Network computing power resource management method and device

Abstract

The application discloses a network computing resource management method and device. The method comprises the following steps: obtaining a first traffic sequence and a first computing sequence of each of a plurality of first links in a first distributed computing network in a first time period; constructing a first space-time graph by taking the plurality of first links as nodes, and determining a first adjacency matrix of the first space-time graph according to the first traffic sequence of each first link; analyzing each first traffic sequence, each first computing sequence and the first adjacency matrix by using a target traffic prediction model to obtain a first traffic prediction sequence of each first link in a second time period; and managing the computing resources of the first distributed computing network according to the first traffic prediction sequence of each first link in the second time period. The application solves the technical problem that, in the related art, the accuracy of the network traffic prediction result is low, so that the computing resources cannot be reasonably allocated and scheduled.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network management, in particular to a network computing resource management method and device. BACKGROUND

[0002] In modern communication networks, with the rapid growth of user numbers and data demands, network traffic prediction becomes particularly important. Accurate traffic prediction not only helps to effectively allocate network resources, but also improves the overall performance of the network, reduces latency and congestion. However, traditional traffic prediction methods only predict based on historical traffic data, such as the computing power usage of each node in the network. However, as the network scale expands, the computing power differences between nodes in the network become more and more obvious. These differences may be due to the performance differences of hardware devices, the complexity of network topology, and the changes in the load of surrounding nodes, etc. Therefore, traditional network traffic prediction methods do not consider other information in distributed networks, which may lead to a decrease in traffic prediction accuracy in high-dynamic and complex network environments. Further, it is unable to assist managers in identifying potential traffic bottlenecks in advance, and it is also unable to reasonably adjust the allocation and scheduling of computing power resources.

[0003] At present, there is no effective solution to the above problems. SUMMARY

[0004] The embodiments of the present application provide a network computing resource management method and device to at least solve the technical problem that the related art cannot reasonably allocate and schedule computing power resources due to the low accuracy of network traffic prediction results relied on when performing network computing resource management.

[0005] According to an aspect of an embodiment of the present application, a network computing resource management method is provided, comprising: obtaining a first traffic sequence and a first computing power sequence of each of a plurality of first links in a first distributed computing power network in a first time period; constructing a first space-time graph corresponding to the first distributed computing power network by taking the plurality of first links as nodes, and determining a first adjacency matrix corresponding to the first space-time graph according to the first traffic sequence of each first link; analyzing each first traffic sequence, each first computing power sequence and the first adjacency matrix by using a pre-trained target traffic prediction model to obtain a first traffic prediction sequence of each first link in a second time period, wherein the second time period is a time period after the first time period; and managing the computing power resources of the first distributed computing power network according to the first traffic prediction sequence of each first link in the second time period.

[0006] Optionally, the obtaining the first traffic sequence and the first computing power sequence of each of the plurality of first links in the first distributed computing power network in the first time period comprises: determining the plurality of first links in the first distributed computing power network; for each first link, collecting an initial traffic sequence and an initial computing power sequence of the first link in the first time period by using a preset protocol, wherein the initial traffic sequence comprises a traffic value of the first link at each time point in the first time period, and the traffic value is used to reflect a data amount successfully transmitted by the first link at the corresponding time point, and the initial computing power sequence comprises computing power information of the first link at each time point in the first time period, and the computing power information is determined by at least computing capability, communication capability, memory capability, and storage capability; and pre-processing the initial traffic sequence and the initial computing power sequence to obtain the corresponding first traffic sequence and the first computing power sequence.

[0007] Optionally, the determining the first adjacency matrix corresponding to the first space-time graph according to the first traffic sequence of each first link comprises: calculating a distance value between the first traffic sequences of each two first links in the plurality of first links, and determining a traffic similarity between each two first links according to the distance value, wherein the greater the distance value is, the smaller the corresponding traffic similarity is, and the smaller the distance value is, the greater the corresponding traffic similarity is; and taking each first link as a row and a column of a matrix, and taking an inverse of the traffic similarity between each two first links in the plurality of first links as a matrix element to obtain the first adjacency matrix corresponding to the first space-time graph.

[0008] Optionally, the training process of the target traffic prediction model comprises: obtaining a training sample set and a sample label set, wherein the training sample set comprises a training sample composed of the second adjacency matrix corresponding to the second space-time graph of the second distributed computing power network in the third time period, the second traffic sequence and the second computing power sequence of the second link in the second distributed computing power network in the third time period, and the sample label set comprises a sample label corresponding to the training sample, which is the second traffic sequence of the second link in the fourth time period, and the third time period is a time period before the fourth time period, and the third time period is a time period before the first time period; constructing an initial traffic prediction model, wherein the initial traffic prediction model comprises at least a plurality of space-time graph convolution modules, and each space-time graph convolution module comprises at least two gated convolution networks and one graph convolution network; and iteratively training the initial traffic prediction model by using the training sample set and the sample label set to obtain the target traffic prediction model.

[0009] Optionally, the initial traffic prediction model is iteratively trained by using the training sample set and the sample label set to obtain a target traffic prediction model, including: in each iteration training process, inputting the training sample in the training sample set into the initial traffic prediction model to obtain a second predicted traffic sequence output by the initial traffic prediction model; constructing a target loss function according to the second predicted traffic sequence output by the initial traffic prediction model and the second traffic sequence corresponding to each training sample; minimizing the target loss function by using a preset optimization algorithm until the model converges to obtain the target traffic prediction model.

