Heterogeneous computing power network task allocation and path optimization system based on neural network

Abstract

The invention relates to the technical field of network task allocation, and discloses a heterogeneous computing power network task allocation and path optimization system based on a neural network, and the system comprises a graph construction module which is used for constructing a traffic dispersion graph, and carrying out the modeling of network topology and node data through the traffic dispersion graph, performing node and edge modeling in combination with graph index and graph displacement, and extracting network communication flow and node features; the node task distribution module is used for forming a graph index set by the graph indexes of the nodes, inputting the graph index set into the improved graph neural network model, performing intelligent prediction and distribution on node tasks, and outputting task distribution prediction results of the nodes; the path optimization module is combined with the node task prediction result to minimize transmission delay and balance node computing power to obtain a path optimization result; and the system optimization module is used for optimizing the overall performance of the system based on the path optimization result and the task allocation prediction result. The task allocation and path optimization efficiency of the heterogeneous computing power network is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of network task allocation, and in particular to a heterogeneous computing power network task allocation and path optimization system based on neural networks. Background Art

[0002] With the rapid development of information technology, computing networks are playing a key role in numerous fields, such as cloud computing, edge computing, and big data processing. In heterogeneous computing networks, different types of computing nodes (such as CPUs, GPUs, and FPGAs) have unique performance characteristics, and the network topology is complex and diverse.

[0003] Currently, task allocation and path optimization face numerous challenges in heterogeneous computing networks. Traditional task allocation and path planning methods often fail to fully account for the real-time performance changes of network nodes and the complex topology, leading to problems such as unbalanced task allocation, high transmission latency, and low resource utilization. For example, in large-scale data processing tasks, improper task allocation can overload some nodes while leaving other nodes idle, impacting overall processing efficiency. In applications with high real-time requirements (such as autonomous driving and telemedicine), high transmission latency can lead to untimely system responses, resulting in serious consequences.

[0004] Neural networks have advantages in processing complex data relationships, but traditional neural networks also have limitations when applied to task allocation and path optimization in heterogeneous computing networks. These limitations include insufficient processing capabilities for graph-structured data and a tendency to oversmooth as depth increases. These limitations make it difficult to effectively handle the dynamic changes and complex topologies found in heterogeneous computing networks. Therefore, a new technical solution is urgently needed to address these issues and improve the efficiency of task allocation and path optimization in heterogeneous computing networks. Summary of the Invention

[0005] The present invention aims to solve the technical problems of unbalanced task allocation, high transmission delay and low resource utilization in traditional task allocation and path optimization in heterogeneous computing power networks, and proposes a task allocation and path optimization system for heterogeneous computing power networks based on neural networks.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] The present invention provides a neural network-based heterogeneous computing power network task allocation and path optimization system, comprising:

[0008] The graph construction module is used to construct a traffic dispersion graph from the captured network data stream, model the network topology and node data through the traffic dispersion graph, and model the nodes and edges by combining graph index and graph displacement to extract network communication traffic and node features;

[0009] The node task assignment module is used to input the graph index set composed of the graph indexes of each node into the improved graph neural network model, perform intelligent prediction and assignment of node tasks, and output the node task assignment prediction results;

[0010] The path optimization module combines the node task prediction results to select the optimal data transmission path to minimize transmission delay and balance node computing power to obtain the path optimization result;

[0011] The system optimization module optimizes the overall system performance based on the path optimization result and the task allocation prediction result.

[0012] In some preferred embodiments, the graph construction module includes:

[0013] Network data flow capture unit, used to collect performance data of heterogeneous computing nodes in real time, including communication traffic, throughput and path information;

[0014] Individual flow graph construction units are used to construct a graph structure of network nodes and communication paths, where nodes represent hosts or computing nodes in the network, edges represent communication flow paths between nodes, and edge weights represent the data flow throughput between nodes;

[0015] The traffic dispersion graph generation unit is used to merge multiple individual flow graphs to generate a traffic dispersion graph representing the overall topological structure of the network, including: using an adjacency matrix to represent the connection relationship between nodes; and identifying abnormal nodes based on the flow threshold of the throughput node.

