CN114900441B

CN114900441B - Network performance prediction method, performance prediction model training method and related devices

Info

Publication number: CN114900441B
Application number: CN202210467299.6A
Authority: CN
Inventors: 崔勇; 王莫为; 惠林博; 梁铷; 刘振华
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2022-04-29
Filing date: 2022-04-29
Publication date: 2024-04-26
Anticipated expiration: 2042-04-29
Also published as: CN114900441A

Abstract

The application relates to the field of artificial intelligence, in particular to a network performance prediction method, a training method of a network performance prediction model and a related device. Wherein predicting comprises: acquiring a first state information of a heterogeneous relation diagram at the current moment aiming at a target network; processing the heterogeneous relation graph at the current moment and the first state information of the heterogeneous relation graph at the current moment through a network performance prediction model of the target network to obtain second state information of the heterogeneous relation graph at the current moment, wherein the second state information comprises performance indexes of the target network at the current moment; and determining the first state information of the heterogeneous relation diagram at the current moment at the next moment based on the second state information and the first state information of the heterogeneous relation diagram at the current moment. By adopting the embodiment of the application, the performance index of the network can be accurately predicted, and the performance index changing along with time can be obtained.

Description

Network performance prediction method, performance prediction model training method and related devices

Technical Field

The application relates to the field of artificial intelligence, in particular to a network performance prediction method, a performance prediction model training method and a related device.

Background

Network researchers and operators are both working on increasing network quality of service to meet various network service requirements, often relying on network performance models to achieve efficient network design, operation, and optimization. In particular, the network performance model provides a new means for performance monitoring of online scenes and performance optimization of offline scenes. Operators can acquire real network performance and monitor possible anomalies through the performance model in real time; a designer can evaluate whether performance under a given topology, configuration and flow meets design requirements through a performance model; researchers can guide the design and evaluation of optimization schemes such as flow scheduling, routing strategies and the like through performance models.

With the rapid development of machine learning technology, deep learning methods based on deep neural networks are applied to the fields of computer vision, machine translation and the like and have been greatly successful. Researchers have begun to focus on modeling network performance based on data-driven methods, and have made some progress, representative of which are: 1. based on the graph neural network, a network performance prediction model is constructed to determine the relationship between links and paths, so that the steady-state time delay of path stages under different topologies, routing schemes, link bandwidths and flows can be evaluated. 2. A network performance prediction model is constructed based on a variation self-encoder (variationalautoencoder, VAE) and long-short-term memory (longshorttermmemory, LSTM) to determine the relationship between path delay and traffic, so that the path delay and packet loss distribution can be predicted through a traffic matrix of a time sequence under a given topology.

These two network performance prediction models lack systematic expressive power and timing state modeling power. There are a variety of network entities in a real network that have a variety of configurations in either local or global scope. For example, the cache size of a switch affects the length of queues locally only, while routing policies affect the traffic distribution throughout the network. Network performance metrics are also presented on different network entities, such as path level delay and packet loss for wide area network scenarios, and flow level metrics such as flow completion time (flowcompletiontime, FCT) for data center scenarios. Because the existing network performance prediction model lacks systematic expression capability, the performance index of the network can only be predicted based on the relation between the link and the path, and the performance index determined in this way is inaccurate. In addition, the network performance is in a continuous change process, the buffer memory of the network device and the data packet being transmitted make the state of the next moment of the network depend on the state of the current moment, the change process of the network performance can be accurately reflected through continuous performance evaluation, so that some network performance anomalies (such as time delay increase and throughput decrease) can be identified, and therefore, the model needs to output the performance index changing with time. Because the existing network performance prediction model does not have time sequence state modeling capability, the existing network performance prediction model cannot output performance indexes changing along with time.

Disclosure of Invention

The embodiment of the application provides a network performance prediction method, a network performance prediction model training method and a related device, and the scheme of the application can overcome the problems.

In a first aspect, an embodiment of the present application provides a network performance prediction method, including:

Acquiring a heterogeneous relation graph at the current moment and first state information of the heterogeneous relation graph at the current moment aiming at a target network, wherein the heterogeneous relation graph at the current moment is used for representing a plurality of nodes participating in network operation in the target network at the current moment and relations among the plurality of nodes; processing the heterogeneous relation graph at the current moment and the first state information of the heterogeneous relation graph at the current moment through a network performance prediction model of the target network to obtain second state information of the heterogeneous relation graph at the current moment, wherein the second state information comprises performance indexes of the target network at the current moment; and determining the first state information of the heterogeneous relation diagram at the current moment at the next moment based on the second state information and the first state information of the heterogeneous relation diagram at the current moment.

In one possible embodiment, the plurality of nodes of the target network comprise: network devices, links, queues, and paths. In some scenarios, the plurality of nodes of the target network further comprise a flow. The network devices may include switches and servers.

By introducing a heterogeneous relationship diagram of the network, the network entity, the configuration and the relationship among the network entity and the configuration are characterized, so that the structural description of the network system is realized. The first state information of the heterogeneous relation diagram at the current moment is processed through the network performance prediction model, so that hidden state information with global state information can be obtained, and more accurate network performance indexes can be obtained based on the hidden state information with the global state information. The first state information of the heterogeneous relation diagram of the current moment of the target network at the next moment is predicted based on the hidden state information, and according to a recursion mode, the performance index of the target network at the next moment and the first state information of the heterogeneous relation diagram of the current moment at the next moment can be predicted based on the first state information of the heterogeneous relation diagram of the current moment, so that the time sequence expression of the performance index of the target network is realized, namely, the network performance prediction model of the target network has systematic expression capability.

In one possible embodiment, the network performance prediction model of the target network includes an encoding network, an information delivery network and a decoding network, and the processing, by the network performance prediction model of the target network, the heterogeneous relationship graph at the current time and the first state information of the heterogeneous relationship graph at the current time to obtain the second state information of the heterogeneous relationship graph at the current time includes:

Performing coding operation on the first state information of the heterogeneous relation diagram at the current moment through a coding network to obtain third state information of the heterogeneous relation diagram at the current moment; carrying out multi-round information transfer operation on the heterogeneous relation diagram at the current moment, the third state information of the heterogeneous relation diagram at the current moment and the hidden state information of the heterogeneous relation diagram at the previous moment through an information transfer network so as to obtain the hidden state information of the heterogeneous relation diagram at the current moment; and decoding the hidden state information of the heterogeneous relation diagram at the current moment through a decoding network to obtain second state information of the heterogeneous relation diagram at the current moment.

It should be understood that the heterogeneous relationship graph at the current time includes a plurality of nodes and side information among the plurality of nodes, and the encoding operation is performed on the first state information of the heterogeneous relationship graph at the current time through the encoding network, so as to obtain the third state information of the heterogeneous relationship graph at the current time, specifically, the encoding operation is performed on the first state information of each node in the heterogeneous relationship graph at the current time.

Because the dimensions of the first state information of different nodes in the heterogeneous relationship graph at the current moment may be different, in order to facilitate subsequent processing, the encoding network encodes the first state information of each node in the heterogeneous relationship graph at the current moment to obtain the third state information of each node, where the dimensions of the third state information of different nodes are the same. The information transmission network is used for carrying out multi-round information transmission, so that the hidden state information of each node comprises the state information of other nodes, and further, more accurate performance indexes can be obtained based on the hidden state information.

In one possible embodiment, the heterogeneous relationship graph at the current time includes a plurality of nodes of the target network and side information between the plurality of nodes, the side information between the plurality of nodes is used to represent a relationship between the plurality of nodes, and multiple rounds of information transfer operations are performed on the heterogeneous relationship graph at the current time, third state information of the heterogeneous relationship graph at the current time and hidden state information of the heterogeneous relationship graph at a previous time through the information transfer network to obtain the hidden state information of the heterogeneous relationship graph at the current time, where the operations include:

Obtaining fourth state information of the heterogeneous relation diagram at the current moment based on the third state information of the heterogeneous relation diagram at the current moment and the hidden state information of the heterogeneous relation diagram at the previous moment; the fourth state information of the heterogeneous relation graph at the current moment comprises fourth state information of each node in the heterogeneous relation graph at the current moment; when the ith round of information transfer is carried out, acquiring the state information S _i of the neighbor node of each node in the heterogeneous relation diagram at the current moment based on the side information among a plurality of nodes; i is an integer greater than 0; obtaining the state information S _i+1 of each node based on the state information S _i of the neighbor node of each node and the state information S _i of each node; when i=1, the state information S _i of each node in the heterogeneous relationship diagram at the current time is the fourth state information of each node; when the ith round of information transfer is the last round of information transfer in the multiple rounds of information transfer, the state information S _i+1 of each node is the hidden state information of each node, and the hidden state information of the heterogeneous relationship graph at the current moment comprises the hidden state information of each node in the heterogeneous relationship graph at the current moment.

In order to obtain more accurate performance indexes, before information transmission is performed, the hidden state information of the previous moment is required to be introduced, the hidden state information of the previous moment and the third state information of the heterogeneous relation diagram of the current moment are integrated together, fourth state information of the heterogeneous relation diagram of the current moment is obtained, and then information transmission is performed based on the fourth state information of the heterogeneous relation diagram of the current moment.

In one possible embodiment, obtaining the state information S _i+1 of each node based on the state information S _i of the neighboring node of each node and the state information S _i of each node includes:

Performing aggregation operation on the state information S _i of the neighbor node of each node to obtain aggregated state information of each node; converting the aggregated state information of each node to obtain converted state information of each node; the state information S _i of each node is updated based on the converted state information of each node to obtain the state information S _i+1 of each node.

Since the aggregate information of each node is the state information S _i for the neighboring node of each node, the meaning expressed by the aggregate information of each node is not consistent with the meaning expressed by the state information S _i of the node, and thus the aggregated information needs to be converted so that the meaning expressed by the converted state information is consistent with the meaning expressed by the state information S _i of the node. The state information S _i of the node is updated based on the converted state information, so that the obtained state information S _i+1 includes the state information of the neighboring node thereof. After multiple rounds of information transfer, information with global state is available for each node.

In one possible embodiment, the first state information of the current time heterogeneous relationship graph includes static information and current time dynamic information, the second state information further includes a state information change amount, the dynamic information change amount is a change amount between the next time dynamic information and the current time dynamic information, and determining the first state information of the current time heterogeneous relationship graph at the next time based on the second state information and the first state information of the current time heterogeneous relationship graph includes:

Determining dynamic information variation from second state information of the heterogeneous relation diagram at the current moment; and obtaining the first state information of the heterogeneous relation diagram at the current moment at the next moment according to the dynamic information variation quantity, the dynamic information at the current moment and the static information.

After the first state information of the heterogeneous relation diagram at the current moment at the next moment is obtained, the performance index of the target network at the next moment and the first state information of the heterogeneous relation diagram at the current moment at the next moment can be predicted according to the first state information of the heterogeneous relation diagram at the current moment at the next moment in a recursion mode, and even the performance index of the target network at each subsequent moment and the first state information of the heterogeneous relation diagram at the current moment at each subsequent moment are obtained.

In a second aspect, an embodiment of the present application provides a training method for a network performance prediction model, including:

acquiring a plurality of heterogeneous relationship graphs and fifth state information of each heterogeneous relationship graph in the plurality of heterogeneous relationship graphs, wherein the plurality of heterogeneous relationship graphs are acquired at different moments aiming at the same network, and each heterogeneous relationship graph is used for representing the relationship among a plurality of nodes participating in network operation and a plurality of nodes in the network at the moment of acquiring each heterogeneous relationship graph; and training the initial prediction model for multiple rounds based on the heterogeneous relation diagrams and fifth state information of each heterogeneous relation diagram in the heterogeneous relation diagrams so as to obtain a network performance prediction model.

In one possible embodiment, training the initial predictive model for multiple rounds based on the plurality of heterogeneous relationship graphs and fifth state information for each of the plurality of heterogeneous relationship graphs to obtain the network performance predictive model includes:

dividing the plurality of heterogeneous relationship maps into a plurality of groups, each group including at least one heterogeneous relationship map; when the p-th training is carried out, fifth state information and hidden state information H' _p of the p-th heterogeneous relation diagram in a plurality of groups are input into a performance prediction model M _p for processing to obtain at least one sixth state information corresponding to the p-th heterogeneous relation diagram, wherein the sixth state information comprises performance indexes and dynamic information variation, and p is an integer larger than 0; inputting the performance index and the dynamic information variation included in the at least one sixth state information into a loss function for calculation to obtain a loss value L _p; adjusting parameters of the performance prediction model M _p based on the loss value L _p to obtain a performance prediction model M _p+1; when p=1, the performance prediction model M _p is an initial prediction model, and the hidden state information H' _p is an empty set; when p is greater than 1, the hidden state information H' _p is obtained according to the last heterogeneous relationship diagram in the p-1 group, and when the p-th training is the last training, the performance prediction model M _p+1 is a network performance prediction model.