[0010] Optionally, the computing resource of the first distributed computing power network is managed according to the first traffic prediction sequence of each first link in the second time period, including: for each first link, determining the demand computing power information of the first link at each time in the second time period according to the first traffic prediction sequence of the first link in the second time period; for the first time in the second time period, determining whether the demand computing power information of the first link at the first time in the second time period is greater than the first computing power information of the first link at the last time in the first time period; if yes, determining a first difference value between the demand computing power information of the first link at the first time in the second time period and the first computing power information of the first link at the last time in the first time period, and scheduling the computing resource corresponding to the first difference value from other idle links; if not, determining a second difference value between the first computing power information of the first link at the last time in the first time period and the demand computing power information of the first link at the first time in the second time period, and obtaining the released computing power information of the first link at the first time in the second time period, and taking the sum of the computing power information corresponding to the second difference value and the released computing power information as the first computing power information of the first link at the first time in the second time period, and repeating the above judgment process until all times in the second time period are traversed or the first computing power information of the first link is zero.

[0011] Optionally, the demand computing power information of the first link at each time in the second time period is determined according to the first traffic prediction sequence of the first link in the second time period, including: matching the first traffic prediction value at each time in the first traffic prediction sequence with a preset traffic-computing power mapping table to determine the demand computing power information of the first link at each time in the second time period, wherein the traffic-computing power mapping table records the mapping relationship between a plurality of traffic values and demand computing power information.

[0012] According to another aspect of the embodiments of the present application, a network computing resource management apparatus is also provided, comprising: an acquisition module configured to acquire a first traffic sequence and a first computing resource sequence of each of a plurality of first links in a first distributed computing network within a first time period; a determination module configured to construct a first space-time graph corresponding to the first distributed computing network by taking the plurality of first links as nodes, and determine a first adjacency matrix corresponding to the first space-time graph according to the first traffic sequence of each of the first links; a prediction module configured to analyze each of the first traffic sequence, each of the first computing resource sequence, and the first adjacency matrix by using a pre-trained target traffic prediction model, to obtain a first traffic prediction sequence of each of the first links within a second time period, wherein the second time period is a time period after the first time period; and a management module configured to manage computing resources of the first distributed computing network according to the first traffic prediction sequence of each of the first links within the second time period.

[0013] According to another aspect of the embodiments of the present application, a computer program product is also provided, comprising: a computer program, wherein the computer program is executed by a processor to implement the network computing resource management method described above.

[0014] According to another aspect of the embodiments of the present application, a non-volatile storage medium is also provided, which stores a computer program, wherein a device in which the non-volatile storage medium is located executes the network computing resource management method described above by running the computer program.

[0015] In the embodiments of the present application, by taking the first traffic sequence and the first computing resource sequence of each of the first links within the first time period in the first distributed computing network, and the first adjacency matrix of the first space-time graph corresponding to the first distributed computing network as inputs of the model, the model can capture the space-time characteristics of the link traffic and its association with the node computing resource state, and generate the first traffic prediction sequence of each of the first links within the second time period with high precision, and then dynamically adjust the allocation of the computing resources in the network according to the prediction sequences, effectively cope with the traffic peak, avoid resource waste, and ensure the efficient operation of the network. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which are included to provide a further understanding of the present application, constitute a part of the present application, illustrate the exemplary embodiments of the present application and specific examples, and serve to explain the present application, and do not constitute improper limitations on the present application. In the drawings:

[0017] Figure 1 is a flow schematic diagram of an optional network computing resource management method according to an embodiment of the present application;

[0018] Figure 2 is a schematic diagram of a topological structure of an optional distributed computing network according to an embodiment of the present application;

[0019] Figure 3 is a schematic diagram of a topological structure and a space-time diagram of an optional distributed computing power network according to an embodiment of the application;

[0020] Figure 4 is a structural schematic diagram of an optional initial traffic prediction model according to an embodiment of the application;

[0021] Figure 5 is a structural schematic diagram of an optional network computing power resource management device according to an embodiment of the application;

[0022] Figure 6 is a structural schematic diagram of an optional electronic device according to an embodiment of the application. DETAILED DESCRIPTION

[0023] In order for those skilled in the art to better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0024] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] In order to better understand the embodiments of the present application, first, the translation explanation of some nouns or terms appearing in the description of the embodiments of the present application is as follows:

[0026] Distributed computing power network is a new type of computing model, which changes the traditional centralized computing architecture into distributed computing architecture. By connecting various computing nodes, a huge computing network is formed, so as to not only improve computing efficiency and reduce computing cost, but also effectively cope with challenges such as large-scale data processing, high-concurrency computing tasks, complex model training, etc.

[0027] Dynamic Time Warping (DTW): A dynamic time warping algorithm is used to measure the similarity of time series of different lengths.

[0028] Gated Convolution Network (GCN): Gated convolution is a convolutional network technique in deep learning that controls the flow of information through convolution operations by introducing a gating mechanism. It applies two independent convolutions on the input features to generate feature maps and gating signals, and selectively allows important information to pass through by gating signals. Therefore, the calculation formula of gated convolution network can be represented as:

[0029] Γ* τ X = P ⊙ σ(Q)

[0030] Where Γ represents the convolution kernel, Y represents the input data of the gated convolution network, P and Q represent the input of the gated linear unit, and the Hadamard product of P and Q after passing through the Sigmoid function is the final convolution output. This mechanism enhances the selective processing ability of the model for important information.

[0031] Graph Convolutional Networks (GCN): A deep learning model specifically designed for processing graph-structured data, it combines the topological structure of the graph with the node features, and extracts the relationship and feature information between nodes through convolution operations. Therefore, the calculation formula of graph convolution network can be represented as:

[0032]

[0033] Where Y represents the convolution input, A represents the adjacency matrix of the graph, and θ k represents the parameters of the graph convolution module. Therefore, the graph convolution network can effectively capture the local dependency relationship and global structure characteristics between nodes in the graph.

