Fine-grained micro-service resource prediction method based on heterogeneous graph neural network

Abstract

The invention discloses a fine-grained micro-service resource prediction method based on a heterogeneous graph neural network, relates to the field of resource prediction, and is used for improving the accuracy of micro-service resource prediction. According to the method, resource utilization rate sequences of two dimensions of a physical node and a container of a micro-service are coded to obtain an initial hidden state, the initial hidden state is input into a heterogeneous graph neural network for adjustment, and the adjusted hidden state is decoded to obtain a resource prediction result; the heterogeneous graph neural network is constructed according to a heterogeneous graph, and the heterogeneous graph takes a physical node and a container of the micro-service as heterogeneous nodes respectively and takes a relationship between the physical node and the container as an edge. According to the invention, the accuracy and universality of micro-service resource prediction are improved.

Description

Technical Field

[0001] The present invention relates to the field of resource prediction, and in particular to a fine-grained microservice resource prediction method based on heterogeneous graph neural network. Background Art

[0002] In recent years, microservices architecture has become a popular software architecture paradigm, adopted by many organizations and enterprises. Its appeal lies in the ability to decompose large, complex applications into small, independent service units, each of which focuses on performing a specific function or task. In this architecture, each microservice usually has unique tasks and responsibilities, resulting in different resource requirements. These differences span computing resources such as CPU and memory, storage resources, network bandwidth, etc.

[0003] Effective resource management and forecasting are critical to ensuring the continued performance and availability of microservice architectures. Monitoring, analyzing, and forecasting the resource requirements of microservices can enable better resource configuration and allocation, ensuring that the individual needs of microservices are met while avoiding resource waste. However, resource requirements in microservice architectures are subject to dynamic influences. The multifaceted nature of these dynamic factors makes it challenging to accurately forecast resource requirements, especially in the face of sudden load changes, requiring smarter and more accurate resource management and forecasting strategies that strike a balance between performance and cost. To meet these challenges, forecasting the resource requirements of microservices becomes a critical task.

[0004] The problem of microservice resource prediction involves time series analysis, while traditional time series prediction methods are mainly oriented towards single time series. The resource utilization of a microservice is affected by multiple factors, including not only its own historical resource utilization, but also the load of other microservices related to its business, the load occupancy of other microservices under the current server node, the load changes of other instances of the current microservice, and the physical environment changes of the server. Therefore, it is far from enough to only consider the resource utilization time series data of a single microservice node.

[0005] The patent document with the Chinese publication number CN115242797A proposes a client load balancing method and system for microservice architecture, which includes: after receiving the service request information sent by the user, the client obtains the instance registry of the service deployed in the cluster from the registration center; obtains the historical information of the resource utilization of all servers deploying this service from the information collector module, and uses the server resource prediction algorithm based on the improved ARIMA model to predict the resource utilization of the corresponding node in the next second; the server calculates the dynamic weight load method according to the historical resource utilization and the predicted resource utilization in the next second, obtains the corresponding server load value, returns the load value information to the client and compares and selects the server with the best load value to execute the request. Although this method improves the utilization of cluster resources, it is based on the traditional ARIMA time series analysis, which is a method based on statistical laws. However, the laws of different time periods of microservice operation are not constant, and this method is not suitable for predicting the utilization of microservice resources when the microservice load fluctuates greatly.

[0006] The patent document with the Chinese publication number CN117891559A proposes a method for intelligent prediction of K8s resource scheduling, which includes: collecting and preprocessing the workload and related node time series data on the Kubernetes platform, classifying and analyzing the existing time series data, and obtaining the historical characteristics of the workload and related nodes; using the traditional time series prediction algorithm to predict resources; using the machine learning algorithm to generate an algorithm model based on massive time series data, and perform algorithm performance evaluation; based on the resource prediction results, using the optimization algorithm to formulate Pod migration and resource allocation plans, and generate scheduling decisions; executing the scheduling decision on the Kubernetes platform, and feeding back the execution results to the monitoring data. This method can save labor costs, but can only analyze a single sequence, and is also not suitable for microservice resource prediction problems. Summary of the invention

[0007] The purpose of the present invention is to provide a fine-grained microservice resource prediction method based on heterogeneous graph neural network to solve all or part of the above-mentioned problems, and to predict microservice resources from multi-dimensional features to improve the accuracy of microservice resource prediction.