In one possible embodiment, the initial prediction model includes an initial encoding network, an initial information delivery network, and an initial decoding network, and the performance prediction model M _p includes an encoding network M _1p, an information delivery network M _2p, and a decoding network M _3p; each heterogeneous relation graph comprises a plurality of nodes and side information among the plurality of nodes; the fifth state information of each heterogeneous relationship graph includes fifth state information of each node in the heterogeneous relationship graph, the fifth state information of the p-th heterogeneous relationship graph and the hidden state information H' _p in the plurality of groups are input into the performance prediction model M _p for processing, so as to obtain at least one sixth state information corresponding to the p-th heterogeneous relationship graph, including:

the fifth state information of the j-th heterogeneous relation chart in the p-th group is processed as follows to obtain sixth state information corresponding to the j-th heterogeneous relation chart:

Processing the fifth state information of each node in the j-th heterogeneous relation diagram through the coding network M _1p to obtain seventh state information of the j-th heterogeneous relation diagram; based on the side information among the plurality of nodes included in the jth heterogeneous relation diagram, carrying out multi-round information transfer operation on the seventh state information of the jth heterogeneous relation diagram and the hidden state information H _j of the jth-1 heterogeneous relation diagram through an information transfer network M _2p so as to obtain the hidden state information H _j+1 of the jth heterogeneous relation diagram; decoding the hidden state information H _j+1 of the j-th heterogeneous relation diagram through a decoding network M _3p to obtain sixth state information corresponding to the j-th heterogeneous relation diagram; when j=1, the hidden state information H _j of the j-1 st heterogeneous relationship diagram is the hidden state information H' _p; when p=1, the encoding network M _1p is an initial encoding network, the information delivery network M _2p is an initial information delivery network, and the decoding network M _3p is an initial decoding network.

In one possible embodiment, based on the side information between the plurality of nodes included in the jth heterogeneous relationship graph, performing, through the information delivery network M _2p, multiple rounds of information delivery operations on the seventh state information of the jth heterogeneous relationship graph and the hidden state information H _j of the jth-1 heterogeneous relationship graph to obtain the hidden state information H _j+1 of the jth heterogeneous relationship graph, including:

Obtaining eighth state information of the j-th heterogeneous relation diagram based on the seventh state information of the j-th heterogeneous relation diagram and the hidden state information H _j of the j-1-th heterogeneous relation diagram; the eighth state information of the jth heterogeneous relation diagram comprises the eighth state information of each node in the jth heterogeneous relation diagram; when the kth round of information transfer is carried out, acquiring the state information S _k of the neighbor node of each node in the jth heterogeneous relation diagram based on the side information among the multiple nodes included in the jth heterogeneous relation diagram; the neighbor node of each node is a node with an edge between the neighbor node and the node; k is an integer greater than 0; obtaining the state information S _k+1 of each node based on the state information S _k of the neighbor node of each node and the state information S _k of each node in the j-th heterogeneous relation diagram; when k=1, the state information S _k of the neighbor node is eighth state information of the neighbor node, and the state information S _k of each node is eighth state information of each node; when the kth round of information transfer is the last round of information transfer in the multiple rounds of information transfer, the state information S _i+1 of each node is the hidden state information of each node, and the hidden state information H _j+1 of the jth heterogeneous relationship graph includes the hidden state information of each node in the jth heterogeneous relationship graph.

In one possible embodiment, obtaining the state information S _k+1 of each node based on the state information S _k of the neighbor node of each node and the state information S _k of each node in the jth heterogeneous relationship graph includes:

Performing aggregation operation on the state information S _k of the neighbor node of each node to obtain aggregated state information of each node; converting the aggregated state information of each node to obtain converted state information of each node; the state information of each node is updated based on the converted state information of each node to obtain the state information of each node S _k+1.

In one possible embodiment, the plurality of nodes of the network comprise: network devices, links, queues, and paths. In some scenarios, the plurality of nodes of the target network further comprise a flow. The network devices may include switches and servers.

It will be appreciated that the method according to the second aspect may be used to train a network performance prediction model according to the first aspect, and that the training process is consistent with the process of predicting performance indicators using the network performance prediction model, and therefore the advantages of the second aspect will be described with reference to the advantages of the first aspect and will not be described.

In a third aspect, an embodiment of the present application provides a prediction apparatus, including:

The acquisition unit is used for acquiring a heterogeneous relation graph at the current moment and first state information of the heterogeneous relation graph at the current moment aiming at the target network, wherein the heterogeneous relation graph at the current moment is used for representing the relation among a plurality of nodes participating in network operation in the target network at the current moment and a plurality of nodes;

the prediction unit is used for processing the heterogeneous relation graph at the current moment and the first state information of the heterogeneous relation graph at the current moment through a network performance prediction model of the target network to obtain second state information of the heterogeneous relation graph at the current moment, wherein the second state information comprises performance indexes of the target network at the current moment;

And the determining unit is used for determining the first state information of the heterogeneous relation diagram at the current moment at the next moment based on the second state information and the first state information of the heterogeneous relation diagram at the current moment.

In one possible embodiment, the network performance prediction model of the target network includes an encoding network, an information delivery network, and a decoding network, and the prediction unit is specifically configured to:

In one possible embodiment, the heterogeneous relationship graph at the current time includes a plurality of nodes of the target network and side information between the plurality of nodes, the side information between the plurality of nodes is used to represent a relationship between the plurality of nodes, and in performing, through the information delivery network, multiple rounds of information delivery operations on the heterogeneous relationship graph at the current time, third state information of the heterogeneous relationship graph at the current time, and hidden state information of the heterogeneous relationship graph at a previous time, so as to obtain an aspect of the hidden state information of the heterogeneous relationship graph at the current time, the prediction unit is specifically configured to:

In one possible embodiment, the prediction unit is specifically configured to, based on the state information S _i of the neighboring node of each node and the state information S _i of each node, obtain the state information S _i+1 of each node:

In one possible embodiment, the first state information of the heterogeneous relationship graph at the current time includes static information and dynamic information at the current time, the second state information further includes a state information change amount, the dynamic information change amount is a change amount between dynamic information at the next time and dynamic information at the current time, and the determining unit is specifically configured to include:

In a fourth aspect, an embodiment of the present application provides a training apparatus, including:

The system comprises an acquisition unit, a storage unit and a processing unit, wherein the acquisition unit is used for acquiring a plurality of heterogeneous relation diagrams and fifth state information of each heterogeneous relation diagram in the plurality of heterogeneous relation diagrams, the plurality of heterogeneous relation diagrams are acquired at different moments aiming at the same network, and each heterogeneous relation diagram is used for representing the relation between a plurality of nodes and a plurality of nodes which participate in network operation in the moment of acquiring each heterogeneous relation diagram;

The training unit is used for carrying out multi-round training on the initial prediction model based on the heterogeneous relation diagrams and the fifth state information of each heterogeneous relation diagram in the heterogeneous relation diagrams so as to obtain a network performance prediction model.

In a possible embodiment, the training unit is specifically configured to:

In one possible embodiment, the initial prediction model includes an initial encoding network, an initial information delivery network, and an initial decoding network, and the performance prediction model M _p includes an encoding network M _1p, an information delivery network M _2p, and a decoding network M _3p; each heterogeneous relation graph comprises a plurality of nodes and side information among the plurality of nodes; the fifth state information of each heterogeneous relationship graph includes fifth state information of each node in the heterogeneous relationship graph, and in the aspect that the fifth state information and the hidden state information H' _p of the p-th heterogeneous relationship graph in multiple groups are input into the performance prediction model M _p for processing, so as to obtain at least one sixth state information corresponding to the p-th heterogeneous relationship graph, the training unit is specifically configured to:

In one possible embodiment, in the aspect of performing, through the information delivery network M _2p, multiple rounds of information delivery operations on the seventh state information of the jth heterogeneous relationship graph and the hidden state information H _j of the jth-1 heterogeneous relationship graph based on the side information between the plurality of nodes included in the jth heterogeneous relationship graph, to obtain the hidden state information H _j+1 of the jth heterogeneous relationship graph, the training unit is specifically configured to:

In one possible embodiment, the training unit is specifically configured to, based on the state information S _k of the neighbor node of each node in the jth heterogeneous relationship graph and the state information S _k of each node, obtain the state information S _k+1 of each node:

In a fifth aspect, an embodiment of the present application further provides an electronic device, including a processor and a memory, where the processor is connected to the memory, where the memory is configured to store program code, and the processor is configured to invoke the program code to perform part or all of the method in the first aspect or the second aspect.

In a sixth aspect, an embodiment of the present application further provides a chip system, where the chip system is applied to an electronic device; the system-on-chip includes one or more interface circuits, and one or more processors; the interface circuit and the processor are interconnected through a circuit; the interface circuit is configured to receive a signal from a memory of the electronic device and to send the signal to the processor, the signal including computer instructions stored in the memory; when the processor executes the computer instructions, the electronic device performs part or all of the method of the first or second aspects.

In a seventh aspect, embodiments of the present application also provide a computer readable storage medium storing a computer program for execution by a processor to implement part or all of the method of the first or second aspects.

In an eighth aspect, embodiments of the present application also provide a computer program to be executed to implement part or all of the method of the first or second aspects.

These and other aspects of the application will be more readily apparent from the following description of the embodiments.

Drawings

In order to more clearly illustrate the embodiments of the application or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1a is a schematic diagram of a system architecture according to an embodiment of the present application;

FIG. 1b is a schematic diagram of another system architecture according to an embodiment of the present application;

Fig. 2 is a flow chart of a network performance prediction method according to an embodiment of the present application;

FIG. 3a is a schematic diagram of an isomerism relationship diagram according to an embodiment of the present application;

FIG. 3b is a schematic diagram of another heterogeneous relationship diagram according to an embodiment of the present application;

Fig. 4 is a specific flowchart of a step in a network performance prediction method according to an embodiment of the present application;

FIG. 5 is a schematic diagram of another heterogeneous relationship diagram according to an embodiment of the present application;

FIG. 6 is a schematic diagram of a process for using a network performance prediction model according to an embodiment of the present application;

FIG. 7 is a schematic diagram of a data center network according to an embodiment of the present application;

FIG. 8 is a schematic diagram of a heterogeneous relationship diagram of a data center network according to an embodiment of the present application;

FIG. 9 is a schematic flow chart of a training method for a network performance prediction model according to an embodiment of the present application;

fig. 10 is a schematic structural diagram of a prediction apparatus according to an embodiment of the present application;

FIG. 11 is a schematic structural diagram of a training device according to an embodiment of the present application;

Fig. 12 is a schematic structural diagram of an electronic device according to an embodiment of the present application;

Fig. 13 is a schematic structural diagram of another electronic device according to an embodiment of the present application.

Detailed Description

The terms "first," "second," "third," and "fourth" and the like in the description and in the claims and drawings are used for distinguishing between different objects and not necessarily for describing a particular sequential or chronological order. Furthermore, the terms "comprise" and "have," as well as any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those listed steps or elements but may include other steps or elements not listed or inherent to such process, method, article, or apparatus.

Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearances of such phrases in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. Those of skill in the art will explicitly and implicitly appreciate that the embodiments described herein may be combined with other embodiments.

"Plurality" means two or more. "and/or", describes an association relationship of an association object, and indicates that there may be three relationships, for example, a and/or B, and may indicate: a exists alone, A and B exist together, and B exists alone. The character "/" generally indicates that the context-dependent object is an "or" relationship.

Embodiments of the present application are described below with reference to the accompanying drawings.

Referring to fig. 1a, an embodiment of the present application provides a system architecture 100. The data collection device 160 is configured to collect a plurality of heterogeneous relationship graphs and a plurality of status information corresponding to the plurality of heterogeneous relationship graphs, and store the plurality of status information in the database 130. The heterogeneous relationship diagrams are determined at different times for the same network, and each heterogeneous relationship diagram in the heterogeneous relationship diagrams is used for representing the relationship among a plurality of nodes participating in the operation of the network at the time corresponding to the heterogeneous relationship diagram and the plurality of nodes. Each heterogeneous relationship graph comprises a plurality of nodes and side information among the plurality of nodes, wherein the side information among the plurality of nodes is used for representing the relationship among the plurality of nodes. Training device 120 generates network performance prediction model 101 based on the state information of the plurality of heterogeneous relationship maps maintained in database 130. How the training device 120 obtains the network performance prediction model 101 based on the plurality of heterogeneous relationship diagrams and the state information thereof can be referred to the relevant description of the subsequent embodiments, which will not be described herein. The network performance prediction model 101 can predict the performance index of the target network at the current time based on the heterogeneous relationship map of the current time of the input target network and the first state information thereof. Further, the network performance prediction model 101 may also predict the state information of the heterogeneous relationship map at the next time based on the heterogeneous relationship map at the current time of the input target network and the first state information thereof.