[0034] RMSProp (Root Mean Square Propagation) optimization algorithm: An adaptive learning rate optimization algorithm, mainly used for parameter updating in deep learning, aims to solve the problem of Adagrad algorithm in deep learning training process, where the learning rate gradually decreases until it cannot be further learned. The core idea is to use different learning rates for each parameter, which are adjusted adaptively according to the size of the recent gradient of the parameter. Specifically, RMSProp uses an exponentially weighted moving average of squared gradients to adjust the learning rate, making the adjustment of the learning rate more smooth.

[0035] Stochastic Gradient Descent (SGD) algorithm: an optimization algorithm widely used in machine learning and deep learning, used to minimize the objective function, i.e. the loss function. Its core idea is to randomly select a sample or a small batch of samples to calculate the gradient in each iteration process, rather than calculating the gradient on the entire dataset. This method can significantly speed up the learning process and is more efficient when dealing with large-scale datasets.

[0036] Embodiment 1

[0037] According to the embodiments of the present application, a network computing resource management method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0038] Figure 1 is a flowchart of a network computing resource management method according to an embodiment of the present application, as Figure 1 shown, the method comprises the following steps:

[0039] Step S102, obtaining the first traffic sequence and the first computing power sequence of each first link in the first distributed computing power network in the first time period.

[0040] In the technical solution provided in the above step S102, the first link in the first distributed computing power network is a physical or logical path connecting two nodes in the network, and the network nodes can be client nodes, router nodes, computing power server nodes, switch nodes, etc. As Figure 2 shown, among them, the client node (such as Figure 2 Client1, Clinent2 in) can be understood as a terminal user equipment, used to send computing task request and receive computing result; the router node (such as Figure 2 Router1, Router2, Router3, Router4, Router5 in) is responsible for data packet forwarding and routing selection to ensure efficient data transmission in the network; the computing power server node (such as Figure 2 Server1, Server2 in) is the core computing resource provider, with powerful computing ability and storage resources, used to process complex computing tasks. Therefore, the first link in the embodiments of the present application can be the link Link1 between the router Router1 and the router Router3, the link Link8 between the router Router5 and the computing power server Server2, etc.

[0041] For each first link in the first distributed computing power network, the network devices (i.e., routers Router and computing power servers Server) at both ends of the first link are managed / monitored to collect a first traffic sequence of the first link in a first time period and a first computing power sequence of the first link in the first time period, wherein the first traffic sequence includes traffic values of the first link at each time in the first time period, and the first computing power sequence includes computing power resources of the first link at each time in the first time period.

[0042] In step S104, a first space-time graph corresponding to the first distributed computing power network is constructed with the plurality of first links as nodes, and a first adjacency matrix corresponding to the first space-time graph is determined according to the first traffic sequence of each first link.

[0043] In the technical solution provided in step S104, the plurality of first links in the first distributed computing power network are taken as independent graph nodes, and the connection relationship between each first link is taken as an edge to construct a first space-time graph corresponding to the first distributed computing power network. Then, a first adjacency matrix corresponding to the first space-time graph is determined according to the first traffic sequence of each first link, wherein each matrix element in the first adjacency matrix can reflect the connection weight between each graph node.

[0044] In step S106, the pre-trained target traffic prediction model is used to analyze each first traffic sequence, each first computing power sequence, and the first adjacency matrix to obtain a first traffic prediction sequence of each first link in a second time period.

[0045] In the technical solution provided in step S106, each first traffic sequence (time information), each first computing power sequence (resource state information), and the first adjacency matrix (spatial information) are jointly input into the target traffic prediction model to perform space-time traffic prediction on the traffic through the model to obtain a first traffic prediction sequence of each first link in a second time period. Wherein:

[0046] The first adjacency matrix reflects the connection structure of each node in the topology structure corresponding to the first distributed computing power network, i.e., the spatial dependency relationship between links, so that through the first adjacency matrix, the model can understand which nodes are directly connected, which nodes are indirectly connected, and the connection strength between nodes. This spatial dependency relationship is crucial for predicting link traffic, because the traffic in the network is often transmitted through multiple links, and the traffic of a link can be affected by the traffic of its directly or indirectly connected links.

[0047] The first traffic sequence of each first link is time series data, which reflects the trend and periodicity of the traffic value on the link over time. Therefore, inputting the first traffic sequence into the model enables the model to learn the change pattern of the traffic, including short-term fluctuations, long-term trends, etc. These dynamic characteristics over time are crucial for predicting future traffic, because traffic prediction is essentially a prediction on a time series. Therefore, by analyzing the first traffic sequence of each first link, the model can capture the periodicity and trend of the traffic, thereby making more accurate predictions.

[0048] The first computing power sequence of each first link contains the computing power information of each first link at each time in the first time period. In this embodiment, the maximum value of the computing power information in the nodes connected by the link is used as the computing power information on the link. This strategy is mainly considered that the efficiency and capacity of data transmission on the link are often limited by the weaker side of the computing power of the nodes at both ends of the link. Therefore, using the computing power information of the strongest node can represent the computing power state of the link, thereby assisting the traffic prediction model to more accurately predict the traffic of the link. These computing power information is also important for traffic prediction, especially when network traffic is closely related to the computing power usage of the computing power server. For example, when the computing power resource is tight, it may affect the data processing speed, thereby affecting the traffic of the link; on the contrary, when the computing power resource is sufficient, the data processing speed is accelerated, and the traffic value on the link may increase. Therefore, the computing power information provides another dimension for traffic prediction, so that the model can consider the resource state and traffic dynamics of the first link to improve the accuracy and robustness of the prediction.

[0049] Therefore, the prediction method of fusing information can more comprehensively understand the complex characteristics of the distributed computing power network traffic, thereby improving the prediction effect, especially in the scene where the network condition is variable and the resource allocation is uneven.

[0050] In step S108, the computing power resources of the first distributed computing power network are managed according to the first traffic prediction sequence of each first link in the second time period.

