0 in will become 0, and non-negative elements remain unchanged. Further, Instructions The output of the function is subjected to the reciprocal square root operation. The output of is a positive number, then its reciprocal square root is a real number.

[0128] Through the above sparse normalization function, due to The function exists if Subtract the elements of If the number is negative, the contribution of these elements to the final output will be set to 0. This means that only those The elements are large enough to cancel the Lagrange constant Then, through the above sparse normalization function, we can The function can generate a sparse probability distribution. In the process of predicting performance indicators, the contribution of irrelevant features can be effectively reduced to zero, thereby improving the interpretability and efficiency of the model. It should be noted that in the above variable selection network, the reference storage performance features of all time steps can share the same set of parameters to achieve efficient use of parameters.

[0129] Based on the above sparse normalization function, the current weight vector is obtained In the case of And the current feature matrix The target storage performance characteristics can be obtained as shown in the following formula:

[0130] ,

[0131] in, represents the current weight vector, is the vector The components is the current feature matrix No. vectors.

[0132] Through the above-mentioned implementation mode of the present application, the current storage performance characteristics are obtained from multiple storage performance characteristics; the current storage performance characteristics are input into the indicator weight module in the variable selection network to obtain the current weight vector matching the current storage performance characteristics; the product result between the current weight vector and the current feature matrix is ​​determined as the target storage performance characteristics matching the current storage performance characteristics, and then a sparse indicator selection method is used to comprehensively process all input indicator characteristics, and all input variables are intelligently integrated according to the importance of the features, thereby effectively reducing the contribution of irrelevant features to zero, improving the efficiency of feature extraction, and significantly enhancing the interpretability of the model.

[0133] exist Figure 5 When the variable selection network 501 processes the storage performance feature sequence to obtain the target storage performance feature sequence, the output of the target storage performance feature sequence is then input into the temporal convolutional network 502. The feature processing and parsing process in the temporal convolutional network 502 is described below:

[0134] In the above step S208, the target performance feature sequence composed of multiple target storage performance features is input into the temporal convolutional network in the target performance prediction model to obtain the performance indicator prediction result at the target time, including:

[0135] S1, inputting the current performance feature sequence into the current feature processing module in the temporal convolutional network, wherein the current performance feature sequence is obtained by processing the target performance feature sequence via at least one feature processing module, the temporal convolutional network includes a plurality of stacked feature processing modules, and the feature processing module includes: two dilated causal convolution submodules, two layer regularization submodules, and an activation function submodule located between the two dilated causal convolution submodules;

[0136] S2, obtaining the current output result of the current feature processing module based on the current performance feature sequence;

[0137] S2-1, when the current feature processing module is connected to the next feature processing module, the current output result is input to the next feature processing module;

[0138] S2-2, when the current feature processing module is not connected to the next feature processing module, the current output result is processed according to the linear transformation layer to obtain the performance indicator prediction result at the target moment.

[0139] It can be understood that in the temporal convolutional network 502, the temporal convolutional network receives the input of the variable selection network, extracts the temporal dependency between the integrated variables at different time steps and predicts the output value of the performance indicator. The temporal convolutional network (TCN) is composed of stacked residual blocks and output layers based on dilated causal convolutions, and the structure of one of the feature processing modules is as follows Figure 6 As shown in the feature processing module 601, the residual block consists of two dilated causal convolutional layers, two layer regularizations and a ReLU activation function in the middle.

[0140] Furthermore, after the current performance feature sequence is input into the current feature processing module in the temporal convolutional network, the method further includes:

[0141] S1, obtain the convolution kernel and dilation factor that match the current feature processing module;

[0142] S2, performing a convolution operation based on the convolution kernel and the current performance feature sequence, and a dilation factor, wherein the dilation factor is used to skip the input value corresponding to the current moment, and to apply the operation area corresponding to the convolution operation to the target feature area, and the step range corresponding to the target feature area is larger than the step range corresponding to the convolution kernel.

[0143] The dilated causal convolution module in the feature processing module 601 can be a strictly time-constrained convolution network, in which the output of the next layer at time t depends only on the value of the previous layer at time t and before, and it skips the input value with a certain step size (dilation factor d) and applies the convolution kernel to the area exceeding its own size, so that it can have a larger receptive field even when the number of stacked layers is small, thus achieving effective optimization of the number of parameters. The convolution kernel c, the calculation formula of the dilated causal convolution is as follows:

[0144] ,

[0145] In the above formula, the collision causal convolution operation ensures that the output Depends only on current and past input values , and does not depend on future input values. This is achieved by using zero-padding when the convolution kernel is applied, ensuring that the output time point does not "see" information after the current time point.

[0146] Expansion rate This allows the model to increase its receptive field without adding additional parameters, i.e., the model is able to capture dependencies over a larger time range. For example, if , then each weight in the convolution kernel is multiplied by the adjacent time points; if , then the weights in the convolution kernel will be multiplied by the input at every other time point, thereby increasing the time range that the model can consider.

[0147] It should be noted that in order for the receptive field of TCN to cover a historical sequence of T time steps, TCN needs at least After multiple convolutional layers and residual connection processing, the output layer of the TCN model will generate multiple performance indicator prediction values ​​through linear transformation, and these prediction values ​​correspond to the values ​​of the next moment of the input time series.

[0148] The following is an explanation of a method for training the target performance prediction model. Optionally, before the target device obtains the target performance prediction model sent by the server in step S202, the following further includes:

[0149] S1, obtaining a first reference prediction model in a training stage from a server, wherein the first reference prediction model includes a reference variable selection network in a training stage and a reference time convolution network in a training stage, and the first reference prediction model is described according to a first model parameter set;

[0150] S2, obtaining a training sample library determined according to the historical operating state of the target device, wherein the training sample library includes training samples respectively matching multiple historical time nodes, and the training samples include historical storage performance feature sequences matching the historical time nodes, and historical label storage performance features matching the historical time nodes;

[0151] S3, training the first reference prediction model according to the training sample library to obtain a second reference prediction model that meets the local training conditions;

[0152] S4: Send a second model parameter set indicating a second reference prediction model to the server.

[0153] In an optional implementation, the acquiring of the first reference prediction model in the training stage from the server includes: acquiring first encryption parameter information from the target encryption channel; decrypting the first encryption parameter information to obtain a first model parameter set for indicating the first reference prediction model;

[0154] The above-mentioned sending of the second model parameter set for indicating the second reference prediction model to the server includes: encrypting the second model parameter set to obtain second encrypted parameter information; and sending the second encrypted parameter information to the server through a target encrypted channel.

[0155] It is understandable that after the cloud server completes model building, it packages the model parameters and configuration information into a format that is easy to transmit and deploy, such as a compressed file or a specific model format, and then sends the model package to all participating edge devices through the established secure communication channel. After receiving the model, the device verifies the integrity and starts training using local data after confirming that it is correct.

[0156] In an optional implementation, during the training process, the first reference prediction model is trained according to the training sample library to obtain a second reference prediction model that meets the local training conditions, including:

[0157] S1, in the current training round, obtaining a reference training loss gradient set corresponding to a training sample subset matching the current training round, wherein the training sample library includes multiple training sample subsets, and the reference training loss gradient set includes first training loss gradients corresponding to multiple training samples in the training sample subsets respectively;

[0158] S2, performing gradient clipping on the first training loss gradients in the reference training loss gradient set to obtain second training loss gradients corresponding to the first training loss gradients;

[0159] S3, determining a third training loss gradient according to the reference noise information and the plurality of second training loss gradients;

[0160] S4: According to the third training loss gradient, adjust the first model parameter set of the first reference prediction model to obtain an updated first model parameter set corresponding to the current training round.