The network performance prediction model 101 obtained by training the device 120 may be applied in different systems or devices. In fig. 1a, the executing device 110 is configured with an I/O interface 112, and interacts with external devices, and a "user" may input data to the I/O interface 112 through the client device 140, where the data includes a heterogeneous relationship diagram of the current time of the target network and state information thereof.

The execution device 110 may call data, code, etc. in the data storage system 150, or may store data, instructions, etc. in the data storage system 150.

The computing module 111 processes the input data using the network performance prediction model 101, where the processing specifically includes:

Performing coding operation on the first state information of the heterogeneous relation diagram at the current moment to obtain third state information of the heterogeneous relation diagram at the current moment; carrying out multi-round information transfer operation based on the third state information of the heterogeneous relation diagram at the current moment and the hidden state information of the heterogeneous relation diagram at the previous moment to obtain the hidden state information of the heterogeneous relation diagram at the current moment; decoding the hidden state information of the heterogeneous relation diagram at the current moment to obtain second state information of the heterogeneous relation diagram at the current moment, wherein the second state information of the heterogeneous relation diagram at the current moment comprises performance indexes of the target network at the current moment, and determining the first state information of the heterogeneous relation diagram at the current moment at the next moment based on the second state information of the heterogeneous relation diagram at the current moment and the first state information of the heterogeneous relation diagram at the current moment.

Finally, the I/O interface 112 returns the processing results, such as the performance index of the target network at the current time, to the client device 140 for the user.

Further, the training device 120 may generate the corresponding network performance prediction model 101 for different networks based on heterogeneous relationship diagrams of the different networks and state information thereof to provide better results to the user.

In the case shown in fig. 1a, the user may manually specify the heterogeneous relationship diagram of the current time of the target network in the input execution device 110 and its state information, for example, to operate in an interface provided by the I/O interface 112. In another case, the client device 140 may automatically input the heterogeneous relationship diagram of the current time and the state information thereof of the target network to the I/O interface 112 and obtain the performance index of the target network at the current time and the state information of the heterogeneous relationship diagram at the next time, and if the client device 140 automatically inputs data, the user may set corresponding rights in the client device 140 if the user needs to obtain the authorization of the user. The user may view the results output by the execution device 110 at the client device 140, for example, the performance index of the target network at the current time and the state information of the heterogeneous relationship diagram at the next time, and the specific presentation form may be a specific manner of display, sound, action, and the like. The client device 140 may also be used as a data collection end to store the collected heterogeneous relationship diagram of the current time of the target network and its state information as training data into the database 130.

It should be noted that fig. 1a is only a schematic diagram of a system architecture provided by an embodiment of the present invention, and the positional relationship among devices, apparatuses, modules, etc. shown in the drawing is not limited in any way, for example, in fig. 1a, the data storage system 150 is an external memory with respect to the execution device 110, and in other cases, the data storage system 150 may be disposed in the execution device 110.

Referring to fig. 1b, fig. 1b is a schematic diagram of a system architecture according to an embodiment of the present application. As shown in fig. 1b, the system architecture comprises a prediction means 10 and a calculation means 20,

The prediction apparatus 10 may be a terminal device, such as a smart phone, a desktop computer, a notebook computer, a tablet computer, etc. The computing device 20 may be a server, a server cluster, a cloud server, or other form of device with computing capabilities.

The prediction apparatus 10 transmits a model acquisition request to the calculation apparatus 20. In one example, the model acquisition request includes an identification of the target network, and the computing device 20 stores a plurality of network performance prediction models corresponding to the plurality of networks. The computing device 20 obtains a network performance prediction model corresponding to the target network from the plurality of network performance prediction models according to the identification of the target network. In another embodiment, the model acquisition request carries training data, where the training data includes a plurality of different time heterogeneous relationship graphs and state information thereof acquired for the target network, and the computing device 20 obtains the network performance prediction model corresponding to the target network based on the plurality of different time heterogeneous relationship graphs and state information thereof. The computing device 20 sends a response message carrying the network performance prediction model corresponding to the target network to the prediction device 10. The prediction apparatus 10 predicts the performance index of the target network at the current time and the first state information of the heterogeneous relationship map at the current time at the next time based on the heterogeneous relationship map at the current time of the target network and the first state information thereof based on the network performance prediction model corresponding to the target network.

In some embodiments, the prediction apparatus 10 has a strong computing power, and the prediction apparatus 10 obtains a network performance prediction model corresponding to the target network based on the heterogeneous relationship diagrams at a plurality of different moments and the state information thereof. The prediction apparatus 10 predicts the performance index at the current time of the target network and the state information of the heterogeneous relationship map at the next time based on the heterogeneous relationship map at the current time of the target network and the first state information thereof based on the network performance prediction model corresponding to the target network.

For a network performance prediction model corresponding to a target network, predicting a performance index of the target network at the current moment and first state information of the heterogeneous relation graph at the current moment at the next moment based on the heterogeneous relation graph at the current moment of the target network and the first state information thereof, specifically including:

It can be seen that by introducing a heterogeneous relationship diagram of the network, for characterizing network entities, configurations and relationships therebetween, a structural description of the network system is achieved. And carrying out information transfer on the state information of the heterogeneous relation graph based on the heterogeneous relation graph, so that the state information of each node in the heterogeneous relation graph contains the state information of other nodes in the heterogeneous relation graph, that is, the state information of each node in the heterogeneous relation graph is global state information, and carrying out performance index prediction based on the global state information, thereby being capable of predicting and obtaining accurate performance indexes. Based on the state information of the current moment, the state information of the next moment and the state information of the subsequent moment can be predicted, and then the performance index of the target network at the next moment and the performance index of the target network at the subsequent moment can be predicted.

Referring to fig. 2, fig. 2 is a flow chart of a network performance prediction method according to an embodiment of the present application. As shown in fig. 2, the method includes:

S201, a prediction device acquires a heterogeneous relation graph at the current moment and first state information of the heterogeneous relation graph at the current moment aiming at a target network, wherein the heterogeneous relation graph at the current moment is used for representing the relation among a plurality of nodes participating in network operation in the target network at the current moment.

Alternatively, the prediction means may be the execution device 110 in fig. 1a, or the prediction means 10 in fig. 1 b.

It should be noted that the nodes in the heterogeneous relationship graph include physical nodes and virtual nodes. Wherein a physical node represents a network node with configuration information including, but not limited to, switches with buffers, servers, links (ports), queues. Virtual nodes represent nodes related to network performance including, but not limited to, flows, paths.

The state information is not the same for different nodes. Status information such as switches includes, but is not limited to: cache size, cache management policies, etc. Status information of links (ports) includes, but is not limited to: bandwidth, propagation delay, queue scheduling policy, cache management control factors, etc. Status information of the queue includes, but is not limited to: scheduling weights, active queue management (activequeuemanagement, AQM) policies, parameters, etc. The state information of the path includes, but is not limited to: aggregate traffic size, hop count, latency, throughput, jitter, packet loss, etc. The state information of the flow includes, but is not limited to: start time, stream size, residual stream size, stream lifetime, etc.

The nodes comprised by the heterogeneous relationship graph at different times may not be identical. The heterogeneous relationship graph is used for representing a plurality of nodes in the network participating in the network operation at a certain moment and relationships among the plurality of nodes. Since the nodes participating in the network operation in the target network may be different at different times, and the relationships between the nodes participating in the network operation may also be different, the heterogeneous relationship diagrams at different times may be different.

The side information between the nodes is used to characterize the relationships between the nodes. Whether there is a relationship between two nodes is determined based on domain knowledge. If there is a relationship between two nodes, the two nodes are neighboring nodes to each other.

For example, assume that the target network includes switch S0 and switch S1, wherein switch S0 includes link L0 and link L1, and switch S1 includes link L2 and link L3. Link L0 is configured with queue Q0 and queue Q1, link L1 is configured with queue Q2 and queue Q3, link L2 is configured with queue Q4 and queue Q5, and link L3 is configured with queue Q6 and queue Q7.

At a first time, path P0 passes through queues Q0 and Q4, and flows F0 and F1 pass through path P0. Path P1 passes through queue Q1 and queue Q7, and flow F2 passes through path P1. Fig. 3a illustrates a diagram of the heterogeneous relationship at a first moment. As shown in fig. 3a, node S0 is associated with link L0, link L0 is associated with both queue Q0 and queue Q1, node S1 is associated with link L2, link L2 is associated with queue Q4, link L3 is associated with queue Q7, path P0 is associated with queue Q0 and queue Q4, path P1 is associated with queue Q1 and queue Q7, path P0 is associated with flow F0 and flow F1, and path P1 is associated with flow F2. The dashed lines in fig. 3a indicate that the corresponding node is not participating in the operation of the network at the first moment.

At a second time, path P0 passes through queues Q0, Q1 and Q4, and flow F0 is transmitted through path P0. Path P1 passes through queue Q2, queue Q3, and queue Q7, and flow F2 and flow F3 pass through path P1. Fig. 3b illustrates a diagram of the heterogeneous relationship at the second moment. As shown in fig. 3b, node S0 has a relationship with link L0 and link L1, link L0 has a relationship with queue Q0 and queue Q1, and link L1 has a relationship with queue Q2 and queue Q3. Node S1 is associated with link L2 and link L3, link L2 is associated with queue Q4, link L3 is associated with queue Q7, path P0 is associated with queue Q0, queue Q1 and queue Q4, path P1 is associated with queue Q2, queue Q3 and queue Q7, path P0 is associated with flow F0, and path P1 is associated with flow F2 and flow F3. The dashed lines in fig. 3b indicate that the corresponding node is not participating in the operation of the target network at the second moment.

As can be seen from fig. 3a and 3b, the heterogeneous relationship diagrams at different times may be different for the same network.

S202, the predicting device processes the heterogeneous relation graph at the current moment and the first state information of the heterogeneous relation graph at the current moment through a network performance predicting model of the target network to obtain second state information of the heterogeneous relation graph at the current moment, wherein the second state information comprises performance indexes of the target network at the current moment.

The network performance prediction model comprises an encoding network, an information transmission network and a decoding network. The specific process of S202 is shown in fig. 4, and includes:

S401, the prediction device performs coding operation on the first state information of the heterogeneous relation diagram at the current moment through a coding network so as to obtain the third state information of the heterogeneous relation diagram at the current moment.

Since the heterogeneous relationship graph includes a plurality of different types of nodes, the state information thereof is different for the different types of nodes. Accordingly, the dimensions of the state information may not be the same. In the present application, the state information is represented by a vector. For different types of nodes, the dimensions of vectors corresponding to the state information may be different, and in order to facilitate subsequent processing, the state information of each node in the heterogeneous relationship graph needs to be processed, so that the dimensions of the state information processed by all the nodes in the heterogeneous relationship graph are the same. When the vector is used for representing the state information, the vector corresponding to the state information of each node in the heterogeneous relation graph is processed, so that the dimensions of the vectors corresponding to the state information of all the nodes in the heterogeneous relation graph are the same. The specific processing mode is as follows:

The prediction device processes the first state information of each node in the heterogeneous relation graph at the current moment by using the coding network so as to obtain the third state information of each node in the heterogeneous relation graph at the current moment. The third state information of the heterogeneous relationship graph at the current time includes third state information of each node in the heterogeneous relationship graph at the current time.

Alternatively, the encoding network is implemented based on a neural network, such as a graph neural network.

S402, the prediction device performs multiple rounds of information transfer operation on the heterogeneous relation diagram at the current moment, the third state information of the heterogeneous relation diagram at the current moment and the hidden state information of the heterogeneous relation diagram at the previous moment through the information transfer network so as to obtain the hidden state information of the heterogeneous relation diagram at the current moment.

It should be noted that, in order to obtain a more accurate performance index, before performing multiple rounds of information transfer operations, the prediction device needs to process, such as splice, the third state information of the current heterogeneous relationship graph and the hidden state information of the previous heterogeneous relationship graph to obtain the fourth state information of the current heterogeneous relationship graph.