[0051] In the technical solution provided in the above step S108, by analyzing the first traffic prediction sequence of each first link in the second time period, and combining the computing power information of each first link, the dynamic scheduling and optimization of the computing power resources are performed, such as increasing the computing power support in the peak time of predicted traffic and reducing the waste of computing power in the low peak period, thereby realizing the efficient use of computing power resources.

[0052] Based on the scheme defined in steps S102 to S108, by taking the first traffic sequence and the first computing power sequence of each first link in the first distributed computing power network within the first time period, and the first adjacency matrix of the first space-time graph corresponding to the first distributed computing power network as the input of the model, the model captures the space-time characteristics of the link traffic and its association with the node computing power state, and generates the first traffic prediction sequence of each first link within the second time period with high precision, and then dynamically adjusts the allocation of computing power resources in the network according to the prediction sequences, effectively deals with traffic peaks, avoids resource waste, and ensures the efficient operation of the network.

[0053] The steps of the network computing power resource management method will be described below in combination with a specific implementation process.

[0054] As an optional implementation, in the technical scheme provided in step S102, the method can include:

[0055] Step S1021, determining a plurality of first links in the first distributed computing power network.

[0056] Specifically, by analyzing the topological structure of the first distributed computing power network, all nodes (such as servers, routers, switches, etc.) in the network and their connection modes are identified, so as to obtain each link (i.e. the communication channel connecting two network nodes).

[0057] Step S1022, for each first link, the first traffic sequence and the first computing power sequence of the corresponding computing power server can be obtained by the following method, including:

[0058] First, a preset protocol is used to collect the initial traffic sequence and the initial computing power sequence of the first link within the first time period.

[0059] The above-mentioned preset protocol is the Simple Network Management Protocol (SNMP), which is a network management standard protocol applied to TCP (Transmission Control Protocol) / IP (Internet Protocol) networks. The protocol can support network management systems to monitor whether the network devices connected to the distributed computing power network have any conditions that cause attention in management. In addition, SNMP uses a polling mechanism, and SNMP uses User Datagram Protocol (UDP) messages as the carrier, so it is supported by most devices, and at the same time, it ensures the transmission of management information between any two points, which is convenient for administrators to retrieve information at any node on the network and troubleshoot faults.

[0060] Therefore, the embodiments of the present application can use the SNMP protocol and related components for data collection, wherein the related components include a network management system NMS (Network Management System), an agent process (Agent), a managed object (Managed Object), and a management information base MIB (Management Information Base), wherein: the Agent is an agent process in the managed device, which is used to maintain the information data of the managed device and respond to the request from the NMS, and report the management data to the NMS sending the request; the Managed Object refers to the managed object. Each device can contain multiple managed objects, and the managed object can be a certain hardware in the device, or a set of parameters configured on the hardware and software (such as a routing protocol). For example, in the link data collection of the distributed computing power network shown in the figure, the agent process (Agent) is installed and started on the router (Router) and the computing power server (Server), and is connected to the network management system NMS. Every time interval Δt (i.e. multiple sub-time intervals in the first time interval), the traffic data collection and computing power related data collection of each device interface in the whole network are performed, wherein: Figure 2

[0061] The initial traffic sequence includes the traffic value (also known as "throughput") of the first link at each time point in the first time interval, which is used to reflect the amount of data successfully transmitted by the first link at the corresponding time point, and is usually measured in bits per second (bps), bytes per second (Byte / s), or packets per second (pps).

[0062] The initial computing power sequence includes the computing power information of the first link at each time point in the first time interval, which is a measure of the comprehensive capability of the computing power server. Therefore, the computing power information is determined by the computing capability, the communication capability, the memory capability, and the storage capability, so the computing power information of a single first link can be represented by the following formula:

[0063] h=β1h c +β2h n +β3h m +β4h s

[0064] Wherein, h represents the computing power information of a certain computing power server; h c represents the computing capability, specifically the CPU (Central Processing Unit, Central Processing Unit) utilization rate; h n represents the communication capability, specifically the network bandwidth; h m represents the memory capability, specifically the memory capacity; h s ​denotes the storage capability, specifically the storage capacity; β1, β2, β3, β4 respectively denote the weight coefficients of h c , h n , h m , h s .

[0065] Then, the initial traffic sequence and the initial computing power sequence are preprocessed respectively to obtain a corresponding first traffic sequence and a first computing power sequence, wherein the preprocessing operation includes but is not limited to: missing value filling, normalization processing, etc.

[0066] As an optional implementation, in the technical solution provided in the above step S104, the first adjacency matrix can be constructed by the following method, including:

[0067] Step S1041, the distance value between the first traffic sequence of each two first links in the plurality of first links is calculated, and the traffic similarity between each two first links is determined according to the distance value.

[0068] Specifically, the embodiment of the application first calculates the distance value between the first traffic sequence of each two first links in the plurality of first links by using the dynamic time warping algorithm (Dynamic Time Warping, DTW). For example, for the traffic time sequence of two nodes x and y, the accumulated distance before each pair of (x i , y j ) points is calculated as follows:

[0069] d i，j =dist(x i , y j )+min(d i-1，j , d i，j-1 , d i-1，j-1 )

[0070] Wherein dist() represents the Euclidean distance. Therefore, the accumulated distance matrix is obtained by using the above calculation formula. m，n , m, n are the lengths of the two traffic time sequences x and y to be compared.

[0071] Then, the traffic similarity between each two first links is determined according to the distance value, wherein the greater the distance value, the smaller the corresponding traffic similarity, and the smaller the distance value, the greater the corresponding traffic similarity.

[0072] In addition to the implementation solutions listed above, based on the basic concept of the present application, those skilled in the art can also calculate the distance between time series through other technical solutions, for example, those skilled in the art can transform the above-mentioned implementation solutions, which should also be within the protection scope of the present application.