[0016] In some preferred embodiments, the graph construction module further includes:

[0017] The graph index and graph displacement construction unit is used to construct the graph index and graph displacement; wherein,

[0018] The graph index includes computing capacity, load level and throughput, and is used to represent attribute information of nodes;

[0019] The graph displacement is used to realize the translation operation of the graph index on the graph. The weighted matrix Laplace is selected as the graph displacement to perform weighted combination of node signals and their neighboring nodes, extract the structural features of the graph, and finally realize local smoothing of the graph index.

[0020] In some preferred embodiments, the graph construction module further includes:

[0021] The graph Fourier transform unit is used to perform graph Laplace eigendecomposition on the weighted matrix Laplace, extract the frequency domain features of the graph index, and then project the graph index into the frequency domain of the graph through graph Fourier transform to analyze the global structure and transmission characteristics between nodes.

[0022] In some preferred embodiments, the improved graph neural network model introduces identity mapping and residual connection, and updates node features through multi-layer iteration to finally obtain the task allocation prediction result of the node.

[0023] In some preferred embodiments, the node feature H (l) The update formula is as follows:

[0024]

[0025] Among them, H (l) : node feature matrix of the lth layer; H (0) =X: input initial node feature matrix; μ: identity mapping coefficient, used to balance the initial feature H (0) The influence of ρ: residual connection coefficient, used to introduce the previous layer feature H (l-1) the impact of; The normalized adjacency matrix of the graph,

[0026]

[0027] Where A: original adjacency matrix; I: identity matrix; D: degree matrix of the graph;

[0028] σ: activation function, the expression is:

[0029]

[0030] where x' is an input variable; θ1∈(0,∞) and

[0031] In some preferred embodiments, the node task allocation module also includes a unit for defining a node task allocation objective function, which is used to measure node load balance and resource utilization, and the objectives include maximizing overall computing efficiency, load balancing, and minimizing network communication overhead.

[0032] In some preferred embodiments, the path optimization module is used to set a path optimization objective function, which is composed of a transmission delay item from any node to other nodes and a load balancing weight item of any node, wherein the load balancing weight item of any node includes the task prediction result of any node, the current task load of any node and the computing power information of any node; minimize the transmission delay and balance the node computing power load, and finally select the optimal data transmission path.

[0033] In some preferred embodiments, the path optimization module further performs path optimization in combination with the frequency characteristics of the graph Fourier transform domain to find the path with the best load balancing, including:

[0034] Input node load balancing weight and transmission delay;

[0035] Perform graph Fourier transform on the load balancing weights of nodes;

[0036] Filter out high-frequency components and retain low-frequency components;

[0037] Inverse transform back to the node domain to obtain the smoothed node load balancing weight;

[0038] The optimal path is solved based on the updated node load balancing weights and transmission delays.

[0039] In some preferred embodiments, the system optimization module constructs a loss function for task scheduling and system optimization, wherein the loss function includes a transmission delay loss term and a load balancing loss term, and is used to minimize the sum of delays on the task transmission path and the variance of the node load respectively;

[0040] The system optimization module also includes: using a gradient descent method to iteratively update weight parameters to optimize the overall performance of the system.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) Through the improved graph neural network model, the present invention can more accurately predict and allocate node tasks, avoiding the problem of unbalanced task allocation in traditional methods; by combining physical information and graph indexes, the system can analyze the performance data of heterogeneous computing nodes in real time, thereby making more reasonable task allocation decisions and improving task allocation efficiency.

[0043] (2) Through the path optimization module, the present invention can select the optimal data transmission path and minimize transmission delay, especially in application scenarios with high real-time requirements, such as autonomous driving and telemedicine, effectively reducing transmission delay. The application of graph Fourier transform enables the system to analyze the global structure and transmission characteristics between nodes, further optimizing path selection.