Specific ways of splicing the third state information of the heterogeneous relationship diagram at the current moment and the hidden state information of the heterogeneous relationship diagram at the previous moment include, but are not limited to:

And for any node X in the heterogeneous relation diagram at the current moment, if the node X exists in the heterogeneous relation diagram at the previous moment, obtaining fourth state information of the node X according to the third state information and the hidden state information of the node X. For example, the hidden state information of the node X is represented using a 3-dimensional vector, the third state information of the node X is represented using a 4-dimensional vector, and the fourth state information of the node X is represented by a 7-dimensional vector obtained from the 3-dimensional vector representing the hidden state information of the node X and the 4-dimensional vector representing the third state information of the node X. If the heterogeneous relation diagram at the previous moment does not exist the node X, determining the third state information of the node X as the fourth state information of the node X.

According to the mode, the fourth state information of each node in the heterogeneous relation diagram at the current moment can be obtained. The fourth state information of the current state heterogeneous relationship graph comprises third state information of each node in the current time heterogeneous relationship graph.

After the fourth state information of the heterogeneous relationship graph at the current moment is obtained, the state information of each node only contains the state information related to the node and does not contain the state information of other nodes in the heterogeneous relationship graph, so that in order to improve the accuracy of the predicted performance index, the prediction device performs multiple rounds of information transfer operation on the fourth state information of the heterogeneous relationship graph at the current moment based on the heterogeneous relationship graph at the current moment, so that the state information of each node of the heterogeneous relationship graph at the current moment contains the state information of other nodes. The heterogeneous relationship graph at the current moment comprises a plurality of nodes of the target network and side information among the plurality of nodes, wherein the side information among the plurality of nodes is used for representing the relationship among the plurality of nodes. The specific mode is as follows:

When the ith round of information is carried out, acquiring state information S _i of neighbor nodes of each node in the heterogeneous relation diagram based on side information among a plurality of nodes included in the heterogeneous relation diagram at the current moment; the neighbor node of each node is a node with an edge between the neighbor node and the node; obtaining the state information S _i+1 of each node based on the state information S _i of the neighbor node of each node and the state information S _i of each node; i is an integer greater than 0;

When i=1, each node S _i in the heterogeneous relationship diagram at the current time is fourth state information of each node; when the ith round of information transfer is the last round of information transfer in the multiple rounds of information transfer, the state information S _i+1 of each node is the hidden state information of the node, and the hidden state information of the heterogeneous relationship graph at the current moment comprises the hidden state information of each node in the heterogeneous relationship graph at the current moment.

Further, obtaining the state information S _i+1 of each node based on the state information S _i of the neighboring node of each node and the state information S _i of each node includes:

The aggregation operation is carried out on the state information S _i of the neighbor node of each node, and the aggregation operation can be realized through an aggregation function, and the aggregation function can realize the summation operation and the function of the cyclic neural network (recurrent neural network, RNN) so as to obtain the aggregated state information of each node; converting the aggregated state information of each node, which can be realized by a conversion function, so as to obtain the converted state information of each node; the state information S _i of each node is updated, for example, by an update function, based on the converted state information of each node, so as to obtain the state information S _i+1 of each node.

Since the aggregate information of each node is the state information S _i for the neighboring node of each node, the meaning expressed by the aggregate information of each node is not consistent with the meaning expressed by the state information S _i of the node, and thus the aggregated information needs to be converted so that the meaning expressed by the converted state information is consistent with the meaning expressed by the state information S _i of the node. Through the conversion operation, the characteristic information of the neighbor node of each node can be mapped into the same hidden space corresponding to each node, and the modeling capability of the graph neural network is not limited when the characteristic information of the neighbor node is uniformly operated by using the update function later.

It is noted here that the information transfer network comprises 3 parts for implementing the functions of the aggregation function, the conversion function and the update function, respectively. That is, the method is implemented by using a graph neural network when the state information S _i+1 of each node is obtained based on the state information S _i of the neighboring nodes of each node and the state information S _i of each node. It should be appreciated that the aggregation function, the conversion function, and the update function may be implemented based on three neural networks, respectively.

It should be understood that the heterogeneous relationship graph may include the same type of nodes, for example, the plurality of nodes of the corresponding switch in the heterogeneous relationship graph are the same type of nodes, and the plurality of nodes of the corresponding path are the same type of nodes. For multiple nodes of the same type, the aggregation function, the conversion function and the update function are the same when doing the above operation, i.e. the graph neural network used is the same.

For example, assume that the heterogeneous relationship map at the present time includes 7 nodes as shown in fig. 5. The 3 rounds of information transfer operation is performed on the third state information of the heterogeneous relationship diagram shown in fig. 5, and specifically includes:

In the 1 st information transfer operation, for the node a, the neighboring nodes of the node a include nodes B and D, the prediction device obtains the state information S ₁ of the node a according to the third state information of the node B, the node D, and the node a, and the state information S ₁ of the node a includes the third state information of the nodes A, B and D. For node B, the neighboring nodes of node B include nodes a and C, and the prediction means obtains the state information S ₁ of node B from the third state information of nodes B, C and a, where the state information S ₁ of node B includes the third state information of nodes A, B and C. For the node C, the neighboring nodes of the node C include the nodes B, D, E and G, the prediction device obtains the state information S ₁ of the node C according to the third state information of the nodes B, C, D, E and G, and the state information S ₁ of the node C includes the third state information of the nodes B, C, D, E and G. For the node D, the neighboring nodes of the node D include the nodes A, C and G, the prediction device obtains the state information S ₁ of the node D according to the third state information of the nodes A, C, D and G, and the state information S ₁ of the node D includes the third state information of the nodes A, C, D and G. For the node E, the neighboring nodes of the node E include nodes F and C, the prediction device obtains the state information S ₁ of the node E according to the third state information of the node C, the node E and the node F, and the state information S ₁ of the node E includes the third state information of the nodes E, F and C. For the node F, the neighboring nodes of the node F include the node E, and the prediction device obtains the state information S ₁ of the node F according to the third state information of the node E and the node F, where the state information S ₁ of the node F includes the third state information of the node E and the node F. For the node G, the neighboring nodes of the node G include nodes D and C, the prediction device obtains the state information S ₁ of the node G according to the third state information of the node C, the node D and the node G, and the state information S ₁ of the node G includes the third state information of the nodes C, G and D.

In the case of the round 2 information transfer operation, for the node a, the prediction apparatus obtains the state information S ₂ of the node a from the state information S ₁ of the node B, the node D, and the node a, the state information S ₂ of the node a includes the state information S ₁ of the nodes A, B and D, that is, the state information S ₂ of the node a includes the third state information of A, B, C, D and G. For node B, the prediction means obtains the state information S ₂ of node B from the state information S ₁ of node B, node C and node a, the state information S ₂ of node B including the state information S ₁ of nodes A, B and C, that is, the state information S ₂ of node B including the third state information of A, B, C, D, E and G. For node C, the prediction means obtains the state information S ₂ of node C from the state information S ₁ of nodes B, C, D, E and G, the state information S ₂ of node C including the state information S ₁ of nodes B, C, D, E and G, that is, the state information S ₂ of node C including the third state information of a, B, C, D, E, F and G. For node D, the prediction means obtains the state information S ₂ of node D from the state information S ₁ of node A, C, D and G, the state information S ₂ of node D including the state information S ₁ of node A, C and G, that is, the state information S ₂ of node D including the third state information of a, B, D, E and G. For node E, the prediction apparatus obtains the state information S ₂ of node E according to the state information S ₁ of node C, node E and node F, where the state information S ₂ of node E includes the state information S ₁ of node F and node C, that is, the state information S ₂ of node E includes the third state information of A, B, C, E and G. For the node F, the neighboring nodes of the node F include the node E, the prediction device obtains the state information S ₂ of the node F according to the state information S ₁ of the node E and the node F, the state information S ₂ of the node F includes the state information S ₁ of the node E and the node F, that is, the state information S ₂ of the node F includes the third state information of the nodes C, E and F. For the node G, the prediction apparatus obtains the state information S ₂ of the node G according to the state information S ₁ of the node C, the node D and the node G, the state information S ₂ of the node G includes the state information S ₂ of the nodes C, D and G, that is, the state information S ₁ of the node G includes the third state information of the nodes A, B, C, D, E and G.

In the 3 rd round of information transfer operation, for the node a, the prediction device obtains the state information S ₃ of the node a according to the state information S ₂ of the node B, the node D and the node a, and the state information S ₃ of the node a includes the state information S ₂ of the nodes A, B and D. For node B, the prediction means obtains the state information S ₃ of node B from the state information S ₂ of node B, node C and node a, and the state information S ₃ of node B includes the state information S ₂ of nodes A, B and C. For node C, the prediction means obtains the state information S ₃ of node C from the state information S ₂ of nodes B, C, D, E and G, and the state information S ₃ of node C contains the state information S ₂ of nodes B, C, D, E and G. For node D, the prediction apparatus obtains state information S ₃ of node D from state information S ₂ of nodes A, C, D and G, and state information S ₃ of node D includes state information S ₂ of nodes A, C and G. For node E, the prediction apparatus obtains state information S ₃ of node E from state information S ₂ of node C, node E, and node F, and state information S ₃ of node E includes state information S ₂ of node F and node C. For the node F, the neighboring nodes of the node F include the node E, and the prediction device obtains the state information S ₃ of the node F according to the state information S ₂ of the node E and the node F, where the state information S ₃ of the node F includes the state information S ₂ of the node E and the node F. For the node G, the neighboring nodes of the node G include nodes D and G, the prediction device obtains the state information S ₃ of the node G according to the state information S ₂ of the node C, the node D and the node G, and the state information S ₃ of the node G includes the state information S ₂ of the nodes C, D and G. At this time, the state information S ₃ of each node in the heterogeneous relationship diagram at the current time includes the third state information of the other nodes.

S403, the prediction device decodes the hidden state information of the heterogeneous relation diagram at the current moment through the decoding network to obtain second state information of the heterogeneous relation diagram at the current moment.

The second state information of the heterogeneous relation diagram at the current moment comprises performance indexes of the target network at the current moment.

Alternatively, the decoding network may be implemented by a neural network, such as a graph neural network.

S203, the prediction device determines the first state information of the heterogeneous relation diagram at the current moment at the next moment based on the second state information and the first state information of the heterogeneous relation diagram at the current moment.

Specifically, the first state information of the heterogeneous relation graph at the current moment comprises static information and dynamic information at the current moment, the second state information further comprises state information variable quantity, the dynamic information variable quantity is the variable quantity between the dynamic information at the next moment and the dynamic information at the current moment, and the prediction device determines the dynamic information variable quantity from the second state information of the heterogeneous relation graph at the current moment; and obtaining the first state information of the heterogeneous relation diagram at the current moment at the next moment according to the dynamic information variation quantity, the dynamic information at the current moment and the static information.

After obtaining the first state information of the heterogeneous relation diagram at the current moment at the next moment, the prediction device may recursively process the first state information of the heterogeneous relation diagram at the current moment at the next moment based on the modes of S202 and S203 so as to predict the performance index of the target network at the next moment and the first state information of the heterogeneous relation diagram at the current moment at the next moment, and even the performance index of the target network at each subsequent moment and the first state information of the heterogeneous relation diagram at the current moment at each subsequent moment.

It should be understood that the static information is information that does not change with time in the target network, such as total buffer size of the switch, link bandwidth, propagation delay, number of path hops, flow start time, etc. The dynamic information is time-varying information in the target network, such as path traffic size, residual flow size, flow lifetime, etc. If the configuration in the network changes over time, it should also be considered as dynamic information such as active queue management parameters, queue weights, etc.

It can be seen that in the scheme of the application, the heterogeneous relation diagram of the network is introduced to characterize network entities, configurations and relations among the network entities, so that structural description of the network system is realized. And carrying out information transfer on the state information of the heterogeneous relation graph based on the heterogeneous relation graph, so that the state information of each node in the heterogeneous relation graph contains the state information of other nodes in the heterogeneous relation graph, that is, the state information of each node in the heterogeneous relation graph is global state information, and carrying out performance index prediction based on the global state information, thereby being capable of predicting and obtaining accurate performance indexes. The first state information of the isomerism relation graph obtained at the current moment in the next moment can be predicted based on the first state information of the isomerism relation graph obtained at the current moment, and according to a recursion mode, the performance index of the target network at the next moment and the first state information of the isomerism relation graph at the current moment in the next moment can be predicted based on the first state information of the isomerism relation graph at the current moment in the next moment, so that the time sequence expression of the performance index of the target network is realized.