[0008] The technical solution adopted by the present invention is as follows:

[0009] A fine-grained microservice resource prediction method based on heterogeneous graph neural network, comprising:

[0010] The resource utilization sequences of the physical nodes and containers of the microservices are encoded to obtain an initial hidden state, which is input into a heterogeneous graph neural network for adjustment, and the adjusted hidden state is decoded to obtain a resource prediction result; the heterogeneous graph neural network is constructed according to a heterogeneous graph, and the heterogeneous graph uses the physical nodes and containers of the microservices as heterogeneous nodes, respectively, and uses the relationships between the physical nodes and containers as edges.

[0011] Furthermore, the relationship between the physical nodes and the containers is used as an edge, including an edge constructed based on the relationship between physical nodes, an edge constructed based on the relationship between physical nodes and containers, and an edge constructed based on the relationship between containers.

[0012] Furthermore, the edges constructed based on the relationship between containers include edges constructed based on the relationship between containers under the same physical node and edges constructed based on the relationship between containers under different physical nodes.

[0013] Furthermore, the method of constructing edges based on the relationship between physical nodes includes:

[0014] Calculate the correlation between resource utilization sequences between two physical nodes;

[0015] A first type of edge is generated between two physical nodes whose correlation reaches or exceeds a first threshold.

[0016] Furthermore, the method of constructing edges based on the relationship between physical nodes and containers includes:

[0017] The second type of edge is generated between a physical node and each container deployed on it.

[0018] Furthermore, the method of constructing edges based on the relationship between containers includes:

[0019] The third type of edge is generated between containers deployed on the same physical node, and the fourth type of edge is generated between containers deployed on different physical nodes.

[0020] Furthermore, a bidirectional recurrent gated neural network Bi-GRU is used to encode the resource utilization sequence of the physical nodes and containers of the microservices to obtain the hidden state.

[0021] Furthermore, a sliding window is used to select the historical resource utilization sequences of the physical nodes and containers of the microservices for encoding.

[0022] Furthermore, the sliding window selects historical resource utilization sequences of two dimensions, physical nodes and containers of the microservice, in a time-series continuous manner for encoding.

[0023] Furthermore, the resource utilization sequences of the physical nodes and containers of the microservices are encoded to obtain the initial hidden state, including:

[0024] Compress the resource utilization sequences of the physical nodes and containers of the microservice into resource utilization information of the same size as the hidden layer state;

[0025] Encoding the resource utilization information using a hidden layer to obtain a first hidden state;

[0026] The resource utilization information residual is connected to the first hidden state to obtain an initial hidden state.

[0027] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0028] The resource utilization of microservices is affected by multiple factors, including not only the historical resource utilization of the microservices themselves, but also the load of other microservices related to the business, the load occupancy of other microservices under the current server node, the load changes of other instances in the current microservice, and the physical environment changes of the server. Traditional time series prediction methods mainly face a single time series, and it is difficult to accurately predict the multiple influencing factors faced by microservice resource utilization. The present invention adopts a data collection strategy for multi-node resource utilization, analyzes the resource utilization relationship between physical nodes, and introduces more fine-grained features, which not only considers the resource utilization information of physical nodes, but also combines the resource utilization information at the container level. Under this design concept, a resource prediction model of encoder-heterogeneous graph neural network-decoder is designed, so that the model can deeply mine the relationship between physical nodes, between physical nodes and containers, and between containers, thereby improving the accuracy of microservice resource prediction. Experiments have proved that the method of the present invention performs extremely well in the short-term prediction of microservice resource utilization, and can still maintain a low error level in long-term prediction, with strong universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present invention will now be described by way of example with reference to the accompanying drawings, in which:

[0030] Figure 1 This is the architecture diagram of the resource prediction model of the encoder-heterogeneous graph neural network-decoder structure.

[0031] Figure 2 This is the encoder structure diagram.

[0032] Figure 3 It is a schematic diagram of heterogeneous nodes and four types of edges of a heterogeneous graph.

[0033] Figure 4 This is the decoder structure diagram.

[0034] Figure 5 It is a performance diagram of the comparative test between the method of the present application and the benchmark method.

[0035] Figure 6 It is a performance diagram of the method of the present application in long-term prediction. DETAILED DESCRIPTION

[0036] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0037] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other equivalent or alternative features having similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0038] The embodiment of the present application provides a fine-grained microservice resource prediction method based on a heterogeneous graph neural network. The model architecture for implementing the method is as follows: Figure 1 As shown in the figure. This method encodes the resource utilization sequence of the physical nodes and containers of the microservice to obtain the initial hidden state, inputs it into the heterogeneous graph neural network for adjustment, and decodes the adjusted hidden state to obtain the resource prediction result. The heterogeneous graph neural network is constructed based on the heterogeneous graph, and the heterogeneous graph uses the physical nodes and containers of the microservice as heterogeneous nodes, and the relationships between the physical nodes and containers as edges.