[0161] It is understandable that in this embodiment, after the local training of the target device is completed, the edge device needs to send the new local model parameters to the cloud server. However, even if the original data is not directly shared, some sensitive information may be leaked through the update of the model parameters. In order to further protect data privacy and security, this embodiment introduces a differential privacy mechanism to add random noise to the model gradient information to confuse the data. To this end, after completing the calculation of the gradient of the loss function of a batch, it is necessary to first clip the calculated gradient, and then add an appropriate amount of noise to the sum of the gradients of the batch to achieve differential privacy protection.

[0162] Specifically, the first training loss gradients in the reference training loss gradient set are respectively subjected to gradient clipping to obtain second training loss gradients corresponding to the first training loss gradients, including:

[0163] S1, obtaining a first ratio between the second-order norm corresponding to the first training loss gradient and the target constant;

[0164] S2, when the first ratio is greater than or equal to 1, determining a second ratio between the first training loss gradient and the first ratio as a second training loss gradient;

[0165] S3: When the first ratio is less than 1, determine the first training loss gradient as the second training loss gradient.

[0166] In this embodiment, the first training loss gradient is obtained by the following formula:

[0167] ,

[0168] In the above formula, the first training loss gradient is at the time step and samples,

[0169] The loss function calculated above About model weights ,

[0170] The gradient of . This gradient represents the rate of change of the loss function under the current weights and data points;

[0171] Furthermore, after obtaining the above-mentioned first training loss gradient, the method of clipping the first training loss gradient to obtain the second training loss gradient can be obtained by the following method:

[0172] ,

[0173] In the above formula, is the gradient of the first training loss The result of cropping. The specific cropping method is based on the first training loss gradient Divide by its 2-norm With constant If this ratio is greater than 1, then The ratio between and this ratio is determined as the second training loss gradient obtained after clipping, where, Set to the median norm of the current batch gradient. This ensures that the norm of the gradient does not exceed , thus limiting the size of the gradient and preventing excessive gradient updates.

[0174] Further, when the second training loss gradient is obtained, the third training loss gradient is determined according to the reference noise information and the plurality of second training loss gradients, including:

[0175] S1, obtaining reference noise information determined according to a normal distribution random variable;

[0176] S2, determining a weighted sum of multiple second training loss gradients as a third reference noise gradient;

[0177] S3: Determine a weighted sum of the third reference noise gradient and the reference noise information as a third training loss gradient.

[0178] Specifically, the third training loss gradient after adding noise can be obtained as follows.

[0179] ,

[0180] In the above formula, is the second training loss gradient after clipping The result of adding noise information to . The noise information can be a kind of information with a mean of 0 and a variance of A normally distributed random variable , is the identity matrix. The above noise will be added to the sum of gradients of the entire batch, thus achieving differential privacy.

[0181] When the third training loss gradient is obtained, the first model parameter set of the first reference prediction model is adjusted according to the third training loss gradient to obtain an updated first model parameter set corresponding to the current training round, including:

[0182] S1, determining the product of the target learning rate and the third training loss gradient as the parameter adjustment degree corresponding to the current training round;

[0183] S2, adjusting the model parameters in the first model parameter set according to the parameter adjustment degree to obtain an updated first model parameter set.

[0184] In this implementation, the local model parameter update formula considering differential privacy is as follows:

[0185] ,

[0186] in, is the learning rate.

[0187] Through the above implementation of the present application, in the current training round, a reference training loss gradient set corresponding to a subset of training samples matching the current training round is obtained; the first training loss gradients in the reference training loss gradient set are subjected to gradient clipping to obtain second training loss gradients corresponding to the first training loss gradients; the third training loss gradient is determined based on the reference noise information and multiple second training loss gradients; the first model parameter set of the first reference prediction model is adjusted based on the third training loss gradient to obtain an updated first model parameter set corresponding to the current training round, thereby integrating differential privacy technology in the local training process to further enhance data privacy protection. Differential privacy obfuscates the original gradient information by adding an appropriate amount of noise to mask the contribution of specific samples to the model update, thereby protecting individual privacy while still being able to extract useful statistical information.

[0188] In an optional implementation manner, after sending the reference model parameter for indicating the second reference prediction model to the server, the method further includes:

[0189] S1, when a third reference prediction model determined according to the second reference prediction model satisfies a target training condition, obtaining a target performance prediction model sent by a server, wherein a model parameter of the third reference prediction model is obtained by processing reference model parameters sent by multiple edge devices based on a federated average algorithm by the server, and a ratio between a first number of edge devices sending reference model parameters and a second number of devices in an edge device set is a target value, and the edge device set includes all edge devices connected to the server;

[0190] S2, when the third reference prediction model determined according to the second reference prediction model does not meet the target training condition, obtaining the third reference prediction model sent by the server;

[0191] S3: training a third reference prediction model according to the training sample library.

[0192] In this embodiment, after receiving the model parameters sent by the edge device, the cloud server uses the federated average algorithm to aggregate these parameters to obtain an updated global model. The global model update method is calculated as follows:

[0193] ,

[0194] in, For edge device subset The total number of training samples in for The updated global model parameters are then distributed to all participating edge devices. Each edge device uses the updated model parameters to continue the next round of local model training on its own local dataset.

[0195] In the cloud server, the effect of the trained third reference prediction model can also be evaluated according to a certain method, and then whether to continue training the model can be determined according to the evaluation result to improve the prediction accuracy. This embodiment uses the root mean square error (RMSE) to measure the difference between the model prediction value and the actual performance index value, and the formula is as follows:

[0196] ,

[0197] in, represents the number of samples, Indicates The actual value of the samples, Represents the model for The predicted value of a sample.

[0198] In addition, if the global training rounds are sufficient, the model accuracy can be further improved through personalized training. That is, based on the global model, each participant can perform local fine-tuning (LFT) on the model through local data to adapt to the characteristics of its local data and generate a personalized model. That is, after the federated training is completed, only a small number of rounds of local model iterative training are performed using local data.

[0199] In one embodiment of the present application, a prediction method of a storage performance prediction model is provided, comprising:

[0200] S1, constructing a first reference prediction model in a training stage, wherein the first reference prediction model includes a reference variable selection network in a training stage and a reference time convolution network in a training stage, the first reference prediction model is described according to a first model parameter set, the variable selection network is used to determine the indicator weights corresponding to the indicator performance sub-features respectively matched by the multiple storage performance indicators in the storage performance feature, and the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the edge device;

[0201] S2, when at least one target device for training the first reference prediction model is determined from the plurality of edge devices, sending the first reference prediction model to the target device, wherein the first reference prediction model is described according to the first model parameter set;

[0202] S3, obtaining a second model parameter set returned by multiple target devices, wherein the second model parameter set is used to describe a second reference prediction model, and the second reference prediction model is obtained by training the first reference prediction model by the target device based on a training sample library;

[0203] S4. Determine a target model parameter set for describing a target performance prediction model based on at least one second model parameter set.

[0204] The following combination Figure 7 The training process of a complete prediction model is specifically described. The implementation method of the present application can be applied to the training system of the prediction model shown in Figure 3. The overall architecture can be shown in Figure 3, including a cloud server 300 and M edge devices. (Which includes edge device 301, edge device 302 and edge device 303.) Cloud server 300 performs model parameter distribution and scheduling tasks, receives parameters from edge nodes, and updates the global model according to the parameter aggregation algorithm.