The following describes a method for predicting a performance index of a target network at a current moment by using a network performance prediction model of the target network with reference to the accompanying drawings.

As shown in fig. 6, the network performance prediction model includes an encoding network, an information delivery network, and a decoding network. The state information of the current moment is the first state information of the heterogeneous relation diagram of the current moment acquired by the pointer to the target network. The heterogeneous relationship graph at the current moment comprises a plurality of nodes and side information among the nodes, and the first state information of the heterogeneous relationship graph at the current moment comprises the first state information of each node in the heterogeneous relationship graph. Because the dimensions of the first state information of different nodes in the heterogeneous relation diagram may be inconsistent, in order to facilitate subsequent processing, the first state information of each node of the heterogeneous relation diagram at the current moment is input into the coding network for processing so as to obtain the third state information of each node; the dimensions of the third state information of all nodes in the heterogeneous relation graph at the current moment are the same.

The specific process of obtaining the fourth state information of the heterogeneous relationship diagram at the current time based on the hidden state information at the previous time and the third state information of the heterogeneous relationship diagram at the current time can be referred to in the related description of S402, which is not described herein. And carrying out N rounds of prediction on the fourth state information of the relation graph at the current moment through the information transfer model to obtain the hidden state information of the heterogeneous relation graph at the current moment, so that the hidden state information of each node in the heterogeneous relation graph at the current moment not only comprises the first state information of the node, but also comprises the first state information of other nodes in the heterogeneous relation graph at the current moment. The specific process may be referred to in the related description of step S402, which will not be described here.

And decoding the hidden state information of the heterogeneous relation diagram at the current moment through a decoding network to obtain second state information of the heterogeneous relation diagram at the current moment. The first state information of the heterogeneous relation diagram at the current moment comprises a performance index and dynamic information variation of the target network at the current moment. The first state information of the heterogeneous relation graph at the current moment comprises static information and dynamic information of the heterogeneous relation graph at the current moment, and the dynamic information change amount is the change amount of the dynamic information of the heterogeneous relation graph at the current moment relative to the dynamic information of the heterogeneous relation graph at the current moment. The dynamic information of the heterogeneous relationship graph at the current moment at the next moment can be obtained based on the dynamic information and the dynamic information variation quantity of the heterogeneous relationship graph at the current moment, and the first state information of the heterogeneous relationship graph at the current moment at the next moment can be obtained based on the static information of the heterogeneous relationship graph at the current moment and the dynamic information of the heterogeneous relationship graph at the next moment.

In one specific example, FIG. 7 is a schematic diagram of a data center network having a leaf-spine topology. As shown in fig. 7, the target network includes a switch S0, a switch S1, a switch S2, a switch S3, a switch S4, a switch S5, a server H0, a server H1, a server H2, a server H3, a server H4, a server H5, a server H6, and a server H7. The buffer memory of the switch is shared among all ports, namely all ports of the switch are shared buffer memory, the queue of each port is managed by a scheduling policy, the flow is sent from a source server to a destination server, a data center transmission control protocol (datacentertransmissioncontrolprotocol, DCTCP) is operated on the server for congestion control, an explicit congestion notification (explicit congestion notification, ECN) mechanism is arranged on the queue of each switch, and a network performance prediction model needs to predict the completion time of different flows (starting time and size) under different configurations.

As shown in fig. 7, the flow F0 and the flow F1 are configured to be transmitted from the queue Q0 of the server H0 to the server H5 through the path P0, the flow F2 is configured to be transmitted from the queue Q1 of the server H0 to the server H4 through the path P1, and the flow F3 is configured to be transmitted from the queue Q1 of the server H1 to the server H5 through the path P2, wherein the path P0 is through the queue Q0 of the link L0 of the switch S0, the queue Q0 of the link L0 of the switch S4, and the queue Q0 of the link L0 of the switch S2, and the path P1 is through the queue Q1 of the link L0 of the switch S0, the queue Q1 of the link L0 of the switch S4, and the queue Q1 of the link L1 of the switch S2; path P2 passes through queue Q1 of link L0 of switch S0, queue Q1 of link L0 of switch S4, and queue Q1 of link L0 of switch S2.

Based on the above information, a heterogeneous relationship diagram of the data center network can be constructed, as shown in fig. 8. To accurately model different levels of network characteristic information, 5 different types of nodes are used to describe the network, including switch nodes S0, S2, and S4, port (link) nodes L0 and L1, queue nodes Q0 and Q1, path nodes P1, P2, and P3, and flow nodes F0, F1, F2, and F3. The connection between nodes indicates that there is a direct impact relationship between the nodes. For example, the queue node has an ECN threshold and priority weight, the port node has a queue policy, and the queue node and the port node are connected together. The switch node has a control variable which shares the cache and acts on the port node, and the switch node needs to be connected with the port node. The flow node and the path node it passes through are connected together. Thus, local relationships can be described by disjoint edges. To establish a global relationship, such as a route, a path node needs to be connected to the queue node it passes through. Note that not all physical nodes in a data center network have to be instantiated, as some of the devices may not flow through, where they have no impact on performance of interest.

The first state information of the heterogeneous relationship diagram shown in fig. 8 is obtained, and includes first state information of each switch node, first state information of each link (port) node, first state information of each queue node, first state information of each path node, and first state information of each flow node. In this example, the switch node, link node, and queue node each contain only static information, and the path and flow nodes contain both static and dynamic information. In particular, the first state information of the switch node includes, but is not limited to, a cache size and a cache policy; the first state information of the link node includes, but is not limited to, bandwidth, propagation delay, queue scheduling policy, cache management control factor; the first state information of the queue node includes, but is not limited to, queue scheduling weight, priority; the first state information of the path node comprises static information such as hop count and dynamic information such as path flow size; the first state information of the flow node includes static information including, but not limited to, a flow start time, a flow size, and dynamic information including, but not limited to, a remaining flow size, a flow lifetime. Since the dimensions of the first state information are different for different types of nodes, the dimensions of the third state information for all nodes in the heterogeneous relationship diagram shown in fig. 8 are identical by encoding the first state information for each node in the heterogeneous relationship diagram shown in fig. 8 to obtain the third state information for each node. In order to make the prediction of the performance index more accurate, the hidden information of the previous heterogeneous relationship graph and the third state information of the current heterogeneous relationship graph are processed, for example, spliced, so as to obtain the fourth state information of the current heterogeneous relationship graph, and the specific splicing process can refer to the related description of S402, which is not described herein.

In order to make the state information of each node in the heterogeneous relationship diagram shown in fig. 8 contain the first state information of other nodes, the fourth state information of the heterogeneous relationship diagram shown in fig. 8 is subjected to multiple rounds of information transfer so as to obtain the hidden state information of the heterogeneous relationship diagram shown in fig. 8. And for each path node, aggregating information of all flows passing through the path by using an addition mode, respectively converting the aggregated information into a hidden space of the state information of the path node by using all queue information passing through the RNN aggregated path node through a conversion function corresponding to the path node, splicing together fourth state information of the path node, converted flow information and converted queue information, and transmitting the spliced information to an updating function for updating the state information of the path node.

For each flow node, converting the fourth state information of the path where the flow is located into the hidden space of the state information of the flow node through a conversion function corresponding to the flow node, splicing the state information of the converted path node and the fourth state information of the flow node together, and transmitting the spliced information to an update function corresponding to the flow node for updating the state information of the path node.

For each queue node, aggregating fourth state information of all paths passing through the queue through an aggregation function corresponding to the queue node, converting the fourth state information of the link nodes related to the queue node and the aggregation state information of the path nodes into a hidden space where the queue node is located through a conversion function corresponding to the queue node, splicing the state information of the converted link nodes and the state information of the path nodes together, and transmitting the spliced information to an update function corresponding to the queue node for updating the state information of the queue node.

For each link (port) node, aggregating the fourth state information of all queue nodes related to the link through an aggregation function corresponding to the link node, converting the fourth state information of the switch node and the state information of the aggregated queue nodes into a hidden space where the link node is located through a conversion function corresponding to the link node, splicing the state information of the switch node and the state information of the queue nodes after conversion together, and transmitting the spliced information to an update function corresponding to the link node for updating the state information of the link node.

For each switch node, converting the fourth state information of the link node contained in the switch node into the hidden space where the switch node is located through the conversion function corresponding to the switch node, splicing the state information of the converted link node and the fourth state information of the switch node together, and transmitting the spliced information to the update function corresponding to the switch node for updating the state information of the switch node.

One round of information transfer can be completed according to the above. After the above information transfer process is performed in multiple rounds, the hidden state information of each node in the heterogeneous relationship diagram shown in fig. 8 can be obtained, and the hidden state information of each node in the heterogeneous relationship diagram shown in fig. 8 includes the first state information of other nodes.

Decoding is performed by decoding the hidden state information of the heterogeneous relationship diagram shown in fig. 8 of the network to obtain second state information of the heterogeneous relationship diagram shown in fig. 8. The second state information of the heterogeneous relationship diagram shown in fig. 8 includes performance indicators of the network shown in fig. 7, such as path delay, packet loss rate, and the like. The second state information of the heterogeneous relationship diagram shown in fig. 8 further includes a dynamic information change amount, which is a change amount of dynamic information of the heterogeneous relationship diagram at the current time with respect to dynamic information of the heterogeneous relationship diagram at the current time, such as a flow completion data amount, a flow survival time, and the like at two times. The first state information of the heterogeneous relationship diagram shown in fig. 8 includes static information and dynamic information of the heterogeneous relationship diagram at the current moment, wherein switch nodes, link nodes, queue nodes and path nodes all include only static information, and stream nodes include both static information and dynamic information, specifically, the first state information of the switch nodes includes, but is not limited to, cache size and cache policy; the first state information of the link node includes, but is not limited to, bandwidth, propagation delay, queue scheduling policy, cache management control factor; the first state information of the queue node includes, but is not limited to, queue scheduling weight, priority; the first state information of the path node includes, but is not limited to, hop count; the first state information of the flow node comprises static information such as flow starting time, flow size and the like and dynamic information such as residual flow size, flow survival time and the like. And obtaining the dynamic information of the heterogeneous relation graph at the current moment at the next moment according to the dynamic information and the dynamic information variation quantity of the heterogeneous relation graph at the current moment, and obtaining the first state information of the heterogeneous relation graph at the current moment at the next moment based on the dynamic information and the static information of the heterogeneous relation graph at the current moment at the next moment. When the first state information of the heterogeneous relation diagram at the current moment in the next moment is obtained, the performance index of the network shown in the following moment figure 7 can be predicted. It is noted that in this example, this performance index for FCTs is obtained indirectly through dynamic state variables, specifically, the network performance prediction model predicts FCTs by predicting the amount of data received between time instants, and updating the residual flow size and time-to-live at each time instant. Once the residual flow size is below zero, the network performance prediction model considers the flow to be complete and uses its lifetime as the predicted FCT.

In the above example, one possible use scenario is that the traffic of the target network is not described by a flow, but rather as aggregate traffic information of the passing paths, and the network performance prediction model needs to predict the path performance over time under different configurations and path traffic. At this time, other nodes are unchanged, and the path nodes simultaneously contain static information and dynamic information, wherein the static information comprises hop count and the like, and the dynamic information comprises path flow speed; the output path performance includes delay, packet loss, etc. In this scenario, when a network performance prediction model is used to predict and train the network performance prediction model, the heterogeneous relationship graph used includes only switch nodes, link (port) nodes, queue nodes, and path nodes, excluding flow nodes.

It should be understood that the scheme of the present application is not limited to the applied network field, and can be used in fine-grained performance prediction of a data center network, and also can be applied in path-level performance prediction of a backbone network.

The heterogeneous relationship diagram describes interaction relationship among different entities, and a similar method can be applied to a simulation process in the physical field, wherein entity constraint, number and relationship and an aggregation function may change, and a physical equation may need to be introduced to describe the change process of the entities with time, but the influence relationship among a plurality of entities can still be described by adopting a state transition process.

The structure of the network, configuration information of devices in the network, etc. may be different for different networks, and thus the network performance prediction model used is different for different networks in order to obtain accurate performance indexes. Thus, for different networks, a corresponding network performance prediction model needs to be trained.

The following describes how to train a network performance prediction model of a network.

Referring to fig. 9, fig. 9 is a flowchart of a network performance prediction model training method according to an embodiment of the present application.

As shown in fig. 9, the method includes:

And S901, the training device acquires a plurality of heterogeneous relation diagrams and fifth state information of each heterogeneous relation diagram.