[0073] Step S1042, respectively taking each first link as the row and column of the matrix, and taking the inverse of the traffic similarity between each two first links in the plurality of first links as the matrix element, to construct the first adjacency matrix.

[0074] That is, taking each first link as the row and column of the matrix, and taking the inverse of the traffic similarity between each two first links in the plurality of first links as the matrix element , to obtain the first adjacency matrix corresponding to the first space-time graph.

[0075] For example, Figure 3 is a schematic diagram of a topological structure graph and a space-time graph of an optional distributed computing power network according to an embodiment of the present application, as Figure 3 shown. Figure 3 The left graph in the figure is a topological structure graph G real corresponding to a certain distributed computing power network, and the right graph is a space-time graph G ST constructed by taking each link in the topological structure graph as a graph node, building an edge if two links are connected to the same router, and taking the inverse of the traffic similarity between each link as the weight of the edge.

[0076] Further, input each first traffic sequence, each first computing power sequence, and the first adjacency matrix obtained in step S104 into the pre-trained target traffic prediction model, to perform space-time traffic prediction on the traffic through the model, and obtain the first traffic prediction sequence of each first link in the second time period.

[0077] Optionally, the training process of the above-mentioned target traffic prediction model includes:

[0078] Step 1: Obtain a training sample set and a sample label set.

[0079] The training sample set includes: a training sample composed of a second traffic sequence of a second link in a third time period in a second distributed computing power network, a second computing power sequence of a computing power server corresponding to the second link in the third time period, and a second adjacency matrix corresponding to a second space-time graph of the second distributed computing power network in the third time period, and the sample label set includes: a second traffic sequence of the second link in a fourth time period as a sample label corresponding to the training sample. The third time period is a time period before the fourth time period, and the third time period is a historical time period before the first time period.

[0080] Second step: constructing an initial traffic prediction model.

[0081] wherein the initial traffic prediction model comprises at least a plurality of spatio-temporal graph convolution modules, and each spatio-temporal graph convolution module comprises at least two gated convolution networks (GCN) and one graph convolutional network (GCN).

[0082] Specifically, Figure 4 is an optional initial traffic prediction model according to an embodiment of the present application, wherein the model comprises two spatio-temporal graph convolution modules, and each spatio-temporal graph convolution module comprises a first gated convolution network, a graph convolutional network and a second gated convolution network connected in sequence. Wherein:

[0083] The main function of the first gated convolution network is to preliminarily extract the dynamic characteristics in the time series data (i.e. the first traffic sequence, the first computing power sequence), especially the time dependence in the short term. It can capture the local change trend and pattern of traffic value and computing power information over time through convolution operation and gating mechanism, which is very crucial for understanding the instantaneous change and short-term prediction of traffic. In addition, the gated convolution network can selectively pass or suppress information through its gating unit, effectively filtering out noise and retaining more valuable time features in the sequence. Based on the processing of the first network, the second gated convolution network can further deepen the feature extraction of the time series, especially in capturing long-term dependencies and processing more complex time series patterns. It can learn features from the output of the first gated convolution network again, further adjust the information flow through the gating mechanism, and focus on longer-term traffic trends and periodic changes. Therefore, the combination of the two gated convolution networks can better capture the temporal dynamic characteristics of traffic data through multi-level time feature extraction, including short-term instantaneous changes and long-term trends, as well as the changes of computing power information over time. This design takes advantage of the ability of gated convolution network to process time series data, while also increasing the depth and complexity of the model by stacking multiple network layers, thereby improving the model's ability to process complex time series and prediction accuracy.

[0084] In addition, the graph convolutional network can capture and learn the spatial dependence of the data, especially in graph structured data, it can effectively aggregate the information of nodes and their neighbors, and extract the feature representation of nodes in the graph. In the embodiments of the present application, it is used to consider the correlation of traffic on different links, and the dependence relationship between nodes is measured and learned through the adjacency matrix. This mechanism can capture the global distribution of traffic in the network and the mutual influence between nodes.

[0085] In summary, the two gated convolutional networks are responsible for processing the dynamic characteristics of the time series in the spatio-temporal graph convolution module, ensuring that the model can capture the regularity of traffic changes over time while filtering unnecessary noise; while the graph convolutional neural network focuses on the spatial structure, learning the spatial correlation of traffic on different links by aggregating the information of nodes and their neighbors. This modular design enables the model to handle the spatio-temporal characteristics of network traffic simultaneously, enhancing the accuracy and robustness of traffic prediction from multiple angles.

[0086] Step 3: Use the training sample set and sample label set to iteratively train the initial traffic prediction model to obtain the target traffic prediction model.

[0087] Specifically, the implementation process of the above-mentioned third step includes:

[0088] First, in each round of iterative training process, the training samples in the training sample set are input into the initial traffic prediction model to obtain the second predicted traffic sequence output by the initial traffic prediction model. Therefore, when predicting the second traffic prediction value in the next H time steps based on the first traffic value in the previous M time steps and the computing power information, the expression of the corresponding second predicted traffic sequence can be written as:

[0089]

[0090] wherein, represents the first traffic value on n links at time t; represents the observed value of the computing power information corresponding to the n links at time t; represents the second traffic prediction value on n links at time t.

[0091] Then, the target loss function is constructed according to the second predicted traffic sequence output by the initial traffic prediction model and the second traffic sequence corresponding to each training sample. Therefore, the expression of the above-mentioned target loss function can be written as:

[0092]

[0093] wherein W θ represents all trainable parameters in the model; represents the first traffic observation value on n links at time t; is the observed value of the computing power information corresponding to the n links at time t, represents the second traffic prediction value on n links at time t.

[0094] Finally, a preset optimization algorithm is used to minimize the target loss function until the model converges, and the target traffic prediction model is obtained. The optimization algorithm includes but is not limited to: RMSProp optimization algorithm, Stochastic Gradient Descent (SGD) algorithm, etc.