[0044] (3) By balancing node loads, the present invention avoids situations where some nodes are overloaded while other nodes are idle, thereby improving resource utilization of the overall computing network. By constructing a loss function for task scheduling and system optimization, it is possible to minimize the variance of transmission delay and node load, further improving resource utilization.

[0045] (4) By introducing identity mapping and residual connection, the present invention solves the over-smoothing problem that occurs in traditional graph neural networks as the depth increases, thereby enhancing the robustness and stability of the system.

[0046] In summary, the present invention significantly improves the task allocation and path optimization efficiency of heterogeneous computing power networks, reduces transmission delays, improves resource utilization, and enhances the robustness and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0048] Figure 1 This is a module diagram of a neural network-based heterogeneous computing power network task allocation and path optimization system provided by an embodiment of the present invention;

[0049] Figure 2 1 is a schematic diagram of a unit of the graph construction module 101 provided in an embodiment of the present invention;

[0050] Figure 3 This is a flow chart of a method for task allocation and path optimization of a heterogeneous computing power network based on a neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0052] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limitations on the present invention.

[0053] In the description of the present invention, "several" means one or more, "many" means more than two, "greater than," "less than," and "exceed" are understood to exclude the number itself, while "above," "below," and "within" are understood to include the number itself. The use of "first" and "second" in the description is solely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, implicitly specifying the number of the indicated technical features, or implicitly specifying the order of the indicated technical features.

[0054] Figure 1 This is a module diagram of a neural network-based heterogeneous computing power network task allocation and path optimization system provided by an embodiment of the present invention. Figure 3 This is a flowchart of a method for task allocation and path optimization of a heterogeneous computing network based on a neural network provided by an embodiment of the present invention. Figure 1 and Figure 3 As shown, the present invention provides a neural network-based heterogeneous computing power network task allocation and path optimization system 100, comprising:

[0055] Graph construction module 101 is used to construct a traffic dispersion graph using the captured network data stream, model the network topology and node data using the traffic dispersion graph, and model nodes and edges using graph indexes and graph displacements to extract network communication traffic and node features;

[0056] This is the first step of this system. The network topology and node data are modeled through the traffic dispersion graph, and the nodes and edges are modeled by combining graph index and graph displacement to extract network communication traffic and node characteristics.

[0057] (1) Graph construction

[0058] The graph construction module 101 includes:

[0059] The network data flow capture unit 1011 is used to collect the performance data of heterogeneous computing nodes in real time, including but not limited to communication traffic, throughput and path information;

[0060] By monitoring network packets, the following information can be extracted:

[0061] Source IP address: the IP address of the node sending the data packet sTc .

[0062] Destination IP address: the IP address of the receiving node of the data packet dst .

[0063] Path information: the path P along which the data packet is transmitted src→dst .

[0064] Throughput T: the amount of data transmitted per unit time.

[0065] Individual flow graph construction unit 1012, used to construct a graph structure of network nodes and communication paths, where nodes represent hosts or computing nodes in the network, edges represent communication flow paths between nodes, and edge weights represent data flow throughput between nodes;

[0066] Individual flow graph construction, constructing the graph structure G = (V, E, W) of network nodes and communication paths.

[0067] The communication flow of data packets is modeled as a graph structure, where:

[0068] Node: A host or computing node in a network.

[0069] Edge: The communication traffic path between nodes.

[0070] Edge weight: data flow throughput T between nodes ij .

[0071] Specifically represented as Figure G:

[0072] G=(V,E,W)

[0073] V: node set (all hosts or computing nodes); E: edge set (communication paths between nodes); W: edge weight set, w ij =T ij represents the throughput between node i and node j.

[0074] The traffic dispersion graph generation unit 1013 is used to merge multiple individual flow graphs to generate a traffic dispersion graph representing the overall network topology, including: using an adjacency matrix to represent the connection relationship between nodes, and identifying abnormal nodes based on the flow threshold of the throughput node.

[0075] Traffic dispersion graph generation: multiple individual flow graphs generated by the individual flow graph construction unit 1012 are merged to generate a traffic dispersion graph for representing the overall topological structure of the network.