[0039] In some embodiments, the encoding-decoding process involved in the method uses a Seq2seq (Sequence to Sequence) model as the infrastructure. The Seq2seq model is a neural network machine learning model that maps sequences to sequences and solves the gradient vanishing problem by using LSTM or GRU networks. The Seq2seq model mainly consists of two parts: an encoder and a decoder. The encoder converts the input into a hidden state vector containing the content of the input item. The decoder does the opposite, converting the vector into an output sequence and using the output of the previous step as the input for the next step.

[0040] Since the Seq2seq model is limited to the prediction processing of a single time series and does not consider the resource utilization relationship across different nodes or containers, it is difficult to obtain more accurate prediction accuracy when directly used in the microservice resource utilization prediction task. On this basis, the embodiment of the present application embeds a heterogeneous graph neural network between the encoder and decoder of the Seq2seq model. The heterogeneous graph neural network allows the resource utilization of microservice physical nodes and containers to be used as heterogeneous nodes in the heterogeneous graph, and the relationship between physical nodes and containers as the edge of the heterogeneous graph. The potential relationship between the physical nodes and containers of the microservice is learned, thereby mining more features that affect the resource utilization of the microservice from the resource utilization sequence, and improving the accuracy of resource utilization prediction.

[0041] In some embodiments, a bidirectional recurrent gated neural network Bi-GRU is used to encode the resource utilization sequence of the physical node and container of the microservice to obtain a hidden state. Of course, a bidirectional long short-term memory network Bi-LSTM can also be used for encoding.

[0042] The introduction of the Bi-GRU network further enhances the modeling capability of temporal features. Compared with the traditional unidirectional recurrent neural network, the Bi-GRU network can learn the forward and backward features of the data more comprehensively, while considering historical and future information. This combination makes the resource prediction model of this embodiment have obvious advantages in handling dynamic and complex resource utilization in microservice environments.

[0043] like Figure 2 The structure diagram of Bi-GRU is shown below. The function of the encoder Bi-GRU is to convert the historical sequence x of resource utilization into the hidden state h,x t is the resource utilization feature vector of the time series x of resource utilization at time step t. At time step t, the forward recurrent neural network will input x t The feature vector and the hidden state of the previous time step Transform to the hidden state of the current time step Transformation function of hidden layer of feedforward recurrent neural network as follows:

[0044]

[0045] The specific transformation process of the hidden layer of the recursive neural network is as follows:

[0046] Reset gate: r t =σ(W r ·[h t-1 ,x t ]),

[0047] Update gate: z t =σ(Wz ·[h t-1 ,x t ]),

[0048] New candidate status:

[0049] Hidden state: h t =(1-z t )*h t-1 +z t *h t .

[0050] σ and tanh are activation functions, W r , W z , is the corresponding transposed matrix.

[0051] For the backward recurrent neural network input, use the backward hidden state of the next time step Use function express:

[0052]

[0053] For a time series x of length H, the forward hidden layer state is finally obtained and the backward hidden layer state The two hidden layer states are concatenated and combined to obtain the final hidden layer state h as the initial hidden state.

[0054] Since the encoder needs to compress all the information in the input sequence into the size of the hidden layer state while capturing and retaining information from distant time steps, it is difficult to avoid the loss of fine-grained details and the forgetting of some long-term dependencies. To this end, a simplified version of the residual connection is added based on the hidden layer state, and the original input is added to the final hidden layer state. This makes the model easier to train and better captures long-term dependencies in the sequence when processing long sequences. Specifically, the resource utilization sequence of the physical nodes and containers of the microservice is encoded to obtain the initial hidden state, including:

[0055] Compress the resource utilization sequences of the physical nodes and containers of the microservice into resource utilization information of the same size as the hidden layer state;

[0056] Encoding the resource utilization information using a hidden layer to obtain a first hidden state;

[0057] The resource utilization information residual is connected to the first hidden state to obtain an initial hidden state.

[0058] In addition, in an embodiment of the present application, the resource utilization sequence fed into the encoder is selected from the historical resource utilization sequence through a sliding window, that is, the historical resource utilization sequence of the physical node and container dimensions of the microservice is selected using a sliding window for encoding.