[0095] In addition, after the training of the target traffic prediction model is completed through the above steps, the model performance can also be evaluated using the validation set, and the hyperparameters can be adjusted to improve the model performance. Then, the model is finally evaluated through the test set, and the accuracy and other performance indicators are calculated to ensure that the model has good generalization ability.

[0096] Furthermore, after obtaining the first traffic prediction sequence of each first link in the second time period through the above steps, the distribution and scheduling of the computing resource can be reasonably adjusted using this information to ensure the efficient operation of the network.

[0097] As an optional implementation, in the technical solution provided in the above step S108, the method can be implemented by the following method, comprising:

[0098] Step S1081, for each first link, determining the demand computing resource information of the first link at each time in the second time period according to the first traffic prediction sequence of the first link in the second time period.

[0099] Specifically, in the technical solution provided in the above step S1081, the method comprises: matching the first traffic prediction value at each time in the first traffic prediction sequence with a preset traffic-computing resource mapping table to determine the demand computing resource information of the first link at each time in the second time period.

[0100] That is, by analyzing the correspondence between the traffic level and the computing resource information consumption based on experimental or historical data, a traffic-computing resource mapping table is obtained, and therefore, the traffic-computing resource mapping table records the mapping relationship between multiple traffic values and demand computing resource information. For example, the traffic-computing resource mapping table can record the CPU utilization, network bandwidth consumption, memory usage, etc. of the computing resource server corresponding to the link under different traffic levels, and then the relationship between the traffic and the computing resource value consumption is fitted through regression analysis or a machine learning model (such as a neural network).

[0101] Step S1082, for the first time in the second time period, determining whether the demand computing resource information of the first link at the first time in the second time period is greater than the first computing resource information of the first link at the last time in the first time period. If yes, step S1083 is executed; otherwise, step S1084 is executed.

[0102] In step S1083, a first difference value between the demand computing power information of the first link at the first time in the second time period and the first computing power information of the first link at the last time in the first time period is determined, and computing power resources corresponding to the first difference value are scheduled from other idle links.

[0103] In step S1084, a second difference value between the first computing power information of the first link at the last time in the first time period and the demand computing power information of the first link at the first time in the second time period is determined, the released computing power information of the first link at the first time in the second time period is obtained, and a sum of the second difference value corresponding computing power information and the released computing power information is taken as the first computing power information of the first link at the first time in the second time period, and the above judgment process is repeatedly executed until all times in the second time period are traversed or the first computing power information of the first link is zero.

[0104] The above allocation and scheduling process can be understood as follows: for each time in the second time period, the demand computing power information of the first link at the time is determined according to the first link traffic prediction value at the time through the above mapping relationship. Then, for the first time in the second time period, the demand computing power information is compared with the available computing power information of the computing power server at the last time in the first time period. If the demand computing power information exceeds the available computing power information, a difference value is calculated, and other idle computing power servers in the network are searched to schedule computing power information corresponding to the difference value from other idle links. If the demand computing power information is less than or equal to the available computing power information, a reverse difference value is calculated, which indicates that the computing power server still has remaining available computing power information, and the released additional computing power information at the time is determined, wherein if no computing power resources are released at the time, the corresponding additional computing power information is zero. The available computing power information of the computing power server is updated according to the computing power information corresponding to the second difference value and the released computing power information. Next, for each time in the second time period, the above process is repeated until the second time period ends or the available computing power information of the computing power server is consumed.

[0105] It should be noted that in actual application, a resource scheduling algorithm (such as genetic algorithm, particle swarm optimization algorithm, simulated annealing algorithm, etc.) can be used to optimize the allocation of computing power information to ensure that the demand for traffic prediction is met while minimizing resource waste and improving resource utilization efficiency. In addition, during the computing power resource scheduling process, the changes in traffic demand and the actual working state of the computing power server need to be monitored in real time, and the computing power resource allocation strategy needs to be adjusted in a timely manner to cope with the dynamic changes and uncertainties of the network.

[0106] The network computing power resource management method provided by the scheme has the following technical advantages compared with the prior art:

[0107] (1) Since the existing network traffic prediction method is often based on historical traffic data for prediction, the node computing power information is introduced in the embodiment of the application, which plays a key role in resource scheduling and traffic prediction, and can better adapt to the network environment with dynamic changes in resources. At the same time, the spatio-temporal graph convolution technology is innovatively integrated in the prediction process, which can capture the time series characteristics and spatial dependence of traffic at the same time. This comprehensive prediction method can more accurately reflect the network dynamics and improve the prediction accuracy.

[0108] (2) The embodiment of the application uses the DTW algorithm to calculate the matrix elements (i.e. the association weight between links) in the adjacency matrix corresponding to the spatio-temporal graph, so as to ensure that the dynamic association and spatial structure between links are focused on in the model prediction process, and the robustness and adaptability of the prediction model are improved.

[0109] (3) The embodiment of the application can manage the computing power resources according to the traffic prediction sequence, which provides more intelligent decision support for network managers, helps them make more reasonable and timely computing power resource scheduling decisions, and thus effectively improves the stability and response speed of the network.

[0110] In summary, the scheme of the application can realize more accurate traffic prediction and more efficient computing power resource management, and has obvious technical advantages in improving network performance and resource utilization efficiency.