[0076] Adjacency matrix A: represents the connection relationship between nodes. If there is data flow between nodes i and j, then:

[0077]

[0078] Abnormal node detection: Based on throughput T ij and the node's traffic threshold T th , identify abnormal nodes. The abnormality degree of node i is defined as:

[0079]

[0080] in, The set of neighbor nodes of node i; T ij : actual throughput of nodes i and j; T th : Normal traffic threshold.

[0081] If D(i)>∈(set abnormality threshold), node i is judged to be abnormal.

[0082] The above abnormal node detection mechanism enables the system to promptly identify and handle abnormal situations in the network to ensure the normal operation of the system.

[0083] (2) Constructing graph index and graph displacement

[0084] In some preferred embodiments, the graph construction module 101 further includes:

[0085] The graph index and graph displacement construction unit 1014 is used to construct the graph index and graph displacement; wherein,

[0086] A graph index is a set of vectors indexed by graph nodes, representing the attribute information of nodes. In this system, the graph index x of node i is i =[C i ,L i ,T i ],include:

[0087] Computing ability C i : The computing resources (CPU / GPU performance) of node i.

[0088] Load level L i : The task load currently executed by node i.

[0089] Throughput T i : Data transmission throughput of node i (in bps).

[0090] Graph displacement is used to realize the translation operation of graph index on the graph. The weighted matrix Laplace is selected as the graph displacement to perform weighted combination of node signals and their neighboring nodes, extract the structural features of the graph, and finally achieve local smoothing of the graph index.

[0091] Defined as:

[0092] Laplace=DA

[0093] in:

[0094] D: The degree matrix of the graph, which is a diagonal matrix, D ii =∑ j A ij .

[0095] A: Adjacency matrix of the graph.

[0096] When the graph displacement Laplace acts on the graph index x, it is expressed as:

[0097] Laplacex=(DA)x

[0098] This operation achieves local smoothing of the graph index, performs a weighted combination of node signals and their neighboring nodes, and extracts the structural features of the graph.

[0099] In some preferred embodiments, the graph construction module 101 further includes:

[0100] The graph Fourier transform unit 1015 is used to perform graph Laplace eigendecomposition on the weighted matrix Laplace, extract the frequency domain features of the graph index, and then project the graph index into the frequency domain of the graph through graph Fourier transform to analyze the global structure and transmission characteristics between nodes.

[0101] (3) Graph Fourier transform

[0102] Perform graph Laplace eigendecomposition on the weighted matrix Laplace to extract the frequency domain features of the graph index:

[0103] Laplace=UΛU T

[0104] in:

[0105] U: Eigenvector matrix of the weighting matrix Laplace.

[0106] Λ: Diagonal matrix of eigenvalues of the weighting matrix Laplace.

[0107] The graph Fourier transform projects the graph index x into the frequency domain of the graph to analyze the global structure and transmission characteristics between nodes. The graph Fourier transform of the graph index x is expressed as:

[0108]

[0109] The representation of the graph index x in the graph Fourier domain (frequency component), that is, the frequency characteristics of the graph Fourier transform domain.

[0110] U T : The eigenvector transpose matrix of the graph Laplacian matrix.

[0111] Graph Construction Module 101 models network topology and node data using a traffic dispersion graph. It then combines graph indexes and graph displacements to model nodes and edges, enabling comprehensive extraction of network traffic and node characteristics. An abnormal node detection mechanism enables the system to promptly identify and address network anomalies, ensuring data accuracy and reliability.

[0112] The node task assignment module 102 is used to input the graph index set composed of the graph index of each node into the improved graph neural network model, perform intelligent prediction and assignment of node tasks, and output the node task assignment prediction result;

[0113] (1) Improved graph neural network model

[0114] To solve the over-smoothing problem that occurs when the depth of traditional graph neural networks increases: 1) The improved graph neural network model first introduces an identity mapping to preserve the initial feature H of the node through the coefficient μ (0)Even if the network is deepened, the initial information can still be effectively transmitted to prevent the features from being overly smoothed. 2) The improved graph neural network model also introduces a residual connection to introduce the feature H of the previous layer through the coefficient ρ (l-1) , which enables information to jump between layers and alleviates the gradient vanishing problem.