The heterogeneous relation diagrams are acquired at different moments aiming at the same network, and each heterogeneous relation diagram comprises a plurality of nodes and side information among the nodes; the state information of each heterogeneous relationship graph includes first state information of each node in the heterogeneous relationship graph.

Alternatively, the training means may be the training device 120 of fig. 1a, or the prediction means 10 or the calculation means 20 of fig. 1 b.

It should be noted that the nodes in the heterogeneous relationship graph include physical nodes and virtual nodes. Wherein a physical node represents a network entity with configuration information including, but not limited to, switches, servers, links (ports), queues. The virtual nodes identify entities related to network performance including, but not limited to, flows, paths.

The nodes comprised by the heterogeneous relationship graph at different times may not be identical. The heterogeneous relationship graph is used to characterize entities in the network that are involved in the operation of the network at a certain moment in time and the relationships between these entities. Since the entities in the network that participate in the operation of the network may not be the same at different times, the heterogeneous relationship diagrams at different times may not be the same. The side information between entities is the side information between the nodes.

It should be noted that, the relevant features of the heterogeneous relationship diagram may be referred to in the relevant description of the embodiment shown in fig. 2, and will not be described herein.

S902, the training device carries out multi-round training on the initial prediction model based on the heterogeneous relation diagrams and fifth state information of each heterogeneous relation diagram in the heterogeneous relation diagrams so as to obtain a network performance prediction model.

In one possible embodiment, the predicting device trains the initial prediction model for multiple rounds based on the multiple heterogeneous relationship graphs and the fifth state information of each heterogeneous relationship graph in the multiple heterogeneous relationship graphs to obtain a network performance prediction model, and specifically includes:

In one example, the prediction means divides the plurality of heterogeneous relationship maps into a plurality of groups, each group comprising at least one heterogeneous relationship map; then, carrying out multi-round training by utilizing a plurality of groups of heterogeneous relation graphs and fifth state information thereof, and when carrying out the 1 st round training, inputting fifth state information and hidden state information H' ₁ of the 1 st group of heterogeneous relation graphs in a plurality of groups into a performance prediction model M ₁ (namely an initial prediction model) for processing to obtain at least one sixth state information corresponding to the 1 st group of heterogeneous relation graphs, wherein the hidden state information is an empty set, and the sixth state information comprises performance indexes and dynamic information variation; inputting the performance index and the dynamic information variation quantity which are included in at least one sixth state information corresponding to the heterogeneous relation diagram of the 1 st group into a loss function for calculation to obtain a loss value L ₁; adjusting parameters of the performance prediction model M ₁ based on the loss value L ₁ to obtain a performance prediction model M ₂; … …; when the p-th training is carried out, fifth state information and hidden state information H' _p of the p-th heterogeneous relation diagram in a plurality of groups are input into a performance prediction model M _p for processing to obtain at least one sixth state information corresponding to the p-th heterogeneous relation diagram, wherein the sixth state information comprises performance indexes and dynamic information variation, and p is an integer larger than 0; inputting the performance index and the dynamic information variation included in the at least one sixth state information into a loss function for calculation to obtain a loss value L _p; adjusting parameters of the performance prediction model M _p based on the loss value L _p to obtain a performance prediction model M _p+1; when p is greater than 1, the hidden state information H' _p is obtained according to the last heterogeneous relationship diagram in the p-1 group, and the specific manner is described in the following related description, and after the last training round, the obtained performance prediction model M is the network performance prediction model.

Wherein adjusting the parameters of the performance prediction model M _p based on the loss value L _p may adjust the parameters of the performance prediction model M _p in a gradient-decreasing manner based on the loss value L _p.

In another example, after each obtaining the preset number of heterogeneous relationship diagrams and the fifth state information thereof, the prediction device performs a round of training based on the preset number of heterogeneous relationship diagrams and the fifth state information thereof, and a specific training process may be referred to in the above description.

Further, the initial prediction model includes an initial encoding network, an initial information delivery network, and an initial decoding network, and the corresponding performance prediction model M _p includes an encoding network M _1p, an information delivery network M _2p, and a decoding network M _3p. The fifth state information of each heterogeneous relationship graph includes fifth state information of each node in the heterogeneous relationship graph.

The fifth state information of each heterogeneous relation chart in the p-th group is processed as follows to obtain sixth state information corresponding to each heterogeneous relation chart:

Processing the fifth state information of each node in the 1 st heterogeneous relation diagram in the p-th group through the coding network M _1p to obtain seventh state information of the 1 st heterogeneous relation diagram; based on the side information among the plurality of nodes included in the 1 st heterogeneous relation diagram, carrying out multi-round information transfer operation on the seventh state information of the 1 st heterogeneous relation diagram and the hidden state information H ₁ of the 0 st heterogeneous relation diagram through an information transfer network M _2p so as to obtain the hidden state information H ₂ of the 1 st heterogeneous relation diagram; decoding the hidden state information H ₂ of the 1 st heterogeneous relation diagram through a decoding network M _3p to obtain sixth state information corresponding to the 1 st heterogeneous relation diagram; wherein the hidden state information H ₁ of the 0 th heterogeneous relationship diagram is hidden state information H' _p;

Processing fifth state information of each node in the 2 nd heterogeneous relation diagram in the p-th group through a coding network M _1p based on side information among a plurality of nodes included in the 2 nd heterogeneous relation diagram so as to obtain seventh state information of the 2 nd heterogeneous relation diagram; carrying out multi-round information transfer operation on the seventh state information of the 2 nd heterogeneous relation diagram and the hidden state information H ₂ of the 1 st heterogeneous relation diagram through an information transfer network M _2p so as to obtain the hidden state information H ₃ of the 2 nd heterogeneous relation diagram; decoding the hidden state information H ₃ of the 2 nd heterogeneous relation diagram through a decoding network M _3p to obtain sixth state information corresponding to the 2 nd heterogeneous relation diagram;

……

Processing fifth state information of each node in the j-th heterogeneous relation diagram through a coding network M _1p based on side information among a plurality of nodes included in the j-th heterogeneous relation diagram so as to obtain seventh state information of the j-th heterogeneous relation diagram; carrying out multi-round information transfer operation on the seventh state information of the j-th heterogeneous relation diagram and the hidden state information H _j of the j-1-th heterogeneous relation diagram through an information transfer network M _2p so as to obtain hidden state information H _j+1 of the j-th heterogeneous relation diagram; and decoding the hidden state information H _j+1 of the j-th heterogeneous relation diagram through the decoding network N _3j to obtain sixth state information corresponding to the j-th heterogeneous relation diagram.

In the above manner, the sixth state information of each of the heterogeneous relationship maps in the p-th group can be obtained. When p=1, the encoding network M _1p is an initial encoding network, the information delivery network M _2p is an initial information delivery network, and the decoding network M _3p is an initial decoding network.

In one possible embodiment, based on the side information between the plurality of nodes included in the jth heterogeneous relationship diagram, performing, through the information delivery network M _2p, multiple rounds of information delivery operations on the seventh state information of the jth heterogeneous relationship diagram and the hidden state information H _j of the jth-1 heterogeneous relationship diagram to obtain the hidden state information H _j+1 of the jth heterogeneous relationship diagram, where the operations specifically include:

Obtaining eighth state information of the j-th heterogeneous relation diagram based on the seventh state information of the j-th heterogeneous relation diagram and the hidden state information H _j of the j-1-th heterogeneous relation diagram; the eighth state information of the jth heterogeneous relation diagram comprises the eighth state information of each node in the jth heterogeneous relation diagram;

When the 1 st round of information transfer is carried out, determining neighbor nodes of each node in the j-th heterogeneous relation diagram based on side information among a plurality of nodes of the j-th heterogeneous relation diagram; obtaining the state information S ₂ of each node according to the state information S ₁ of each node and the state information S ₁ of the neighbor nodes thereof; the state information S ₂ of each node includes state information S ₁ of neighboring nodes of the node; the state information S ₁ is eighth state information;

When the 2 nd round of information transfer is carried out, determining neighbor nodes of each node in the j-th heterogeneous relation diagram based on side information among a plurality of nodes of the j-th heterogeneous relation diagram; obtaining the state information S ₃ of each node according to the state information S ₂ of each node and the state information S ₂ of the neighbor nodes thereof; the state information S ₃ of each node includes state information S ₂ of neighboring nodes of the node;

……

When the kth round of information transfer is carried out, acquiring the state information S _k of the neighbor node of each node in the jth heterogeneous relation diagram based on the jth heterogeneous relation diagram; and obtaining the state information S _k+1 of each node based on the state information S _k of the neighbor node of each node and the state information S _k of each node in the j-th heterogeneous relation diagram.

The eighth state information of the jth heterogeneous relationship graph includes eighth state information of each node in the jth heterogeneous relationship graph, and when the kth round of information is transmitted as the last round of information in the multiple rounds of information transmission, the state information S _i+1 of each node is the hidden state information of the node, and the hidden state information H _j+1 of the jth heterogeneous relationship graph includes the hidden state information of each node in the jth heterogeneous relationship graph.

It should be understood that, since the neighboring node of each node in the 1 st heterogeneous relationship graph has been determined based on the side information between the plurality of nodes of the j-th heterogeneous relationship graph at the time of 1 st round of information transfer, the operation of "determining the neighboring node of each node in the j-th heterogeneous relationship graph based on the side information between the plurality of nodes of the j-th heterogeneous relationship graph" may not be performed at the time of subsequent information transfer.

The aggregation operation may be implemented by an aggregation function, the conversion operation may be implemented by a conversion function, and the update operation may be implemented by an update function. The aggregation function, the conversion function, and the update function may each be implemented by a neural network, such as a graph neural network. For the heterogeneous relationship graph, nodes of the same type are included. For nodes of the same type, the corresponding aggregate functions may be the same, the corresponding transfer functions may be the same, and the corresponding update functions may be the same.

In one possible embodiment, the hidden state information H _j of the j-1 th heterogeneous relationship graph includes hidden state information of each node of the j-1 th heterogeneous relationship graph; the obtaining the eighth state information of the jth heterogeneous relationship diagram based on the seventh state information of the jth heterogeneous relationship diagram and the hidden state information H _j of the jth-1 heterogeneous relationship diagram may specifically include:

And for any node Y in the heterogeneous relation diagram of the jth heterogeneous relation diagram, if the jth-1 heterogeneous relation diagram exists in the node Y, obtaining eighth state information of the node Y according to seventh state information of the node Y and hidden state information of the node Y. For example, the hidden state information of the node Y is represented using a 2-dimensional vector, the seventh state information of the node Y is represented using a 3-dimensional vector, and then the eighth state information of the node Y is represented by a 4-dimensional vector obtained from the 2-dimensional vector representing the hidden state information of the node Y and the 3-dimensional vector representing the seventh state information of the node Y. If the j-1-th heterogeneous relation diagram does not exist the node Y, the seventh state information of the node Y is determined to be the eighth state information of the node Y.

In the above manner, the eighth state information of each node in the jth heterogeneous relationship graph can be obtained. The eighth state information of the jth heterogeneous relationship graph includes eighth state information of each node in the jth heterogeneous relationship graph.

According to the training method, a network performance prediction model of the network can be obtained.

Referring to fig. 10, fig. 10 is a schematic structural diagram of a prediction apparatus according to an embodiment of the present application. As shown in fig. 10, the prediction apparatus 1000 includes:

An obtaining unit 1001, configured to obtain, for a target network, a current time heterogeneous relationship graph and first state information of the current time heterogeneous relationship graph, where the current time heterogeneous relationship graph is used to characterize a plurality of nodes and relationships between the plurality of nodes that participate in network operation in the current time target network;

A prediction unit 1002, configured to process, by using a network performance prediction model of the target network, the heterogeneous relationship graph at the current time and the first state information of the heterogeneous relationship graph at the current time, to obtain second state information of the heterogeneous relationship graph at the current time, where the second state information includes a performance index of the target network at the current time;

A determining unit 1003, configured to determine, based on the second state information and the first state information of the heterogeneous relationship diagram at the current time, the first state information of the heterogeneous relationship diagram at the next time at the current time.