[0059] The sliding window allows the model to be trained using historical time series data, capturing past trends and patterns and helping to better predict changes in resource utilization. It is also highly adaptable and can flexibly capture short-term and long-term changes in the sequence, making it suitable for different types of time series data. More importantly, for real-time data, the sliding window can update the model in a timely manner to maintain sensitivity to new observations in the sequence.

[0060] In some embodiments, the sliding window selects the historical resource utilization sequences of the physical nodes and containers of the microservice in a time-series continuous manner for encoding. The so-called time-series continuous manner means that the resource utilization sequence selected by each sliding is continuous in time. For example, the sequence selected in the current time step is {x1, x2, …, x H}, then the sequence selected for the next time step is {x H+1 ,x H+2 ,…,x 2H}.

[0061] The heterogeneous graph neural network is constructed based on the heterogeneous graph, that is, based on the heterogeneous graph, the heterogeneous graph neural network is designed to learn the heterogeneous graph. The heterogeneous graph constructs the physical nodes and containers of the microservices as heterogeneous nodes respectively, and constructs edges based on the relationships between the physical nodes and containers. The edges in the heterogeneous graph include edges constructed based on the relationships between physical nodes, edges constructed based on the relationships between physical nodes and containers, and edges constructed based on the relationships between containers.

[0062] Specifically, the edges constructed for the heterogeneous graph are:

[0063] (1) Edges constructed based on the relationships between physical nodes.

[0064] The edges between physical nodes are Figure 3 (a). There are similar resource utilization characteristics between microservice physical nodes. In some embodiments, the method of constructing edges based on the relationship between physical nodes includes:

[0065] Calculate the correlation between resource utilization sequences between two physical nodes;

[0066] A first type of edge is generated between two physical nodes whose correlation reaches or exceeds a first threshold.

[0067] In some optional methods, the Pearson correlation coefficient is used to measure the correlation between physical nodes to construct the edge weight. The calculation method is as follows:

[0068]

[0069] Where ρ(x,y) is the correlation coefficient of sequences x and y, cov(x,y) is the covariance of sequences x and y, and H is the sliding window length, that is, the length of the sequence at each time step. and σ x , σ y are the mean and standard deviation of the two series respectively.

[0070] The value range of the Pearson correlation coefficient is between [-1, 1]. The closer the absolute value is to 1, the stronger the correlation between the two resource utilization sequences is. A first-class edge from one physical node to another is established between all physical nodes whose correlation coefficient is greater than the given first threshold δ. ij Indicates whether the first type of edge exists between physical nodes. The formula is as follows:

[0071]

[0072] Where, X i ,X j They represent the resource utilization sequences of physical node i and physical node j respectively.

[0073] (2) Build edges based on the relationship between physical nodes and containers.

[0074] The edges between physical nodes and containers are as follows Figure 3 As shown in (b), since each container needs to be physically deployed on a physical node, the resource utilization of the physical node and the container are physically closely related.

[0075] In some embodiments, a method for constructing an edge based on a relationship between a physical node and a container includes:

[0076] The second type of edge is generated between a physical node and each container deployed on it.

[0077] (3) Build edges based on the relationships between containers.

[0078] There are two types of relationships between containers. The first type is strongly related to the containers under the same node in terms of physical deployment. The second type is also related to the containers of the microservice in terms of logical relationship. Therefore, there are two types of edges between containers, including edges built based on the relationship between containers under the same physical node and edges built based on the relationship between containers under different physical nodes. The two types of edges are as follows: Figure 3 (c) and Figure 3 (d) as shown.

[0079] Specifically, the methods for constructing edges based on the relationship between containers include:

[0080] The third type of edge is generated between containers deployed on the same physical node, such as Figure 3 (c) shows that the fourth type of edge is generated between containers deployed on different physical nodes, such as Figure 3 (d) as shown.

[0081] Through the above method, a heterogeneous graph can be constructed using the physical nodes and containers of microservices and the four types of edges between the physical nodes and containers.

[0082] In some embodiments, a resource prediction model is trained based on a constructed heterogeneous graph using RGCN (Relational Graph Convolutional Network) (which can be replaced by other heterogeneous graph neural networks). In the constructed heterogeneous graph, this embodiment uses RGCN in the following way: Each heterogeneous node, whether it represents a physical node or a container, has an initial representation containing features such as CPU and memory utilization. There are four types of edges in a heterogeneous graph, each representing a different relationship. Taking advantage of the heterogeneity of the heterogeneous graph, the RGCN layer is applied to each edge type. The computational process of information propagation and aggregation of the RGCN layer on the heterogeneous graph can be expressed as:

[0083]

[0084] In the formula, is the representation of node i in layer l, R is the set of edge types, N r (i) is the set of neighbor nodes of node i under edge type r, is the normalized coefficient of the edge of type r between nodes i and j, is the weight matrix of the lth layer under edge type r.