[0205] S702, the cloud server builds a performance prediction model in the training phase;

[0206] In step S702, a storage performance indicator prediction model based on variable selection and time convolutional network can be constructed through the cloud server. This prediction model is a multi-input and multi-output system. The input is the historical performance indicator data and time information of the storage system, and the output is the predicted value of multiple performance indicators at the next moment. The variable selection network can intelligently integrate all input variables according to the importance of the features, thereby significantly enhancing the interpretability of the model. At the same time, the introduction of the time convolutional network not only gives the model a powerful long-term time dependency modeling capability, but also achieves the optimization and reduction of the number of parameters. Through the above prediction model, the performance indicators of the storage system can be predicted more accurately, and the efficiency and reliability of the entire system can be improved.

[0207] Specifically, the model structure of the performance prediction model constructed by the cloud server is shown in FIG5, including a variable selection network 501 and a temporal convolutional network 502. The variable selection network 501 further includes an indicator weight module 501-1 and a feature adjustment network 501-2; the temporal convolutional network 502 includes multiple stacked feature processing modules, including a feature processing module 502-1 and a feature processing module 502-N.

[0208] The variable selection network 501 is used to improve the interpretability of the model and identify the importance of different features. This embodiment adopts a novel sparse variable selection method to comprehensively process all input variables. In this process, the present invention first uses entity embedding technology to convert discrete variables (such as hours, weeks, etc.) Then, each continuous performance indicator is also converted into a vector of dimension. The vector of dimension is used as its feature representation. After completing these transformations, the present invention applies variable selection technology to the transformed variables. Suppose the transformed feature variable is ,in yes No. variables, the integrated variable is calculated by the following formula:

[0209] ,

[0210] in, represents the variable weight vector, is the vector The components User-specified transformation function of variables, such as a linear transformation or a multivariate perceptron network. is a sparse normalization function, and its calculation formula is as follows:

[0211] ,

[0212] Where p is the Lagrange multiplier and I is an all-one vector. Compared with the traditional normalization transformation, It has significant advantages. The traditional normalization transformation always produces positive output, while It can generate a sparse probability distribution, effectively reducing the contribution of irrelevant features to zero. It is worth noting that in the variable selection network, the variables of all time steps share the same set of parameters to achieve efficient use of parameters.

[0213] The specific network structure of the temporal convolutional network 502 can be shown in FIG6. The temporal convolutional network 502 is used to receive the input of the variable selection network 501, extract the temporal dependency between the integrated variables at different time steps, and predict the output value of the performance indicator. The temporal convolutional network (TCN) is composed of a stacked residual block and an output layer based on dilated causal convolution. Its structure is as follows: Figure 6 As shown in . The residual block consists of two dilated causal convolutional layers, two regularization layers and a ReLU activation function in the middle. Dilated causal convolution is a strict time-constrained convolutional network. The output of the next layer at time t only depends on the value of the previous layer at time t and before. It skips input values ​​with a certain step size (dilation factor d) and applies the convolution kernel to an area that exceeds its own size. This allows for a larger receptive field even with a small number of stacked layers, achieving effective optimization of the number of parameters. For a size of The convolution kernel c, the calculation formula of the dilated causal convolution is as follows:

[0214] ,

[0215] In order to make the receptive field of TCN cover time steps of history sequence, TCN requires at least After multiple convolutional layers and residual connections, the output layer of the TCN model will generate multiple performance indicator prediction values ​​through linear transformation, which correspond to the values ​​of the next moment of the input time series.

[0216] After the model is built, the cloud server packages the model parameters and configuration information into a format that is easy to transmit and deploy, such as a compressed file or a specific model format, and then sends the model package to all participating edge devices through the established secure communication channel. After receiving the model, the device verifies the integrity and starts training using local data after confirming that it is correct.

[0217] S704, determining a target edge device from multiple edge devices;

[0218] S706, sending the constructed performance prediction model to the target edge device;

[0219] In the above steps S704 and S706, in the federated learning In round-by-round communication, the cloud server Proportionally randomly select a subset of edge devices , and send the latest global model parameters to Next, Each edge node in the To train local performance indicator prediction models;

[0220] S708, the target edge device trains the performance prediction model using local data;

[0221] During the training process, the edge device First, a local training sample set is constructed. During the construction process, the edge device Key performance indicator data, including IOPS, bandwidth, latency, and CPU utilization, can be monitored and collected in real time, and the time information of these data, such as hours, weeks, and months, can be recorded. Then, these data are sorted and stored according to the time series to form a complete time series data set. Next, the data set will be cleaned, the 3σ criterion will be used to remove abnormal data, and linear interpolation will be used to fill missing values ​​to ensure the validity and integrity of the data. In order to eliminate the dimensional differences between different performance indicators, this implementation method converts the performance indicator data to the [0-1] range through Min-max normalization. The conversion formula is as follows:

[0222] ,

[0223] For discrete time variables such as hours, weeks, and months, we use label encoding to represent them. For example, the hour variable uses integers from 0 to 23 to identify the hours of the day.

[0224] After preparing the preprocessed dataset In the case of , the edge device can use the initial global model parameters sent to the server to initialize the local model. In order to further protect data privacy and security, this implementation introduces a differential privacy mechanism to add random noise to the model gradient information to confuse the data. To this end, after completing the loss function gradient calculation for a batch, it is necessary to first clip the calculated gradient, and then add an appropriate amount of noise to the gradient sum of the batch to achieve differential privacy protection. For samples within a batch , Represents the current batch, and the specific calculation formula for gradient update is as follows:

[0225] ,

[0226] in represents the loss function, are the current model parameters, Representation sample The corresponding gradient, Set to the median norm of the current batch gradient, I is an all-one vector, represents a normal distribution, Represents the batch size. The local model parameter update formula considering differential privacy is as follows:

[0227] ,

[0228] in is the learning rate. In addition, in one communication, the local update process will run multiple rounds.

[0229] S710, the target edge device returns the model parameters of the trained performance prediction model to the cloud server;

[0230] It is understandable that after each edge device completes local training, each edge device needs to send new local model parameters to the cloud server.

[0231] S712, the cloud server aggregates parameters according to the model parameters returned by the multiple edge devices to obtain a reference performance prediction model;

[0232] After the cloud server receives the model parameters sent by the edge device, it uses the federated averaging algorithm to aggregate these parameters to obtain an updated global model. The global model update method is calculated as follows:

[0233] ,

[0234] in For edge device subset The total number of training samples in for The number of training samples.

[0235] Next, executing S714, referring to whether the performance prediction model meets the target condition;

[0236] If the target condition is not met, the process returns to step S706, and the updated global model parameters are then distributed to all participating edge devices. Each edge device uses the updated model parameters to continue the next round of local model training on its own local data set.

[0237] The above target conditions can be determined by the evaluation and optimization module, wherein the evaluation and optimization module uses the trained global model to predict the performance indicators of the storage system, compares the prediction results with the actual performance indicators, and evaluates the accuracy of the model. According to the evaluation results, it is judged whether to continue training the model to improve the prediction accuracy. The present invention uses the root mean square error (RMSE) to measure the difference between the model prediction value and the actual performance indicator value, and the formula is as follows:

[0238] ,

[0239] in, represents the number of samples, Indicates The actual value of the samples, Represents the model for The predicted value of a sample.

[0240] When the target conditions are met, S716 is executed to send the model to each edge device and fine-tune it based on the local data of each edge device;

[0241] Specifically, under the premise of sufficient global training rounds, the model accuracy can be further improved through personalized training, that is, based on the global model, each participant can perform local fine-tuning (LFT) on the model through local data to adapt to the characteristics of its local data and generate a personalized model. That is, after the federated training is completed, only a small number of rounds of local model iterative training are performed using local data.