In one possible embodiment, the network performance prediction model of the target network includes an encoding network, an information delivery network, and a decoding network, and the prediction unit 1002 is specifically configured to:

In one possible embodiment, the heterogeneous relationship graph at the current time includes a plurality of nodes of the target network and side information between the plurality of nodes, where the side information between the plurality of nodes is used to represent a relationship between the plurality of nodes, and in performing, through the information delivery network, multiple rounds of information delivery operations on the heterogeneous relationship graph at the current time, third state information of the heterogeneous relationship graph at the current time, and hidden state information of the heterogeneous relationship graph at a previous time, to obtain an aspect of the hidden state information of the heterogeneous relationship graph at the current time, the prediction unit 1002 is specifically configured to:

In one possible embodiment, in obtaining the state information S _i+1 of each node based on the state information S _i of the neighboring node of each node and the state information S _i of each node, the prediction unit 1002 is specifically configured to:

In one possible embodiment, the first state information of the heterogeneous relationship graph at the current time includes static information and dynamic information at the current time, the second state information further includes a state information change amount, the dynamic information change amount is a change amount between dynamic information at the next time and dynamic information at the current time, and the determining unit 1003 is specifically configured to:

The above units (the acquisition unit 1001, the prediction unit 1002, and the determination unit 1003 are configured to perform the relevant steps of the above method, for example, the acquisition unit 1001 is configured to perform S201 relevant contents, the prediction unit 1002 is configured to perform S202 relevant contents, and the determination unit 1003 is configured to perform S203 relevant contents.

In the present embodiment, the prediction apparatus 1000 is presented in the form of a unit. "unit" herein may refer to an application-specific integrated circuit (ASIC), a processor and memory that execute one or more software or firmware programs, an integrated logic circuit, and/or other devices that can provide the above-described functionality. Further, the above acquisition unit 1001, prediction unit 1002, and determination unit 1003 may be implemented by the processor 1201 of the electronic device shown in fig. 12.

Referring to fig. 11, fig. 11 is a schematic structural diagram of a training device according to an embodiment of the present application. As shown in fig. 11, the training apparatus 1100 includes:

an obtaining unit 1101, configured to obtain a plurality of heterogeneous relationship graphs and fifth state information of each of the plurality of heterogeneous relationship graphs, where the plurality of heterogeneous relationship graphs are obtained at different times for the same network, and each heterogeneous relationship graph is used to characterize a relationship between a plurality of nodes and a plurality of nodes participating in network operation in a time network for obtaining each heterogeneous relationship graph;

the training unit 1102 is configured to perform multiple rounds of training on the initial prediction model based on the multiple heterogeneous relationship graphs and the fifth state information of each heterogeneous relationship graph in the multiple heterogeneous relationship graphs, so as to obtain a network performance prediction model.

In one possible embodiment, training unit 1102 is specifically configured to:

In one possible embodiment, the initial prediction model includes an initial encoding network, an initial information delivery network, and an initial decoding network, and the performance prediction model M _p includes an encoding network M _1p, an information delivery network M _2p, and a decoding network M _3p; each heterogeneous relation graph comprises a plurality of nodes and side information among the plurality of nodes; the fifth state information of each heterogeneous relationship graph includes fifth state information of each node in the heterogeneous relationship graph, and in the aspect that the fifth state information of the p-th heterogeneous relationship graph and the hidden state information H' _p in the multiple groups are input into the performance prediction model M _p to be processed, so as to obtain at least one sixth state information corresponding to the p-th heterogeneous relationship graph, the training unit 1102 is specifically configured to:

In one possible embodiment, in performing, through the information delivery network M _2p, multiple rounds of information delivery operations on the seventh state information of the jth heterogeneous relationship graph and the hidden state information H _j of the jth-1 heterogeneous relationship graph based on the side information between the plurality of nodes included in the jth heterogeneous relationship graph to obtain the hidden state information H _j+1 of the jth heterogeneous relationship graph, the training unit 1102 is specifically configured to:

In one possible embodiment, in the aspect of obtaining the state information S _k+1 of each node based on the state information S _k of the neighbor node of each node in the jth heterogeneous relationship diagram and the state information S _k of each node, the training unit 1102 is specifically configured to:

It should be noted that the above units (the obtaining unit 1101 and the training unit 1102) are configured to perform the relevant steps of the above method. For example, the acquisition unit 1101 is configured to execute the related content of S901, and the training unit 1102 is configured to execute the related content of S902.

In this embodiment, the training device 1100 is presented in the form of a unit. "unit" herein may refer to an application-specific integrated circuit (ASIC), a processor and memory that execute one or more software or firmware programs, an integrated logic circuit, and/or other devices that can provide the above-described functionality. Further, the above acquisition unit 1101 and training unit 1102 may be implemented by the processor 1301 of the electronic apparatus shown in fig. 13.

The electronic device 1200 as shown in fig. 12 may be implemented in the structure of fig. 12, the electronic device 1200 comprising at least one processor 1201, at least one memory 1202 and at least one communication interface 1203. The processor 1201, the memory 1202 and the communication interface 1203 are connected via the communication bus and perform communication with each other.

Processor 1201 may be a general purpose Central Processing Unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of the above program.

Communication interface 1203 is configured to communicate with other devices or communication networks such as ethernet, radio Access Network (RAN), wireless local area network (Wireless Local Area Networks, WLAN), etc.

The Memory 1202 may be, but is not limited to, read-Only Memory (ROM) or other type of static storage device that can store static information and instructions, random access Memory (random access Memory, RAM) or other type of dynamic storage device that can store information and instructions, as well as electrically erasable programmable read-Only Memory (ELECTRICALLY ERASABLE PROGRAMMABLE READ-Only Memory, EEPROM), compact disc read-Only Memory (Compact Disc Read-Only Memory) or other optical disc storage, optical disc storage (including compact disc, laser disc, optical disc, digital versatile disc, blu-ray disc, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. The memory may be stand alone and coupled to the processor via a bus. The memory may also be integrated with the processor.

Wherein the memory 1202 is used for storing application program codes for executing the above schemes and is controlled to be executed by the processor 1201. The processor 1201 is configured to execute application code stored in the memory 1202.

The code stored by memory 1202 may perform any of the model search methods provided above, such as:

The electronic device 1300 as shown in fig. 13 may be implemented in the structure of fig. 13, the electronic device 1300 comprising at least one processor 1301, at least one memory 1302 and at least one communication interface 1303. The processor 1301, the memory 1302, and the communication interface 1303 are connected through the communication bus and perform communication with each other.

Processor 1301 may be a general purpose Central Processing Unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of the above program schemes.

Communication interface 1303 for communicating with other devices or communication networks, such as ethernet, radio Access Network (RAN), wireless local area network (Wireless Local Area Networks, WLAN), etc.

The Memory 1302 may be, but is not limited to, read-Only Memory (ROM) or other type of static storage device that can store static information and instructions, random access Memory (random access Memory, RAM) or other type of dynamic storage device that can store information and instructions, as well as electrically erasable programmable read-Only Memory (ELECTRICALLY ERASABLE PROGRAMMABLE READ-Only Memory, EEPROM), compact disc read-Only Memory (Compact Disc Read-Only Memory) or other optical disc storage, optical disc storage (including compact disc, laser disc, optical disc, digital versatile disc, blu-ray disc, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. The memory may be stand alone and coupled to the processor via a bus. The memory may also be integrated with the processor.

The memory 1302 is used for storing application program codes for executing the above schemes, and the execution is controlled by the processor 1301. The processor 1301 is configured to execute application code stored in the memory 1302.

The code stored by memory 1302 may perform any of the model search methods provided above, such as:

The embodiment of the application also provides a computer storage medium, wherein the computer storage medium can store a program, and the program can be executed to include part or all of the steps of any one of the network performance prediction methods described in the above method embodiments.

The embodiment of the application also provides a computer storage medium, wherein the computer storage medium can store a program, and the program can be executed to include part or all of the steps of the training method of any one of the network performance prediction models described in the method embodiment.

It should be noted that, for simplicity of description, the foregoing method embodiments are all described as a series of acts, but it should be understood by those skilled in the art that the present application is not limited by the order of acts described, as some steps may be performed in other orders or concurrently in accordance with the present application. Further, those skilled in the art will also appreciate that the embodiments described in the specification are all preferred embodiments, and that the acts and modules referred to are not necessarily required for the present application.

In the foregoing embodiments, the descriptions of the embodiments are emphasized, and for parts of one embodiment that are not described in detail, reference may be made to related descriptions of other embodiments.

In the several embodiments provided by the present application, it should be understood that the disclosed apparatus may be implemented in other manners. For example, the apparatus embodiments described above are merely illustrative, such as the division of the units, merely a logical function division, and there may be additional manners of dividing the actual implementation, such as multiple units or components may be combined or integrated into another system, or some features may be omitted, or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, or may be in electrical or other forms.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in the embodiments of the present application may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.

The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable memory. Based on this understanding, the technical solution of the present application may be embodied essentially or partly in the form of a software product, or all or part of the technical solution, which is stored in a memory, and includes several instructions for causing a computer device (which may be a personal computer, a server, a network device, or the like) to perform all or part of the steps of the method according to the embodiments of the present application. And the aforementioned memory includes: a usb disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a removable hard disk, a magnetic disk, or an optical disk, or other various media capable of storing program codes.

Those of ordinary skill in the art will appreciate that all or a portion of the steps in the various methods of the above embodiments may be implemented by a program that instructs associated hardware, and the program may be stored in a computer readable memory, which may include: flash disk, read-Only Memory (ROM), random access Memory (Random Access Memory, RAM), magnetic disk or optical disk.

The foregoing has outlined rather broadly the more detailed description of embodiments of the application, wherein the principles and embodiments of the application are explained in detail using specific examples, the above examples being provided solely to facilitate the understanding of the method and core concepts of the application; meanwhile, as those skilled in the art will have variations in the specific embodiments and application scope in accordance with the ideas of the present application, the present description should not be construed as limiting the present application in view of the above.

Claims

1. A method for predicting network performance, comprising:

Acquiring a heterogeneous relation graph at a current moment and first state information of the heterogeneous relation graph at the current moment aiming at a target network, wherein the heterogeneous relation graph at the current moment is used for representing a plurality of nodes participating in network operation in the target network at the current moment and relations among the plurality of nodes;

Processing the heterogeneous relation graph at the current moment and the first state information of the heterogeneous relation graph at the current moment through a network performance prediction model of the target network to obtain second state information of the heterogeneous relation graph at the current moment, wherein the second state information comprises performance indexes of the target network at the current moment;

And determining the first state information of the heterogeneous relation diagram at the current moment at the next moment based on the second state information and the first state information of the heterogeneous relation diagram at the current moment through the network performance prediction model.

2. The method according to claim 1, wherein the network performance prediction model includes an encoding network, an information delivery network, and a decoding network, and the processing, by the network performance prediction model of the target network, the heterogeneous relationship graph at the current time and the first state information of the heterogeneous relationship graph at the current time to obtain the second state information of the heterogeneous relationship graph at the current time includes:

performing coding operation on the first state information of the heterogeneous relation diagram at the current moment through the coding network to obtain third state information of the heterogeneous relation diagram at the current moment;

carrying out multi-round information transfer operation on the heterogeneous relation diagram at the current moment, the third state information of the heterogeneous relation diagram at the current moment and the hidden state information of the heterogeneous relation diagram at the previous moment through the information transfer network so as to obtain the hidden state information of the heterogeneous relation diagram at the current moment;

and decoding the hidden state information of the heterogeneous relation diagram at the current moment through the decoding network to obtain second state information of the heterogeneous relation diagram at the current moment.

3. The method according to claim 2, wherein the current time heterogeneous relationship graph includes a plurality of nodes of the target network and side information between the plurality of nodes, the side information between the plurality of nodes is used to represent a relationship between the plurality of nodes, and the performing, by the information delivery network, a plurality of rounds of information delivery operations on the current time heterogeneous relationship graph, third state information of the current time heterogeneous relationship graph, and hidden state information of a previous time heterogeneous relationship graph to obtain hidden state information of the current time heterogeneous relationship graph includes:

Obtaining fourth state information of the heterogeneous relation diagram at the current moment based on the third state information of the heterogeneous relation diagram at the current moment and the hidden state information of the heterogeneous relation diagram at the previous moment; the fourth state information of the heterogeneous relation graph at the current moment comprises fourth state information of each node in the heterogeneous relation graph at the current moment;

Acquiring state information S _i of neighbor nodes of each node in the heterogeneous relation diagram at the current moment based on side information among the plurality of nodes when the ith round of information transfer is carried out; the i is an integer greater than 0;

Obtaining the state information S _i+1 of each node based on the state information S _i of the neighboring nodes of each node and the state information S _i of each node;

When i=1, the state information S _i of each node in the heterogeneous relationship diagram at the current time is the fourth state information of each node; when the ith round of information transfer is the last round of information transfer in the multiple rounds of information transfer, the state information S _i+1 of each node is the hidden state information of each node, and the hidden state information of the heterogeneous relationship graph at the current moment includes the hidden state information of each node in the heterogeneous relationship graph at the current moment.