[0111] Embodiment 2

[0112] According to the embodiments of the application, a network computing power resource management device for implementing the network computing power resource management method in Embodiment 1 is also provided, as shown in Figure 5 The network computing power resource management device at least includes: an acquisition module 52, a determination module 54, a prediction module 56 and a management module 58, wherein:

[0113] The acquisition module 52 is configured to acquire a first traffic sequence and a first computing power sequence of each of a plurality of first links in a first distributed computing power network within a first time period;

[0114] The determination module 54 is configured to construct a first spatio-temporal graph corresponding to the first distributed computing power network by taking the plurality of first links as nodes, and determine a first adjacency matrix corresponding to the first spatio-temporal graph according to the first traffic sequence of each first link;

[0115] The prediction module 56 is configured to analyze each first traffic sequence, each first computing power sequence and the first adjacency matrix by using a pre-trained target traffic prediction model, to obtain a first traffic prediction sequence of each first link within a second time period, wherein the second time period is a time period after the first time period;

[0116] The management module 58 is configured to manage the computing resource of the first distributed computing network according to the first traffic prediction sequence of each first link in the second time period.

[0117] It should be noted that the modules in the network computing resource management apparatus in the embodiments of the present application correspond to the implementation steps of the network computing resource management method in Embodiment 1 one by one. Since Embodiment 1 has been described in detail, the details not embodied in this embodiment can be referred to Embodiment 1, and will not be described in detail here.

[0118] Embodiment 3

[0119] According to the embodiments of the present application, a computer program product is also provided, which includes a computer program. When the computer program is executed by a processor, the network computing resource management method in Embodiment 1 is implemented.

[0120] According to the embodiments of the present application, a non-volatile storage medium is also provided, which includes a stored computer program. The device where the non-volatile storage medium is located executes the network computing resource management method in Embodiment 1 by running the computer program.

[0121] According to the embodiments of the present application, a processor is also provided, which is used to run a computer program. When the computer program is run, the network computing resource management method in Embodiment 1 is executed.

[0122] According to the embodiments of the present application, an electronic device is also provided, which includes a memory and a processor. The memory stores a computer program, and the processor is configured to execute the network computing resource management method in Embodiment 1 by the computer program.

[0123] Specifically, when the computer program is run, the following steps are implemented: obtaining the first traffic sequence and the first computing resource sequence of each first link in the first time period in the first distributed computing network; constructing a first space-time graph corresponding to the first distributed computing network by taking the plurality of first links as nodes, and determining a first adjacency matrix corresponding to the first space-time graph according to the first traffic sequence of each first link; analyzing each first traffic sequence, each first computing resource sequence and the first adjacency matrix by using a pre-trained target traffic prediction model to obtain a first traffic prediction sequence of each first link in a second time period, wherein the second time period is a time period after the first time period; and managing the computing resource of the first distributed computing network according to the first traffic prediction sequence of each first link in the second time period.

[0124] As an optional implementation, the electronic device can exist in the form of a mobile terminal, a computer terminal or a similar computing device. Figure 6A hardware structure block diagram of an electronic device for implementing a network computing resource management method is shown. As shown in Figure 6 The electronic device 60 can include one or more processors 602 (the processor 602 can include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 604 for storing data, and a transmission device 606 for communication functions. In addition, it can also include a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which can be included as one of the ports in the BUS bus), a network interface, a power supply and / or a camera. Those skilled in the art can understand that Figure 6 The structure shown is only schematic, which does not limit the structure of the above-mentioned electronic device. For example, the electronic device 60 can include more or fewer components than those shown in Figure 6 or have a different configuration than that shown in Figure 6

[0125] It should be noted that the one or more processors 602 and / or other data processing circuits described above can be referred to herein as "data processing circuits" in general. The data processing circuit can be embodied in whole or in part as software, hardware, firmware or any combination thereof. In addition, the data processing circuit can be a single independent processing module, or all or part of any one of the other elements combined into the electronic device 60. As referred to in the embodiments of the present application, the data processing circuit serves as a processor to control (for example, selection of a variable resistance terminal path connected to an interface).

[0126] The memory 604 can be used to store software programs and modules of application software, such as program instructions / data storage devices corresponding to the network computing resource management method in the embodiments of the present application. The processor 602 executes various functional applications and data processing by running the software programs and modules stored in the memory 604, that is, implements the vulnerability detection method of the application program described above. The memory 604 can include a high-speed random access memory, and can 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 604 can further include a memory remotely located with respect to the processor 602, which can be connected to the electronic device 60 through 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 a combination thereof.

[0127] ​The transmission device 606 is configured to receive or send data via a network. The network can include a wireless network provided by a communication provider of the electronic device 60. In an example, the transmission device 606 includes a network interface controller (NIC) that can be connected to other network devices through a base station to communicate with the Internet. In an example, the transmission device 606 can be a radio frequency (RF) module that is configured to communicate with the Internet wirelessly.

[0128] The display can be a liquid crystal display (LCD) that is touch screen type, for example, which can enable a user to interact with a user interface of the electronic device 60.

[0129] The above-mentioned embodiment numbers are only for description, and do not represent the advantages or disadvantages of the embodiments.

[0130] In the above-mentioned embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0131] In several embodiments provided in the present application, it should be understood that the disclosed technical contents can be implemented by other ways. Among them, the above-mentioned device embodiments are only schematic, for example, the division of units can be a logical function division, and actual implementation can have another division way, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, units or modules, which can be electrical or other forms.

[0132] The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on multiple units. According to actual needs, part or all of the units can be selected to achieve the purpose of the present embodiment scheme.

[0133] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be realized in the form of hardware or in the form of software functional unit.

[0134] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0135] The above is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application.

Claims

1. A network computing resource management method, characterized in that, The method comprises the following steps: obtaining a first traffic sequence and a first computing power sequence of each of a plurality of first links in a first distributed computing power network within a first time period; constructing a first space-time graph corresponding to the first distributed computing power network by taking the plurality of first links as nodes, and determining a first adjacency matrix corresponding to the first space-time graph according to the first traffic sequence of each of the first links, comprising: calculating the distance value between the first traffic sequence of each of the plurality of first links, and determining the traffic similarity between each of the first links according to the distance value, wherein the greater the distance value, the smaller the corresponding traffic similarity, and the smaller the distance value, the greater the corresponding traffic similarity; taking each of the first links as a row and a column of a matrix, respectively, and taking the reciprocal of the traffic similarity between each of the plurality of first links as a matrix element to obtain the first adjacency matrix corresponding to the first space-time graph; analyzing each of the first traffic sequence, each of the first computing power sequence, and the first adjacency matrix by using a pre-trained target traffic prediction model to obtain a first traffic prediction sequence of each of the first links within a second time period, wherein the second time period is a time period after the first time period; managing the computing power resources of the first distributed computing power network according to the first traffic prediction sequence of each of the first links within the second time period.