[0242] S718, each edge device predicts the storage performance index of the local device based on the local performance prediction model.

[0243] It is understandable that data storage is the core of modern information services, and performance indicator prediction is crucial to ensure the stable operation of storage systems. Cloud monitoring systems can monitor storage performance in real time and help predict future states, but they face challenges in large-scale data processing and user privacy protection. To address these issues, this implementation proposes a storage system performance indicator prediction method based on federated learning under a cloud-edge collaborative architecture. Specifically:

[0244] This implementation constructs a multi-input multi-output prediction model based on variable selection and temporal convolutional networks, achieving high-precision prediction of storage system performance indicators. The variable selection network can generate sparse probability distributions, effectively eliminate irrelevant features, and reduce their contribution to zero, thereby intelligently and efficiently integrating all features and significantly improving the interpretability of the model. In addition, by introducing a temporal convolutional network, the model is able to capture the long-term temporal dependency of the data and optimize the number of parameters. This innovative model construction not only significantly improves the accuracy of the prediction, but also makes an important contribution to enhancing the efficiency and reliability of the storage system.

[0245] Furthermore, this embodiment proposes a storage performance indicator prediction model training method based on differential privacy federated learning. This method makes full use of the local data sets scattered on the edge devices to perform local training of the model. During the training process, the present invention cleverly integrates the differential privacy technology, and injects an appropriate amount of noise into the gradient information to ensure that even if the attacker has access to the aggregated model update, the training data of any single device cannot be accurately inferred, thereby significantly improving the security of data privacy. After the local training is completed, the cloud server will aggregate the local training model parameters from different edge devices to build a more accurate and comprehensive global model. This method not only avoids the privacy and security risks that may be caused by large-scale data transmission, but also promotes the continuous improvement and enhancement of the global model through continuous iteration and optimization.

[0246] In addition, this implementation supports personalized training based on sufficient global training. Each participant can use local data for local fine-tuning to generate a personalized model that adapts to the characteristics of local data. The generation of this personalized model further improves the prediction accuracy and better meets the demand for high-precision prediction in practical applications.

[0247] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0248] According to another aspect of the embodiment of the present application, a data processing device in a memory for implementing the above performance indicator prediction method is also provided. Figure 8 As shown, the device comprises:

[0249] A first acquisition unit 802 is used to acquire, through a target device, a target performance prediction model sent by a server, wherein target model parameters of the target performance prediction model are determined by the server according to reference model parameters returned by a plurality of edge devices;

[0250] The second acquisition unit 804 is used to acquire the storage performance feature sequence of the target device within the performance detection period, wherein the storage performance feature sequence includes storage performance features respectively matched with a plurality of time nodes, the plurality of time nodes are time nodes within the performance detection period, the storage performance features include indicator performance sub-features respectively matched with a plurality of storage performance indicators, the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the target device at the time nodes matched therewith;

[0251] The feature processing unit 806 is used to sequentially input the plurality of storage performance features into the variable selection network in the target performance prediction model to obtain a set of index weights that match the plurality of storage performance features, and determine a plurality of target storage performance features according to the plurality of storage performance features and the respective matching sets of index weights, wherein the set of index weights includes index weights corresponding to the respective plurality of index performance sub-features;

[0252] The prediction unit 808 is used to input the target performance feature sequence composed of multiple target storage performance features into the time convolution network in the target performance prediction model to obtain the performance indicator prediction result at the target moment, wherein the time convolution network is used to determine the time dependency between at least two of the multiple target storage performance features.

[0253] Optionally, the feature processing unit 806 comprises: an acquisition module for acquiring a current storage performance feature from a plurality of the above-mentioned storage performance features, wherein the current storage performance feature is a current feature matrix composed of a plurality of feature vectors, and the plurality of feature vectors are respectively used to characterize one of the above-mentioned indicator performance sub-features; an input module for inputting the current storage performance feature into the indicator weight module in the above-mentioned variable selection network to obtain a current weight vector matching the current storage performance feature, wherein the current weight vector includes the indicator weights respectively matching the plurality of the above-mentioned feature vectors; and a calculation module for determining the product result between the current weight vector and the current feature matrix as the target storage performance feature matching the current storage performance feature.

[0254] Optionally, the above-mentioned input module is used to: process the above-mentioned current storage performance characteristics through a feature conversion function to obtain a reference storage performance characteristic, wherein the above-mentioned feature conversion function includes one of the following: a linear transformation function, a multivariate perceptron function; process the above-mentioned reference storage performance characteristics through a sparse normalization function to obtain the above-mentioned current weight vector.

[0255] Optionally, the input module is used to: obtain the product result between the Lagrange multiplier and the first reference vector to obtain a second reference vector, wherein the vector elements included in the first reference vector are all 1s; determine the vector difference between the reference storage performance characteristic and the second reference vector as the first intermediate storage performance characteristic; convert each vector element in the intermediate storage performance characteristic into the corresponding reciprocal square root to obtain the second intermediate storage performance characteristic; update the values ​​of the vector elements in the second intermediate storage performance characteristic whose values ​​are less than 0 to 0 to obtain the current weight vector.

[0256] Optionally, the prediction unit 808 is used to: input the current performance feature sequence into the current feature processing module in the temporal convolutional network, wherein the current performance feature sequence is obtained by processing the target performance feature sequence via at least one of the feature processing modules, and the temporal convolutional network includes a plurality of stacked feature processing modules, wherein the feature processing module includes: two dilated causal convolution submodules, two layer regularization submodules, and an activation function submodule located between the two dilated causal convolution submodules; obtain the current output result of the current feature processing module based on the current performance feature sequence; when the current feature processing module is connected to the next feature processing module, input the current output result into the next feature processing module; when the current feature processing module is not connected to the next feature processing module, process the current output result according to the linear transformation layer to obtain the performance indicator prediction result at the target moment.

[0257] Optionally, the prediction unit 808 is used to: obtain a convolution kernel and a dilation factor that match the current feature processing module; perform a convolution operation based on the convolution kernel, the current performance feature sequence, and the dilation factor, wherein the dilation factor is used to skip the input value corresponding to the current moment, and apply the operation area corresponding to the convolution operation to the target feature area, and the step range corresponding to the target feature area is larger than the step range corresponding to the convolution kernel.

[0258] Optionally, the second acquisition unit 804 comprises: a feature extraction module for acquiring the multiple time nodes within the performance detection cycle, and multiple indicator parameters respectively matching the multiple time nodes, wherein the multiple indicator parameters respectively correspond to one storage performance indicator, and the multiple storage performance indicators include at least one of the following: a read / write rate indicator, a bandwidth indicator, a latency indicator, a CPU utilization indicator, and a data throughput indicator, wherein the read / write rate indicator is used to indicate the number of read / write operations performed by the target device within a unit time; according to each of the time nodes, normalizing the multiple indicator parameters corresponding to the time nodes to obtain indicator characteristic values ​​respectively matching the multiple indicator parameters; determining the storage performance characteristics matching the time nodes according to the multiple indicator characteristic values; and determining the storage performance characteristics corresponding to the multiple time nodes as the storage performance characteristic sequence.

[0259] Optionally, the feature extraction module is used to: obtain a current time parameter from a plurality of time parameters corresponding to the time node, wherein the plurality of time parameters are used to describe the time node; convert the current time parameter into a current time feature value according to a time period corresponding to the current time parameter; and determine a time feature vector indicating the time node according to the current time feature values ​​corresponding to each of the plurality of time parameters.