[0115] Improved graph neural network model: By introducing identity mapping and residual connection, and updating node features through multi-layer iteration, the task allocation prediction result of the node is finally obtained.

[0116] Update node features H through multi-layer iteration (l) , and finally obtain the node task allocation prediction result. The core goal of task allocation is to minimize the imbalance of node load through network characteristics H while making full use of the node's computing power resources.

[0117] Among them, the node feature H (l) The update formula is as follows:

[0118]

[0119] Among them, H (l) : node feature matrix of the lth layer; H (0) =X: input initial node feature matrix; μ: identity mapping coefficient, used to balance the initial feature H (0) The influence of ρ: residual connection coefficient, used to introduce the previous layer feature H (l-1) the impact of; The normalized adjacency matrix of the graph is obtained by normalizing the adjacency matrix For node feature H (l-1) Aggregation is performed to achieve effective fusion of node and neighbor information. Defined as:

[0120]

[0121] Where A: original adjacency matrix; I: identity matrix (with self-loops added to ensure the participation of node information); D: degree matrix of the graph;

[0122] σ: activation function, the expression is:

[0123]

[0124] where x' is an input variable; θ1∈(0,∞) and

[0125] (2) Define the node task allocation objective function

[0126] The node task allocation module 102 further includes a unit for defining a node task allocation objective function, which is used to measure node load balance and resource utilization. The objectives include maximizing overall computing efficiency, load balancing, and minimizing network communication overhead.

[0127] The objective function J is defined as follows:

[0128]

[0129] Where J is the comprehensive cost function (needs to be minimized).

[0130] N: total number of nodes.

[0131] C i : The computing power of node i (such as CPU and GPU computing power).

[0132] L i : Current task load of node i.

[0133] T ij : The communication throughput between node i and node j.

[0134] d ij : The communication overhead (such as delay) between nodes i and j, that is, the transmission delay from node i to node j.

[0135] The average load of all nodes in the network.

[0136] ΔL i : Predict the new task load assigned to node i.

[0137] α, β, γ, δ: weight coefficients used to balance the importance of computational efficiency, communication overhead, load balancing, and resource utilization.

[0138] ∈0: A small positive number to avoid the denominator being zero.

[0139] By training the improved graph neural network model, J is minimized and the node task allocation prediction result ΔL is obtained i , which is the new task load of each node. Ensure that node tasks are properly distributed to avoid overload and resource waste.

[0140] The node task assignment module 102 intelligently predicts and assigns node tasks by inputting a graph index set composed of the graph indexes of each node into an improved graph neural network model. This module also defines a node task assignment objective function to measure node load balancing and resource utilization, ensuring the rationality and effectiveness of task assignment.

[0141] The path optimization module 103 selects the optimal data transmission path based on the node task prediction result obtained by the node task allocation module 102 to minimize transmission delay and balance node computing power, thereby obtaining a path optimization result.

[0142] Specifically, in the path optimization stage, combined with the obtained node task prediction result ΔL i ,select the optimal data transmission path, the goal is to minimize the transmission delay and balance the node computing power, to achieve efficient data scheduling and path planning.

[0143] Among them, the path optimization module 103 is used to set the path optimization objective function, which is composed of the transmission delay item from any node to other nodes and the load balancing weight item of any node, wherein the load balancing weight item of any node includes the task prediction result of any node, the current task load of any node and the computing power information of any node; minimize the transmission delay and balance the node computing power load, and finally select the optimal data transmission path.

[0144] (1) Setting path optimization goals

[0145] The task load and node computing power information jointly determine the availability and load of each node in the network. The goal of path optimization is to minimize transmission delay while considering the node computing power load balance. By selecting paths, minimizing transmission delay and load imbalance, the objective function constructed is as follows:

[0146]

[0147] λ: Weight parameter, which adjusts the importance of transmission delay and load balancing.

[0148] P: The node set of the data transmission path.