2. The method of claim 1, wherein, Obtaining a first traffic sequence and a first computing power sequence of each of a plurality of first links in a first distributed computing power network within a first time period, comprising: determining a plurality of first links in the first distributed computing power network; for each of the first links, collecting an initial traffic sequence and an initial computing power sequence of the first link within the first time period by using a preset protocol, wherein the initial traffic sequence includes a traffic value of the first link at each time within the first time period, and the traffic value reflects the amount of data successfully transmitted by the first link at the corresponding time, the initial computing power sequence includes computing power information of the first link at each time within the first time period, and the computing power information is determined by at least computing power, communication capability, memory capability, and storage capability; preprocessing the initial traffic sequence and the initial computing power sequence to obtain the corresponding first traffic sequence and the first computing power sequence.

3. The method of claim 1, wherein, The training process of the target traffic prediction model comprises: obtaining a training sample set and a sample label set, wherein the training sample set includes a second adjacency matrix corresponding to a second space-time graph of a second distributed computing power network within a third time period, a second traffic sequence and a second computing power sequence of a second link in the second distributed computing power network within the third time period, the sample label set includes a second traffic sequence of the second link within a fourth time period as a sample label corresponding to the training sample, and the third time period is a time period before the fourth time period, and the third time period is a time period before the first time period; The initial traffic prediction model is constructed, wherein the initial traffic prediction model at least includes a plurality of spatio-temporal graph convolution modules, and each spatio-temporal graph convolution module at least includes two gated convolution networks and one graph convolution network; The initial traffic prediction model is iteratively trained by using the training sample set and the sample label set to obtain the target traffic prediction model.

4. The method of claim 3, wherein, The initial traffic prediction model is iteratively trained by using the training sample set and the sample label set to obtain the target traffic prediction model, including: In each round of iterative training, the training samples in the training sample set are input into the initial traffic prediction model to obtain a second predicted traffic sequence output by the initial traffic prediction model; A target loss function is constructed according to the second predicted traffic sequence output by the initial traffic prediction model and the second traffic sequence corresponding to each training sample; The target loss function is minimized by using a preset optimization algorithm until the model converges, and the target traffic prediction model is obtained.

5. The method of claim 1, wherein, The first distributed computing power network is managed by using the first traffic prediction sequence of each first link in the second time period, including: For each first link, the demand computing power information of the first link at each time in the second time period is determined according to the first traffic prediction sequence of the first link in the second time period; For the first time in the second time period, it is judged whether the demand computing power information of the first link at the first time in the second time period is greater than the first computing power information of the first link at the last time in the first time period; If yes, a first difference value between the demand computing power information of the first link at the first time in the second time period and the first computing power information of the first link at the last time in the first time period is determined, and the computing power resource corresponding to the first difference value is scheduled from other idle links; If not, a second difference value between the first computing power information of the first link at the last time in the first time period and the demand computing power information of the first link at the first time in the second time period is determined, the released computing power information of the first link at the first time in the second time period is obtained, and the sum of the computing power information corresponding to the second difference value and the released computing power information is taken as the first computing power information of the first link at the first time in the second time period, and the above judgment process is repeatedly executed until all times in the second time period are traversed or the first computing power information of the first link is zero.

6. The method of claim 5, wherein, The demand computing power information of the first link at each time in the second time period is determined according to the first traffic prediction sequence of the first link in the second time period, including: The first traffic prediction sequence in the first traffic prediction sequence is matched with a preset traffic computing power mapping table to determine the demand computing power information of the first link at each time in the second time period, wherein the traffic computing power mapping table records the mapping relationship between a plurality of traffic values and demand computing power information.

7. A network computing resource management apparatus, characterized by, including: An acquisition module is configured to acquire a first traffic sequence and a first computing power sequence of each of a plurality of first links in a first distributed computing power network within a first time period; A determination module is configured to construct a first space-time graph corresponding to the first distributed computing power network by taking the plurality of first links as nodes, and determine a first adjacency matrix corresponding to the first space-time graph according to the first traffic sequence of each of the first links, including: calculating a distance value between the first traffic sequences of each two of the first links, and determining a traffic similarity between each two of the first links according to the distance value, wherein the greater the distance value, the smaller the corresponding traffic similarity, and the smaller the distance value, the greater the corresponding traffic similarity; taking each of the first links as a row and a column of a matrix respectively, and taking the reciprocal of the traffic similarity between each two of the first links as a matrix element, to obtain the first adjacency matrix corresponding to the first space-time graph; A prediction module is configured to analyze each of the first traffic sequences, each of the first computing power sequences, and the first adjacency matrix by using a pre-trained target traffic prediction model, to obtain a first traffic prediction sequence of each of the first links within a second time period, wherein the second time period is a time period after the first time period; A management module is configured to manage computing power resources of the first distributed computing power network according to the first traffic prediction sequence of each of the first links within the second time period.

8. A computer program product, characterised in that, The computer program is executed by a processor to implement the network computing power resource management method of any one of claims 1 to 6. The non-volatile storage medium stores a computer program, and a device in which the non-volatile storage medium is located executes the network computing power resource management method of any one of claims 1 to 6 by running the computer program.

9. A non-volatile storage medium, comprising: ​

Images