[0260] Optionally, the feature extraction module is used to: obtain the vector dimension of the time feature vector used to characterize the time node; convert the multiple indicator feature values ​​into indicator feature vectors matching the vector dimension, wherein the indicator feature vectors are used to indicate the indicator performance sub-features; and determine the feature matrix composed of the multiple indicator feature vectors as the storage performance feature matching the time node.

[0261] Optionally, the prediction device for the above-mentioned performance indicators also includes: a training unit, used to obtain a first reference prediction model in a training stage from the above-mentioned server, wherein the above-mentioned first reference prediction model includes a reference variable selection network in a training stage and a reference time convolutional network in a training stage, and the above-mentioned first reference prediction model is described according to a first model parameter set; obtaining a training sample library determined according to the historical operating status of the above-mentioned target device, wherein the above-mentioned training sample library includes training samples respectively matched with multiple historical time nodes, and the above-mentioned training samples include a historical storage performance feature sequence matching the above-mentioned historical time node, and a historical label storage performance feature matching the above-mentioned historical time node; training the above-mentioned first reference prediction model according to the above-mentioned training sample library to obtain a second reference prediction model that meets local training conditions; and sending a second model parameter set used to indicate the above-mentioned second reference prediction model to the above-mentioned server.

[0262] Optionally, the training unit is used to: in the current training round, obtain a reference training loss gradient set corresponding to a training sample subset matching the current training round, wherein the training sample library includes multiple training sample subsets, and the reference training loss gradient set includes first training loss gradients corresponding to multiple training samples in the training sample subset; perform gradient clipping on the first training loss gradients in the reference training loss gradient set to obtain second training loss gradients corresponding to the first training loss gradients; determine a third training loss gradient based on reference noise information and multiple second training loss gradients; and adjust the first model parameter set of the first reference prediction model based on the third training loss gradient to obtain the updated first model parameter set corresponding to the current training round.

[0263] Optionally, the training unit is used to: obtain a first ratio between the second-order norm corresponding to the first training loss gradient and a target constant; when the first ratio is greater than or equal to 1, determine a second ratio between the first training loss gradient and the first ratio as the second training loss gradient; when the first ratio is less than 1, determine the first training loss gradient as the second training loss gradient.

[0264] Optionally, the above-mentioned training unit is used to: obtain the above-mentioned reference noise information determined according to the normally distributed random variable; determine the weighted summation result between multiple second training loss gradients as the third reference noise gradient; determine the weighted summation result between the above-mentioned third reference noise gradient and the above-mentioned reference noise information as the above-mentioned third training loss gradient.

[0265] Optionally, the training unit is used to: determine the product of the target learning rate and the third training loss gradient as the parameter adjustment degree corresponding to the current training round; adjust the model parameters in the first model parameter set according to the parameter adjustment degree to obtain the updated first model parameter set.

[0266] Optionally, the training unit is used to: when the third reference prediction model determined according to the second reference prediction model meets the target training conditions, obtain the target performance prediction model sent by the server, wherein the model parameters of the third reference prediction model are obtained by the server after processing the reference model parameters sent by multiple edge devices based on the federal averaging algorithm, and the ratio of the first number of edge devices sending the reference model parameters to the second number of devices in the edge device set is a target value, and the edge device set includes all edge devices connected to the server; when the third reference prediction model determined according to the second reference prediction model does not meet the target training conditions, obtain the third reference prediction model sent by the server; and train the third prediction model according to the training sample library.

[0267] Optionally, the above-mentioned training unit is used to: obtain first encryption parameter information from the target encryption channel; decrypt the above-mentioned first encryption parameter information to obtain the above-mentioned first model parameter set used to indicate the above-mentioned first reference prediction model; encrypt the above-mentioned second model parameter set to obtain second encryption parameter information; and send the above-mentioned second encryption parameter information to the above-mentioned server through the above-mentioned target encryption channel.

[0268] According to another aspect of the embodiment of the present application, an electronic device for implementing the above-mentioned method for processing data in a memory is also provided. The electronic device may be Figure 1 The terminal device or server shown in the figure. This embodiment is described by taking the electronic device as a mobile phone or a computer as an example. Fig. 9 As shown, the electronic device includes a memory 902 and a processor 904. The memory 902 stores a computer program, and the processor 904 is configured to execute the steps in any of the above method embodiments through the computer program.

[0269] Optionally, in this embodiment, the electronic device may be located in at least one network device among a plurality of network devices of a computer network.

[0270] Optionally, in this embodiment, the processor may be configured to perform the following steps through a computer program:

[0271] S1, obtaining a target performance prediction model sent by a server through a target device, wherein target model parameters of the target performance prediction model are determined by the server according to reference model parameters returned by multiple edge devices;

[0272] S2, obtaining a storage performance feature sequence of the target device within a performance detection period, wherein the storage performance feature sequence includes storage performance features that are matched with a plurality of time nodes respectively, the plurality of time nodes are time nodes within the performance detection period, the storage performance features include indicator performance sub-features that are matched with a plurality of storage performance indicators respectively, the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the target device at the time nodes that match them;

[0273] S3, sequentially inputting the plurality of storage performance characteristics into the variable selection network in the target performance prediction model, obtaining a set of indicator weights that match the plurality of storage performance characteristics, and determining a plurality of target storage performance characteristics according to the plurality of storage performance characteristics and the respective matching sets of indicator weights, wherein the set of indicator weights includes indicator weights corresponding to the plurality of indicator performance sub-characteristics;

[0274] S4, inputting a target performance feature sequence composed of a plurality of the above target storage performance features into the temporal convolutional network in the above target performance prediction model to obtain a performance indicator prediction result at the target moment, wherein the above temporal convolutional network is used to determine the time dependency between at least two of the above target storage performance features among the plurality of the above target storage performance features.

[0275] Alternatively, a person skilled in the art may understand that: Fig. 9 The structure shown is for illustration only, and the electronic device may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile Internet device (Mobile Internet Devices, MID), a PAD, or other terminal devices. Fig.10 The structure of the electronic device is not limited. Fig.10 More or fewer components (such as network interfaces, etc.) as shown in, or with Fig.10 Different configurations shown.

[0276] Among them, the memory 902 can be used to store software programs and modules, such as the program instructions / modules corresponding to the performance indicator prediction method and device in the embodiment of the present application. The processor 904 executes various functional applications and data processing by running the software programs and modules stored in the memory 902, that is, realizing the above-mentioned performance indicator prediction method. The memory 902 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 902 may further include a memory remotely arranged relative to the processor 904, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Among them, the memory 902 can be specifically, but not limited to, used to store information such as page elements and page styles. As an example, such as Fig. 9 As shown, the memory 902 may include, but is not limited to, the first acquisition unit 802, the second acquisition unit 804, the feature processing unit 806, and the prediction unit 808 in the prediction device for the performance index. In addition, other module units in the prediction device for the performance index may also be included but are not limited to, which will not be described in detail in this example.

[0277] Optionally, the transmission device 906 is used to receive or send data via a network. Specific examples of the network may include a wired network and a wireless network. In one example, the transmission device 906 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices and routers via a network cable so as to communicate with the Internet or a local area network. In one example, the transmission device 906 is a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0278] In addition, the electronic device further includes: a display 908 for displaying the target page; and a connection bus 910 for connecting various module components in the electronic device.

[0279] In other embodiments, the terminal device or server may be a node in a distributed system, wherein the distributed system may be a blockchain system, and the blockchain system may be a distributed system formed by connecting the multiple nodes through network communication. The nodes may form a point-to-point network, and any form of computing device, such as a server, terminal or other electronic device, may become a node in the blockchain system by joining the point-to-point network.