[0149] W i : The load balancing weight of node i, expressed as:

[0150]

[0151] (2) Path optimization based on the Fourier transform domain frequency characteristics

[0152] In the Fourier transform domain of a graph, low-frequency components reflect smooth changes in node load balancing weights, indicating good overall load balancing. High-frequency components reflect dramatic changes in node load balancing weights, indicating that certain nodes in the network are overloaded or have severe transmission bottlenecks. Therefore, it is necessary to filter out high-frequency components and retain low-frequency components to find the path with the best load balancing. Based on the graph's frequency domain information, highly loaded nodes or transmission bottleneck nodes can be eliminated.

[0153] The steps for path optimization based on the frequency characteristics of the graph Fourier transform domain include:

[0154] 2.1 Input node load balancing weight W and transmission delay d ij : Among them, the node load balancing weight vector W represents the ratio of the task load to the computing power of each node in the network. The higher W is, the busier the node is or the more limited its resources are. In path selection, we should try to avoid selecting too many of these nodes.

[0155] 2.2 Perform graph Fourier transform on the load balancing weight of the node:

[0156]

[0157] Here U is the eigenvector matrix of the graph Laplacian matrix, which is used to project the node weights W into the graph Fourier frequency domain.

[0158] 2.3 Retain the low-frequency components after Fourier transformation and filter out the high-frequency components, that is, remove the high-load bottleneck. In the frequency domain, it is expressed as Each of these is the component corresponding to a certain frequency k. Select a filter function H(k) to suppress the high-frequency components in the frequency domain, and calculate the filtered frequency signal by element-wise multiplication:

[0159]

[0160] in, is a parameter that controls the filtering strength.

[0161] 2.4 Inverse transform back to the node domain to obtain the smoothed node load balancing weight W ' :

[0162]

[0163] 2.5 Based on the updated node load balancing weight W ' and transmission delay d ij , solve the optimal path P * :

[0164]

[0165] Among them, arg min P The role is to find the path P * To minimize the objective function, we can use dynamic programming to perform optimal recursion on the weights of the segmented paths to obtain the optimal output path P * , which can not only ensure a small transmission delay, but also consider the balance of node weight distribution.

[0166] The path optimization module 103 combines the node task prediction results to select the optimal data transmission path to minimize transmission delay and balance node computing power. It analyzes the global structure and transmission characteristics between nodes through graph Fourier transform, further optimizes path selection, and improves data transmission efficiency.

[0167] The system optimization module 104 optimizes the overall system performance based on the path optimization results and the task allocation prediction results.

[0168] (1) Constructing a loss function for task scheduling and system optimization

[0169] The system optimization module 104 constructs a loss function for task scheduling and system optimization, with the goal of minimizing transmission delay and balancing node computing load. The loss function includes a transmission delay loss term and a load balancing loss term. It can be expressed as a weighted sum of these two parts:

[0170]

[0171] in: Overall loss function; Transmission delay loss term; Load balancing loss term; γ1 and γ2: weight coefficients in the loss function, used to balance the influence of the two parts.

[0172] Among them, the transmission delay loss term is the time required for the task to be transferred from the source node to the destination node. Assume that the task is transferred through path P and the delay from node i to node j is d ij , transmission delay loss term Defined as:

[0173]

[0174] P: The optimal path for task transmission.

[0175] d ij : The transmission delay between node i and node j.

[0176] The goal of this loss term is to minimize the sum of delays along the task transmission path.

[0177] Load balancing loss item The goal is to make the task load of each node more evenly distributed to avoid overloading of a single node. Defined as:

[0178]

[0179] W i: The load balancing weight of node i.

[0180] The mean of the node weights, representing the average load level of the entire system, is defined as:

[0181]

[0182] N: The set of all nodes in the network.

[0183] The goal of this loss term is to minimize the variance of node load so that the load is evenly distributed throughout the network.

[0184] (2) Weight parameter λ j Adjustments:

[0185] In path optimization and task scheduling, the weight parameter λ j The adjustment of is crucial to improving system accuracy. The weight parameters are iteratively updated using the gradient descent method. The update formula is as follows:

[0186]

[0187] in:

[0188] The weight parameter for the t+1th iteration.