[0085] The RGCN layer allows heterogeneous nodes to communicate and aggregate information with adjacent nodes through multiple rounds of graph convolution operations through different types of edges. In this way, the resource prediction model can more comprehensively understand the relationship between physical nodes, physical nodes and containers, and containers and containers, and improve the prediction accuracy of resource utilization in the microservice architecture. Finally, the RGCN layer is embedded between the encoder and decoder of the Seq2seq model to form an end-to-end heterogeneous graph neural network structure.

[0086] For the decoder, it decodes the hidden state generated by the encoder and adjusted by the heterogeneous graph neural network to obtain the resource utilization prediction value of the future window.

[0087] like Figure 4The figure shows the structure of the decoder. The decoder also uses the GRU network as its basic component and uses the output of the previous time step as the input of the current time step. It can better utilize previous information to generate predictions for the next time step.

[0088] See also Figure 1-Figure 4 In the embodiments of the present application, the training of the resource prediction model requires the use of historical resource utilization sequences, which are then input into the encoder-heterogeneous graph neural network-decoder model to predict resource utilization in the future. Therefore, in some embodiments, the microservice resource prediction method includes the following steps:

[0089] (1) Collect resource utilization sequences of the physical nodes and containers of the microservice. The resource utilization sequences collected here are resource utilization sequences of historical records and are used to train resource prediction models.

[0090] (2) Use the bidirectional recurrent gated neural network Bi-GRU to extract the bidirectional temporal features of the resource utilization sequence to obtain the initial hidden state.

[0091] (3) Two different types of heterogeneous node and edge attributes of the heterogeneous graph are constructed based on the hidden states of resource utilization of two types of physical nodes and containers (i.e., the potential relationship between the two).

[0092] (4) The correlation coefficient is used to measure the correlation between physical nodes of microservices, and edges between physical nodes are constructed for the heterogeneous graph. The edges between physical nodes and containers are constructed based on the physical deployment relationship between microservice containers and physical nodes. Two types of edges are constructed between containers based on the physical connection of the same physical deployment and the logical connection of the same subordinate microservices. A total of four different edges are used to connect two types of heterogeneous nodes to construct a heterogeneous graph.

[0093] (4) A heterogeneous graph neural network is constructed based on the heterogeneous graph. In this embodiment, the heterogeneous graph neural network uses RGCN, and the RGCN graph neural network is used to adjust the hidden state of the resource utilization characteristics of the physical nodes and containers.

[0094] (5) The adjusted hidden state is decoded through the recurrent gated neural network GRU (Gated Recurrent Unit) to obtain the resource prediction value for a window in the future.

[0095] According to an embodiment of the present application, the microservice operation data disclosed by Alibaba is used to verify the feasibility of the method. The data contains more than 40,000 BM nodes, 500 of which are randomly selected here, covering the resource utilization data of approximately 5,839 containers, totaling 6,339 sequences with a total of approximately 14,400 timestamps of resource utilization within 10 days.

[0096] Each sliding window contains a historical sequence X H and a future sequence X F As the input and output of the model. The resource utilization sequence of the first 6 days is used as the training set, the data of the middle 2 days is used as the validation set, and the data of the last 2 days is used as the test set to ensure that the model can fully learn historical information during the training process. The effect of the model was tested from three indicators: MAE (Mean Absolute Error), MAPE (Mean Absolute Percentage Error), and MSE (Mean Squared Error).

[0097] Table 1 and Figure 5 The test comparison results between the method of the embodiment of the present application and the benchmark method are given respectively.

[0098] Table 1 Performance comparison of this application method (HGL-RP) and the benchmark method

[0099]

[0100]

[0101] According to Table 1 and Figure 5 It can be seen that:

[0102] 1. The ARIMA model is based on a linear model, and its ability to simulate nonlinear relationships is relatively limited, and the errors of the three indicators are the largest.

[0103] 2.MLP, LSTM, and GRU models can handle nonlinear relationships and have significant improvements in error compared to the ARIMA model. The LSTM and GRU models have lower errors because they are better at handling time series data.