[0280] According to another aspect of the embodiment of the present application, a training device for a performance indicator prediction model for implementing the training method of the performance indicator prediction model is also provided. Fig.10 As shown, the device comprises:

[0281] A construction unit 1002 is used to construct a first reference prediction model in a training stage, wherein the first reference prediction model includes a reference variable selection network in a training stage and a reference time convolution network in a training stage, the first reference prediction model is described according to a first model parameter set, the variable selection network is used to determine the indicator weights corresponding to the indicator performance sub-features that match the multiple storage performance indicators in the storage performance feature, and the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the edge device;

[0282] A first determining unit 1004 is configured to send the first reference prediction model to the target device when at least one target device for training the first reference prediction model is determined from the plurality of edge devices, wherein the first reference prediction model is described according to the first model parameter set;

[0283] An acquiring unit 1006 is used to acquire a second model parameter set returned by the plurality of target devices, wherein the second model parameter set is used to describe a second reference prediction model, and the second reference prediction model is obtained by training the first reference prediction model by the target device based on a training sample library;

[0284] The second determining unit 1008 is used to determine a target model parameter set for describing a target performance prediction model according to at least one of the second model parameter sets.

[0285] According to another aspect of the embodiment of the present application, there is also provided an electronic device for implementing the above-mentioned performance indicator prediction method or performance indicator prediction model training method, which may be Figure 1 The terminal device or server shown in the figure. This embodiment is described by taking the electronic device as a mobile phone or a computer as an example. Fig.11 As shown, the electronic device includes a memory 1102 and a processor 1104. The memory 1102 stores a computer program, and the processor 1104 is configured to execute the steps in any of the above method embodiments through the computer program.

[0286] Optionally, in this embodiment, the electronic device may be located in at least one network device among a plurality of network devices of a computer network.

[0287] Optionally, in this embodiment, the processor may be configured to perform the following steps through a computer program:

[0288] S1, constructing a first reference prediction model in a training stage, wherein the first reference prediction model includes a reference variable selection network in a training stage and a reference time convolution network in a training stage, the first reference prediction model is described according to a first model parameter set, the variable selection network is used to determine the indicator weights corresponding to the indicator performance sub-features respectively matching the multiple storage performance indicators in the storage performance feature, and the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the edge device;

[0289] S2, when at least one target device for training the first reference prediction model is determined from the plurality of edge devices, sending the first reference prediction model to the target device, wherein the first reference prediction model is described according to the first model parameter set;

[0290] S3, obtaining a second model parameter set returned by the plurality of target devices, wherein the second model parameter set is used to describe a second reference prediction model, and the second reference prediction model is obtained by training the first reference prediction model by the target device based on a training sample library;

[0291] S4: Determine a target model parameter set for describing a target performance prediction model based on at least one of the second model parameter sets.

[0292] Alternatively, a person skilled in the art may understand that: Fig.11 The structure shown is for illustration only, and the electronic device may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile Internet device (Mobile Internet Devices, MID), a PAD, or other terminal devices. Fig.10 The structure of the electronic device is not limited. Fig.10 More or fewer components (such as network interfaces, etc.) as shown in, or with Fig.10 Different configurations shown.

[0293] Among them, the memory 1102 can be used to store software programs and modules, such as the program instructions / modules corresponding to the performance indicator prediction method and device in the embodiment of the present application. The processor 1104 executes various functional applications and data processing by running the software programs and modules stored in the memory 1102, that is, realizing the above-mentioned performance indicator prediction method. The memory 1102 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1102 may further include a memory remotely arranged relative to the processor 1104, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Among them, the memory 1102 can be specifically, but not limited to, used to store information such as page elements and page styles. As an example, such as Fig.11 As shown, the memory 1102 may include, but is not limited to, the construction unit 1002, the first determination unit 1004, the acquisition unit 1006, and the second determination unit 1008 in the training device for the performance indicator prediction model. In addition, other module units in the training device for the performance indicator prediction model may also be included but are not limited to, which will not be repeated in this example.

[0294] Optionally, the transmission device 1106 is used to receive or send data via a network. Specific examples of the network may include a wired network and a wireless network. In one example, the transmission device 1106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices and routers via a network cable so as to communicate with the Internet or a local area network. In one example, the transmission device 1106 is a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0295] In addition, the electronic device further includes: a display 1108 for displaying the target page; and a connection bus 1110 for connecting various module components in the electronic device.

[0296] In other embodiments, the terminal device or server may be a node in a distributed system, wherein the distributed system may be a blockchain system, and the blockchain system may be a distributed system formed by connecting the multiple nodes through network communication. The nodes may form a point-to-point network, and any form of computing device, such as a server, terminal or other electronic device, may become a node in the blockchain system by joining the point-to-point network.

[0297] According to one aspect of the present application, a computer-readable storage medium is provided, and a processor of a computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in the above various optional implementations;

[0298] Optionally, in this embodiment, the computer-readable storage medium may be configured to store a computer program for performing the following steps:

[0299] S1, obtaining a target performance prediction model sent by a server through a target device, wherein target model parameters of the target performance prediction model are determined by the server according to reference model parameters returned by multiple edge devices;

[0300] S2, obtaining a storage performance feature sequence of the target device within a performance detection period, wherein the storage performance feature sequence includes storage performance features that are matched with a plurality of time nodes respectively, the plurality of time nodes are time nodes within the performance detection period, the storage performance features include indicator performance sub-features that are matched with a plurality of storage performance indicators respectively, the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the target device at the time nodes that match them;

[0301] S3, sequentially inputting the plurality of storage performance characteristics into the variable selection network in the target performance prediction model, obtaining a set of indicator weights that match the plurality of storage performance characteristics, and determining a plurality of target storage performance characteristics according to the plurality of storage performance characteristics and the respective matching sets of indicator weights, wherein the set of indicator weights includes indicator weights corresponding to the plurality of indicator performance sub-characteristics;

[0302] S4, inputting a target performance feature sequence composed of a plurality of the above target storage performance features into the temporal convolutional network in the above target performance prediction model to obtain a performance indicator prediction result at the target moment, wherein the above temporal convolutional network is used to determine the time dependency between at least two of the above target storage performance features among the plurality of the above target storage performance features.

[0303] Optionally, in this embodiment, the computer-readable storage medium may also be configured to store a computer program for performing the following steps:

[0304] S1, constructing a first reference prediction model in a training stage, wherein the first reference prediction model includes a reference variable selection network in a training stage and a reference time convolution network in a training stage, the first reference prediction model is described according to a first model parameter set, the variable selection network is used to determine the indicator weights corresponding to the indicator performance sub-features respectively matching the multiple storage performance indicators in the storage performance feature, and the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the edge device;

[0305] S2, when at least one target device for training the first reference prediction model is determined from the plurality of edge devices, sending the first reference prediction model to the target device, wherein the first reference prediction model is described according to the first model parameter set;

[0306] S3, obtaining a second model parameter set returned by the plurality of target devices, wherein the second model parameter set is used to describe a second reference prediction model, and the second reference prediction model is obtained by training the first reference prediction model by the target device based on a training sample library;

[0307] S4: Determine a target model parameter set for describing a target performance prediction model based on at least one of the second model parameter sets.

[0308] Optionally, in the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories), or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0309] Optionally, in this embodiment, a person of ordinary skill in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by instructing hardware related to the terminal device through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0310] If the integrated units in the above embodiments are implemented in the form of software functional units and sold or used as independent products, they can be stored in the above computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling one or more computer devices (which can be personal computers, servers or network devices, etc.) to execute all or part of the steps of the methods of each embodiment of the present application.