4. A method according to claim 3, wherein the obtaining the state information S _i+1 of each node based on the state information S _i of the neighboring node of each node and the state information S _i of each node includes:

performing aggregation operation on the state information S _i of the neighbor node of each node to obtain aggregated state information of each node;

Converting the aggregated state information of each node to obtain converted state information of each node;

and updating the state information S _i of each node based on the converted state information of each node to obtain the state information S _i+1 of each node.

5. The method according to any one of claims 1 to 4, wherein the first state information of the current time heterogeneous relationship map includes static information and dynamic information of the current time, the second state information further includes a dynamic information change amount, the dynamic information change amount is a change amount between dynamic information of a next time relative to the dynamic information of the current time, and the determining the first state information of the current time heterogeneous relationship map at the next time based on the second state information and the first state information of the current time heterogeneous relationship map includes:

determining dynamic information variation from second state information of the heterogeneous relation diagram at the current moment;

And obtaining first state information of the heterogeneous relation diagram at the current moment at the next moment according to the dynamic information variable quantity, the dynamic information at the current moment and the static information.

6. The method of claim 5, wherein the plurality of nodes of the target network comprise: network devices, links, queues, paths, and flows.

7. A method for training a network performance prediction model, comprising:

Acquiring a plurality of heterogeneous relationship graphs and fifth state information of each heterogeneous relationship graph in the plurality of heterogeneous relationship graphs, wherein the plurality of heterogeneous relationship graphs are acquired at different moments aiming at the same network, and each heterogeneous relationship graph is used for representing a plurality of nodes participating in network operation and relations among the plurality of nodes in the network at the moment of acquiring each heterogeneous relationship graph; the plurality of nodes includes: network devices, links, queues, paths, and flows;

And training the initial prediction model for multiple rounds based on the heterogeneous relation diagrams and fifth state information of each heterogeneous relation diagram in the heterogeneous relation diagrams so as to obtain a network performance prediction model.

8. The method of claim 7, wherein the training the initial predictive model based on the plurality of heterogeneous relationship graphs and the fifth state information for each of the plurality of heterogeneous relationship graphs to obtain the network performance predictive model comprises:

Dividing the plurality of heterogeneous relationship maps into a plurality of groups, each group comprising at least one heterogeneous relationship map;

when the p-th training is performed, inputting fifth state information and hidden state information H' _p of the p-th heterogeneous relation diagram in the multiple groups into a performance prediction model M _p for processing to obtain at least one sixth state information corresponding to the p-th heterogeneous relation diagram, wherein the sixth state information comprises performance indexes and dynamic information variation, and p is an integer greater than 0;

Inputting the performance index and the dynamic information variation included in the at least one sixth state information into a loss function for calculation to obtain a loss value L _p;

Adjusting parameters of the performance prediction model M _p based on the loss value L _p to obtain a performance prediction model M _p+1; when p=1, the performance prediction model M _p is an initial prediction model, and the hidden state information H' _p is an empty set; when p is greater than 1, the hidden state information H' _p is obtained according to the last heterogeneous relationship diagram in the p-1 group, and when the p-th training is the last training, the performance prediction model M _p+1 is the network performance prediction model.

9. The method of claim 8, wherein the initial prediction model comprises an initial encoding network, an initial information delivery network, and an initial decoding network, and wherein the performance prediction model M _p comprises an encoding network M _1p, an information delivery network M _2p, and a decoding network M _3p; each heterogeneous relation graph comprises a plurality of nodes and side information among the nodes; the fifth state information of each heterogeneous relationship graph includes fifth state information of each node in the heterogeneous relationship graph, and the inputting the fifth state information and the hidden state information H' _p of the p-th heterogeneous relationship graph in the plurality of groups into the performance prediction model M _p for processing to obtain at least one sixth state information corresponding to the p-th heterogeneous relationship graph includes:

and carrying out the following processing on the fifth state information of the j-th heterogeneous relation diagram in the p-th group to obtain sixth state information corresponding to the j-th heterogeneous relation diagram:

Processing the fifth state information of each node in the jth heterogeneous relation diagram through a coding network M _1p to obtain seventh state information of the jth heterogeneous relation diagram;

Based on the side information among the plurality of nodes included in the jth heterogeneous relationship diagram, carrying out multiple rounds of information transfer operation on the seventh state information of the jth heterogeneous relationship diagram and the hidden state information H _j of the jth-1 heterogeneous relationship diagram through an information transfer network M _2p so as to obtain the hidden state information H _j+1 of the jth heterogeneous relationship diagram;

Decoding the hidden state information H _j+1 of the jth heterogeneous relation diagram through a decoding network M _3p to obtain sixth state information corresponding to the jth heterogeneous relation diagram;

When j=1, the hidden state information H _j of the j-1 st heterogeneous relationship diagram is the hidden state information H' _p; when p=1, the encoding network M _1p is the initial encoding network, the information delivery network M _2p is the initial information delivery network, and the decoding network M _3p is the initial decoding network.

10. The method according to claim 9, wherein the performing, through the information delivery network M _2p, a plurality of rounds of information delivery operations on the seventh state information of the jth heterogeneous relationship graph and the hidden state information H _j of the jth-1 heterogeneous relationship graph based on the side information between the plurality of nodes included in the jth heterogeneous relationship graph to obtain the hidden state information H _j+1 of the jth heterogeneous relationship graph includes:

Obtaining eighth state information of the j-th heterogeneous relation diagram based on the seventh state information of the j-th heterogeneous relation diagram and hidden state information H _j of the j-1-th heterogeneous relation diagram; the eighth state information of the jth heterogeneous relation graph comprises the eighth state information of each node in the jth heterogeneous relation graph;

When the kth round of information transfer is performed, acquiring state information S _k of a neighbor node of each node in the jth heterogeneous relationship diagram based on side information among a plurality of nodes included in the jth heterogeneous relationship diagram; the neighbor node of each node is a node with an edge between the neighbor node and the node; k is an integer greater than 0;

Obtaining the state information S _k+1 of each node based on the state information S _k of the neighbor node of each node in the j-th heterogeneous relation diagram and the state information S _k of each node;

When k=1, the state information S _k of the neighboring node is eighth state information of the neighboring node, and the state information S _k of each node is eighth state information of each node; when the kth round of information transfer is the last round of information transfer in the multiple rounds of information transfer, the state information S _i+1 of each node is the hidden state information of each node, and the hidden state information H _j+1 of the jth heterogeneous relationship graph includes the hidden state information of each node in the jth heterogeneous relationship graph.

11. The method according to claim 10, wherein the obtaining the state information S _k+1 of each node based on the state information S _k of the neighbor node of each node in the jth heterogeneous relationship graph and the state information S _k of each node includes:

Performing aggregation operation on the state information S _k of the neighbor node of each node to obtain aggregated state information of each node;

And updating the state information of each node based on the converted state information of each node to obtain the state information S _k+1 of each node.

12. A prediction apparatus, comprising:

the system comprises an acquisition unit, a processing unit and a processing unit, wherein the acquisition unit is used for acquiring a heterogeneous relation graph at the current moment and first state information of the heterogeneous relation graph at the current moment aiming at a target network, and the heterogeneous relation graph at the current moment is used for representing a plurality of nodes participating in network operation in the target network at the current moment and relations among the plurality of nodes;

and the determining unit is used for determining the first state information of the heterogeneous relation diagram at the current moment at the next moment based on the second state information and the first state information of the heterogeneous relation diagram at the current moment through the network performance prediction model.

13. The apparatus of claim 12, wherein the network performance prediction model comprises an encoding network, an information delivery network, and a decoding network, the prediction unit being specifically configured to

14. The apparatus according to claim 13, wherein the current time heterogeneous relationship graph includes a plurality of nodes of the target network and side information between the plurality of nodes, the side information between the plurality of nodes being used to represent relationships between the plurality of nodes, and the prediction unit is specifically configured to, in the performing, by the information delivery network, a plurality of rounds of information delivery operations on the current time heterogeneous relationship graph, third state information of the current time heterogeneous relationship graph, and hidden state information of a previous time heterogeneous relationship graph to obtain hidden state information of the current time heterogeneous relationship graph:

Acquiring state information S _i of neighbor nodes of each node in the heterogeneous relation diagram at the current moment based on side information among the plurality of nodes when the ith round of information transfer is carried out; the neighbor node of each node is a node with an edge between the neighbor node and the node; the i is an integer greater than 0;

15. The apparatus according to claim 14, wherein, in the aspect that the state information S _i+1 of each node is obtained based on the state information S _i of the neighboring node of each node and the state information S _i of each node, the prediction unit is specifically configured to:

16. The apparatus according to any one of claims 12-15, wherein the first state information of the current time heterogeneous relation graph includes static information and dynamic information of the current time, the second state information further includes a dynamic information change amount, the dynamic information change amount is a change amount between dynamic information of a next time relative to the dynamic information of the current time, the determining unit is specifically configured to determine the first state information of the current time heterogeneous relation graph at the next time based on the second state information and the first state information of the current time heterogeneous relation graph:

17. The apparatus of claim 16, wherein the plurality of nodes of the target network comprise: network devices, links, queues, paths, and flows.

18. A training device, comprising:

an obtaining unit, configured to obtain a plurality of heterogeneous relationship graphs and fifth state information of each heterogeneous relationship graph in the plurality of heterogeneous relationship graphs, where the plurality of heterogeneous relationship graphs are obtained at different times for a same network, and each heterogeneous relationship graph is used to characterize a plurality of nodes participating in network operation in the network at a time when each heterogeneous relationship graph is obtained and a relationship between the plurality of nodes; wherein the plurality of nodes comprises: network devices, links, queues, paths, and flows;

The training unit is used for carrying out multi-round training on the initial prediction model based on the heterogeneous relation diagrams and fifth characteristic information of each heterogeneous relation diagram in the heterogeneous relation diagrams so as to obtain a network performance prediction model.

19. The device according to claim 18, wherein the training unit is specifically configured to:

20. The apparatus of claim 19, wherein the initial prediction model comprises an initial encoding network, an initial information delivery network, and an initial decoding network, and wherein the performance prediction model M _p comprises an encoding network M _1p, an information delivery network M _2p, and a decoding network M _3p; each heterogeneous relation graph comprises a plurality of nodes and side information among the nodes; the fifth state information of each heterogeneous relationship graph includes fifth state information of each node in the heterogeneous relationship graph, in the aspect that the fifth state information and the hidden state information H' _p of the p-th heterogeneous relationship graph in the plurality of groups are input into the performance prediction model M _p for processing, so as to obtain at least one sixth state information corresponding to the p-th heterogeneous relationship graph, the training unit is specifically configured to:

The fifth state information of the j-th heterogeneous relation diagram in the p-th group is processed as follows to obtain sixth state information corresponding to the j-th heterogeneous relation diagram;

Based on the side information among the plurality of nodes included in the jth heterogeneous relationship diagram, carrying out multi-round information transfer operation on the seventh state information of the jth heterogeneous relationship diagram and the hidden state information H _j of the jth-1 heterogeneous relationship diagram through an information transfer network M _2p so as to obtain the hidden state information H _j+1 of the jth heterogeneous relationship diagram;

Decoding the hidden state information H _j+1 of the jth heterogeneous relationship diagram through a decoding network M _3p to obtain sixth state information corresponding to the jth heterogeneous relationship diagram

21. The apparatus according to claim 20, wherein, in the aspect of obtaining the hidden state information H _j+1 of the jth heterogeneous relationship diagram based on the side information among the plurality of nodes included in the jth heterogeneous relationship diagram, the training unit is specifically configured to:

when the kth round of information transfer is carried out, acquiring the state information S _k of the neighbor node of each node in the jth heterogeneous relation diagram based on the jth heterogeneous relation diagram; the neighbor node of each node is a node with an edge between the neighbor node and the node;

Obtaining the state information S _k+1 of each node based on the state information S _k of the neighbor node of each node in the j-th heterogeneous relation diagram and the state information S _k of each node; k is an integer greater than 0;

22. The apparatus of claim 21, wherein, in the aspect of obtaining the state information S _k+1 of each node based on the state information S _k of the neighbor node of each node in the jth heterogeneous relationship diagram and the state information S _k of each node, the training unit is specifically configured to:

23. An electronic device comprising a processor and a memory, wherein the processor and the memory are connected, wherein the memory is for storing program code, the processor is for invoking the program code to implement the method of any of claims 1-11.

24. A computer storage medium comprising computer instructions which, when run on an electronic device, cause the electronic device to perform the method of any one of claims 1 to 11.