[0104] 3. The GraphGRU model introduces the DTW algorithm based on GRU to calculate the similarity between nodes, and adds GAT to model this relationship. However, DTW cannot capture the negative correlation between nodes and ignores the more fine-grained container sequence features. A linear layer is used to process the final hidden state. The performance improvement is small, only slightly better than the GRU model.

[0105] 4. Based on the existing methods, the method of this application comprehensively considers the positive and negative correlation characteristics between physical nodes, adds the resource utilization characteristics of the container level, introduces RGCN to model the complex relationship between sequences, and adopts the encoder-heterogeneous graph neural network-decoder architecture as a whole. These improvements make the method of this application surpass the comparison method in all three indicators. It is particularly outstanding in short-term prediction.

[0106] In addition, the test results of the method of this application in long-term prediction are as follows Figure 6 As shown in the figure, as the prediction time span increases, the performance of the model gradually approaches that of the traditional model. Although it gradually approaches the performance of the traditional model in long-term prediction, it can still maintain a low error level, further proving its superiority in microservice resource prediction tasks.

[0107] In summary, the method of this application outperforms the baseline method in all three indicators in the problem of microservice resource prediction due to the introduction of container-level multi-dimensional features, and the prediction error is reduced by about 58.12%. In the long-term prediction after the prediction window exceeds 90s, the method of this application also has better performance, and the error of long-term prediction is reduced by about 14.88%.

[0108] The present invention is not limited to the above-mentioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A fine-grained microservice resource prediction method based on heterogeneous graph neural network, characterized in that: include: The resource utilization sequences of the physical nodes and containers of the microservices are encoded to obtain an initial hidden state, which is input into a heterogeneous graph neural network for adjustment, and the adjusted hidden state is decoded to obtain a resource prediction result; the heterogeneous graph neural network is constructed according to a heterogeneous graph, and the heterogeneous graph uses the physical nodes and containers of the microservices as heterogeneous nodes, respectively, and uses the relationships between the physical nodes and containers as edges.

2. The fine-grained microservice resource prediction method based on heterogeneous graph neural network according to claim 1 is characterized in that: The relationship between the physical nodes and the containers is used as an edge, including an edge constructed based on the relationship between physical nodes, an edge constructed based on the relationship between physical nodes and containers, and an edge constructed based on the relationship between containers.

3. The fine-grained microservice resource prediction method based on heterogeneous graph neural network as described in claim 2 is characterized in that: The edges constructed based on the relationship between containers include edges constructed based on the relationship between containers under the same physical node and edges constructed based on the relationship between containers under different physical nodes.

4. The fine-grained microservice resource prediction method based on heterogeneous graph neural network according to claim 2 is characterized in that: Methods for building edges based on the relationship between physical nodes include: Calculate the correlation between resource utilization sequences between two physical nodes; A first type of edge is generated between two physical nodes whose correlation reaches or exceeds a first threshold.

5. The fine-grained microservice resource prediction method based on heterogeneous graph neural network according to claim 2 is characterized in that: Methods for building edges based on the relationship between physical nodes and containers include: The second type of edge is generated between a physical node and each container deployed on it.

6. The fine-grained microservice resource prediction method based on heterogeneous graph neural network according to claim 3 is characterized in that: Methods for building edges based on relationships between containers include: The third type of edge is generated between containers deployed on the same physical node, and the fourth type of edge is generated between containers deployed on different physical nodes.

7. The fine-grained microservice resource prediction method based on heterogeneous graph neural network according to claim 1 is characterized in that: The bidirectional recurrent gated neural network Bi-GRU is used to encode the resource utilization sequence of the physical nodes and containers of the microservices to obtain the hidden state.

8. The fine-grained microservice resource prediction method based on heterogeneous graph neural network according to claim 7 is characterized in that: The sliding window is used to select the historical resource utilization sequences of the physical nodes and containers of the microservices for encoding.

9. The fine-grained microservice resource prediction method based on heterogeneous graph neural network according to claim 7 is characterized in that: The sliding window selects historical resource utilization sequences of two dimensions, physical nodes and containers of the microservice, in a time-series continuous manner for encoding.

10. The fine-grained microservice resource prediction method based on heterogeneous graph neural network according to any one of claims 7 to 9, characterized in that: The resource utilization sequence of the physical nodes and containers of the microservices is encoded to obtain the initial hidden state, including: Compress the resource utilization sequences of the physical nodes and containers of the microservice into resource utilization information of the same size as the hidden layer state; Encoding the resource utilization information using a hidden layer to obtain a first hidden state; The resource utilization information residual is connected to the first hidden state to obtain an initial hidden state.

Images