[0311] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0312] In the several embodiments provided in the present application, it should be understood that the disclosed client can be implemented in other ways. Among them, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units or components 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 interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0313] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0314] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0315] The above are only preferred implementations of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

[0316] It should be noted that the above modules can be implemented by software or hardware. For the latter, it can be implemented in the following ways, but not limited to: the above modules are all located in the same processor; or the above modules are located in different processors in any combination.

[0317] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when running.

[0318] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0319] An embodiment of the present application further provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0320] In an exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0321] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any of the above method embodiments are implemented.

[0322] An embodiment of the present application further provides another computer program product, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in any of the above method embodiments are implemented.

[0323] The embodiments of the present application also provide a computer program, which includes computer instructions, which are stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in any one of the above method embodiments.

[0324] For specific examples in this embodiment, reference may be made to the examples described in the above embodiments and exemplary implementation modes, and this embodiment will not be described in detail herein.

[0325] Obviously, those skilled in the art should understand that the above modules or steps of the present application can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present application is not limited to any specific combination of hardware and software.

[0326] The above are only preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting a performance index, characterized in that: include: Acquire, through the target device, a target performance prediction model sent by the server, wherein a target model parameter set of the target performance prediction model is determined by the server according to a reference model parameter set returned by multiple edge devices; Acquire a storage performance feature sequence of the target device within a performance detection period, wherein the storage performance feature sequence includes storage performance features that are matched by a plurality of time nodes respectively, the plurality of time nodes are time nodes within the performance detection period, the storage performance features include indicator performance sub-features that are matched by a plurality of storage performance indicators respectively, the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the target device at the time node that matches the storage performance indicator; Inputting the plurality of storage performance characteristics into the variable selection network in the target performance prediction model in sequence, obtaining a set of indicator weights that match each of the plurality of storage performance characteristics, and determining a plurality of target storage performance characteristics according to the plurality of storage performance characteristics and the respective matching indicator weight sets, wherein the indicator weight set includes indicator weights corresponding to each of the plurality of indicator performance sub-characteristics; A target performance feature sequence composed of multiple target storage performance features is input into the temporal convolutional network in the target performance prediction model to obtain a performance indicator prediction result at the target moment, wherein the temporal convolutional network is used to determine the time dependency between at least two of the multiple target storage performance features.

2. The method according to claim 1, characterized in that The step of sequentially inputting the plurality of storage performance characteristics into the variable selection network in the target performance prediction model to obtain a set of index weights that match each of the plurality of storage performance characteristics, and determining a plurality of target storage performance characteristics according to the plurality of storage performance characteristics and the respective matching index weight sets comprises: Acquire a current storage performance feature from the plurality of storage performance features, wherein the current storage performance feature is a current feature matrix composed of a plurality of feature vectors, and the plurality of feature vectors are respectively used to characterize one of the indicator performance sub-features; Inputting the current storage performance characteristics into the indicator weight module in the variable selection network to obtain a current weight vector matching the current storage performance characteristics, wherein the current weight vector includes the indicator weights respectively matching the plurality of feature vectors; The product result between the current weight vector and the current feature matrix is ​​determined as the target storage performance feature that matches the current storage performance feature.

3. The method according to claim 2, characterized in that The step of inputting the current storage performance characteristics into the indicator weight module in the variable selection network to obtain a current weight vector matching the current storage performance characteristics includes: Processing the current storage performance feature through a feature conversion function to obtain a reference storage performance feature, wherein the feature conversion function includes one of the following: a linear transformation function, a multivariate perceptron function; The reference storage performance characteristic is processed by a sparse normalization function to obtain the current weight vector.

4. The method according to claim 3, characterized in that The step of processing the reference storage performance characteristic by a sparse normalization function to obtain the current weight vector includes: Obtaining a product result between a Lagrange multiplier and a first reference vector to obtain a second reference vector, wherein all vector elements included in the first reference vector are all 1s; determining a vector difference between the reference storage performance characteristic and the second reference vector as a first intermediate storage performance characteristic; Convert each vector element in the first intermediate storage performance characteristic into a corresponding reciprocal square root to obtain a second intermediate storage performance characteristic; The values ​​of the vector elements whose values ​​in the second intermediate storage performance characteristic are less than 0 are updated to 0 to obtain the current weight vector.

5. The method according to claim 2, characterized in that: The target performance feature sequence composed of a plurality of the target storage performance features is input into the temporal convolutional network in the target performance prediction model to obtain the performance indicator prediction result at the target time, including: Inputting the current performance feature sequence into the current feature processing module in the temporal convolutional network, wherein the current performance feature sequence is obtained by processing the target performance feature sequence via at least one feature processing module, the temporal convolutional network includes a plurality of stacked feature processing modules, and the feature processing module includes: two dilated causal convolution submodules, two layer regularization submodules, and an activation function submodule located between the two dilated causal convolution submodules; Obtaining a current output result of the current feature processing module based on the current performance feature sequence; In the case where the current feature processing module is connected to the next feature processing module, inputting the current output result into the next feature processing module; In the case where the current feature processing module is not connected to the next feature processing module, the current output result is processed according to the linear transformation layer to obtain the performance indicator prediction result at the target moment.

6. The method according to claim 5, characterized in that After the current performance feature sequence is input into the current feature processing module in the temporal convolutional network, the method further includes: Obtaining a convolution kernel and a dilation factor that match the current feature processing module; A convolution operation is performed based on the convolution kernel, the current performance feature sequence, and the dilation factor, wherein the dilation factor is used to skip the input value corresponding to the current moment, and to apply the operation area corresponding to the convolution operation to the target feature area, and the step range corresponding to the target feature area is larger than the step range corresponding to the convolution kernel.

7. The method according to claim 2, characterized in that The acquiring of the storage performance characteristic sequence of the target device within the performance detection cycle includes: Acquire a plurality of the time nodes within the performance detection cycle, and a plurality of indicator parameters respectively matched with the plurality of the time nodes, wherein the plurality of indicator parameters respectively correspond to one of the storage performance indicators, and the plurality of storage performance indicators include at least one of the following: a read / write rate indicator, a bandwidth indicator, a latency indicator, a CPU utilization indicator, and a data throughput indicator, wherein the read / write rate indicator is used to indicate the number of read / write operations performed by the target device within a unit time; According to each of the time nodes, normalizing the multiple indicator parameters corresponding to the time node to obtain indicator characteristic values ​​respectively matching the multiple indicator parameters; Determine the storage performance feature matching the time node according to the plurality of indicator feature values; The storage performance characteristics corresponding to each of the multiple time nodes are determined as the storage performance characteristic sequence.

8. The method according to claim 7, characterized in that After acquiring the multiple time nodes within the performance detection period and the multiple indicator parameters respectively matching the multiple time nodes, the method further includes: Obtaining a current time parameter from among a plurality of time parameters corresponding to the time node, wherein the plurality of time parameters are used to describe the time node; According to the time period corresponding to the current time parameter, converting the current time parameter into a current time characteristic value; A time feature vector indicating the time node is determined according to the current time feature values ​​corresponding to each of the multiple time parameters.

9. The method according to claim 7, characterized in that: The determining the storage performance feature matching the time node according to the plurality of indicator feature values ​​includes: Obtaining a vector dimension of a time feature vector used to characterize the time node; Converting the plurality of indicator feature values ​​into indicator feature vectors matching the vector dimension respectively, wherein the indicator feature vectors are used to indicate the indicator performance sub-features; A feature matrix composed of a plurality of the indicator feature vectors is determined as the storage performance feature matching the time node.

10. The method according to claim 1, characterized in that Before obtaining the target performance prediction model sent by the server through the target device, the method further includes: Acquire a first reference prediction model in a training phase from the server, wherein the first reference prediction model includes a reference variable selection network in a training phase and a reference time convolution network in a training phase, and the first reference prediction model is described according to a first model parameter set; Acquire a training sample library determined according to the historical operating state of the target device, wherein the training sample library includes training samples that match multiple historical time nodes respectively, and the training samples include historical storage performance feature sequences that match the historical time nodes, and historical label storage performance features that match the historical time nodes; Training the first reference prediction model according to the training sample library to obtain a second reference prediction model that meets local training conditions; A second model parameter set indicating the second reference prediction model is sent to the server.

11. The method according to claim 10, characterized in that The step of training the first reference prediction model according to the training sample library to obtain a second reference prediction model that meets the local training condition includes: In a current training round, a reference training loss gradient set corresponding to a training sample subset matching the current training round is obtained, wherein the training sample library includes a plurality of the training sample subsets, and the reference training loss gradient set includes first training loss gradients respectively corresponding to a plurality of the training samples in the training sample subsets; Performing gradient clipping on the first training loss gradients in the reference training loss gradient set to obtain second training loss gradients corresponding to the first training loss gradients; Determine a third training loss gradient according to the reference noise information and a plurality of the second training loss gradients; According to the third training loss gradient, the first model parameter set of the first reference prediction model is adjusted to obtain an updated first model parameter set corresponding to the current training round.

12. The method according to claim 11, characterized in that The performing gradient clipping on the first training loss gradients in the reference training loss gradient set to obtain second training loss gradients corresponding to the first training loss gradients respectively includes: Obtaining a first ratio between a second-order norm corresponding to the first training loss gradient and a target constant; When the first ratio is greater than or equal to 1, determining a second ratio between the first training loss gradient and the first ratio as the second training loss gradient; When the first ratio is less than 1, the first training loss gradient is determined as the second training loss gradient.

13. The method according to claim 11, characterized in that The determining the third training loss gradient according to the reference noise information and the plurality of the second training loss gradients comprises: Acquiring the reference noise information determined according to a normally distributed random variable; Determine a weighted sum of the multiple second training loss gradients as a third reference noise gradient; A weighted sum of the third reference noise gradient and the reference noise information is determined as the third training loss gradient.

14. The method according to claim 11, characterized in that The step of adjusting the first model parameter set of the first reference prediction model according to the third training loss gradient to obtain an updated first model parameter set corresponding to the current training round includes: Determine the product of the target learning rate and the third training loss gradient as the parameter adjustment degree corresponding to the current training round; The model parameters in the first model parameter set are adjusted according to the parameter adjustment degree to obtain an updated first model parameter set.

15. The method according to claim 10, characterized in that After sending the second model parameter set for indicating the second reference prediction model to the server, the method further includes: When the third reference prediction model determined according to the second reference prediction model meets the target training condition, the target performance prediction model sent by the server is obtained, wherein the model parameters of the third reference prediction model are obtained after the server processes the reference model parameters sent by the plurality of edge devices based on the federated average algorithm, and the ratio of the first number of edge devices sending the reference model parameters to the second number of devices in the edge device set is a target value, and the edge device set includes all edge devices connected to the server; When the third reference prediction model determined according to the second reference prediction model does not meet the target training condition, obtaining the third reference prediction model sent by the server; The third reference prediction model is trained according to the training sample library.

16. The method according to claim 10, characterized in that The acquiring, from the server, a first reference prediction model in a training phase comprises: Acquire first encryption parameter information from a target encryption channel; decrypting the first encrypted parameter information to obtain the first model parameter set indicating the first reference prediction model; The sending a second model parameter set indicating the second reference prediction model to the server includes: Encrypting the second model parameter set to obtain second encrypted parameter information; The second encryption parameter information is sent to the server through the target encryption channel.

17. A training method for a performance indicator prediction model, characterized in that: Constructing a first reference prediction model in a training phase, wherein the first reference prediction model includes a reference variable selection network in a training phase and a reference time convolution network in a training phase, the first reference prediction model is described according to a first model parameter set, the variable selection network is used to determine the indicator weights corresponding to the indicator performance sub-features respectively matched by the multiple storage performance indicators in the storage performance feature, and the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the edge device; In a case where at least one target device for training the first reference prediction model is determined from a plurality of edge devices, sending the first reference prediction model to the target device, wherein the first reference prediction model is described according to a first model parameter set; Acquire a second model parameter set returned by the plurality of target devices, wherein the second model parameter set is used to describe a second reference prediction model, and the second reference prediction model is obtained by the target device training the first reference prediction model based on a training sample library; According to at least one of the second model parameter sets, a target model parameter set for describing a target performance prediction model is determined.

18. A performance index prediction device, characterized in that: include: A first acquisition unit is used to acquire, through a target device, a target performance prediction model sent by a server, wherein a target model parameter set of the target performance prediction model is determined by the server according to a reference model parameter set returned by a plurality of edge devices; A second acquisition unit is used to acquire a storage performance feature sequence of the target device within a performance detection period, wherein the storage performance feature sequence includes storage performance features that are matched with a plurality of time nodes respectively, the plurality of time nodes are time nodes within the performance detection period, the storage performance features include indicator performance sub-features that are matched with a plurality of storage performance indicators respectively, the indicator performance sub-features are used to indicate the performance of the storage performance indicator of the target device at the time node that matches therewith; A feature processing unit, used to sequentially input the plurality of storage performance features into the variable selection network in the target performance prediction model, obtain an indicator weight set that matches each of the plurality of storage performance features, and determine a plurality of target storage performance features according to the plurality of storage performance features and the respective matching indicator weight sets, wherein the indicator weight set includes indicator weights corresponding to each of the plurality of indicator performance sub-features; A prediction unit is used to input a target performance feature sequence composed of multiple target storage performance features into a temporal convolutional network in the target performance prediction model to obtain a performance indicator prediction result at a target moment, wherein the temporal convolutional network is used to determine a time dependency relationship between at least two of the multiple target storage performance features.

19. A training device for a performance indicator prediction model, characterized in that: A construction unit, configured to construct a first reference prediction model in a training phase, wherein the first reference prediction model includes a reference variable selection network in a training phase and a reference time convolution network in a training phase, the first reference prediction model is described according to a first model parameter set, the variable selection network is configured to determine the indicator weights corresponding to the indicator performance sub-features respectively matched by the multiple storage performance indicators in the storage performance feature, and the indicator performance sub-features are configured to indicate the performance of the storage performance indicator of the edge device; A first determining unit is configured to send the first reference prediction model to the target device when at least one target device for training the first reference prediction model is determined from a plurality of edge devices, wherein the first reference prediction model is described according to a first model parameter set; an acquiring unit, configured to acquire a second model parameter set returned by the plurality of target devices, wherein the second model parameter set is used to describe a second reference prediction model, and the second reference prediction model is obtained by training the first reference prediction model by the target device based on a training sample library; The second determining unit determines a target model parameter set for describing a target performance prediction model according to at least one of the second model parameter sets.

20. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program implements the steps of the method described in any one of claims 1 to 16 or 17 when executed by a processor.

21. An electronic 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 steps of the method described in any one of claims 1 to 16 or 17 are implemented.

22. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in any one of claims 1 to 16 or 17 are implemented.

Images