[0189] η: learning rate, controls the update step size.

[0190] The partial derivative of the loss function with respect to the weight parameters.

[0191] The system optimization module minimizes transmission delay and node load variance by constructing the aforementioned loss function for task scheduling and system optimization, further improving overall system performance. Gradient descent is used to iteratively update weight parameters to optimize overall system performance and ensure system stability and reliability in different application scenarios.

[0192] The above is an introduction to the system embodiment. The following further illustrates the solution of the present invention through a method embodiment.

[0193] Another embodiment of the present invention provides a dynamic demand computing and network fusion scheduling optimization method based on an improved large model, including:

[0194] A traffic dispersion graph is constructed using captured network data streams. The network topology and node data are modeled using the traffic dispersion graph. Nodes and edges are modeled using graph indexes and graph displacements to extract network communication traffic and node features.

[0195] The graph index of each node is combined into a graph index set and input into the improved graph neural network model to intelligently predict and allocate node tasks, and output the node task allocation prediction results;

[0196] Combined with the node task prediction results, the optimal data transmission path is selected to minimize transmission delay and balance node computing power, obtaining the path optimization result;

[0197] Based on the path optimization results and the task allocation prediction results, the overall system performance is optimized.

[0198] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described method steps can refer to the corresponding process in the aforementioned system embodiment and will not be repeated here.

[0199] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the claims of the present invention.

[0200] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present invention can be achieved. This is not limited herein.

[0201] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A neural network-based heterogeneous computing power network task allocation and path optimization system, characterized by: include: The graph construction module is used to construct a traffic dispersion graph from the captured network data stream, model the network topology and node data through the traffic dispersion graph, and model the nodes and edges by combining graph index and graph displacement to extract network communication traffic and node features; The node task assignment module is used to input the graph index set composed of the graph indexes of each node into the improved graph neural network model, perform intelligent prediction and assignment of node tasks, and output the node task assignment prediction results; The path optimization module combines the node task prediction results to select the optimal data transmission path to minimize transmission delay and balance node computing power to obtain the path optimization result; The system optimization module optimizes the overall system performance based on the path optimization result and the task allocation prediction result.

2. The neural network-based heterogeneous computing power network task allocation and path optimization system according to claim 1 is characterized in that: in, The graph construction module includes: Network data flow capture unit, used to collect performance data of heterogeneous computing nodes in real time, including communication traffic, throughput and path information; Individual flow graph construction units are used to construct a graph structure of network nodes and communication paths, where nodes represent hosts or computing nodes in the network, edges represent communication flow paths between nodes, and edge weights represent the data flow throughput between nodes; The traffic dispersion graph generation unit is used to merge multiple individual flow graphs to generate a traffic dispersion graph representing the overall topological structure of the network, including: using an adjacency matrix to represent the connection relationship between nodes; and identifying abnormal nodes based on the flow threshold of the throughput node.

3. The neural network-based heterogeneous computing power network task allocation and path optimization system according to claim 1 is characterized in that: in, The graph construction module further includes: The graph index and graph displacement construction unit is used to construct the graph index and graph displacement; wherein, The graph index includes computing capacity, load level and throughput, and is used to represent attribute information of nodes; The graph displacement is used to realize the translation operation of the graph index on the graph. The weighted matrix Laplace is selected as the graph displacement to perform weighted combination of node signals and their neighboring nodes, extract the structural features of the graph, and finally realize local smoothing of the graph index.

4. The neural network-based heterogeneous computing power network task allocation and path optimization system according to claim 3 is characterized in that: in, The graph construction module further includes: The graph Fourier transform unit is used to perform graph Laplace eigendecomposition on the weighted matrix Laplace, extract the frequency domain features of the graph index, and then project the graph index into the frequency domain of the graph through graph Fourier transform to analyze the global structure and transmission characteristics between nodes.

5. The neural network-based heterogeneous computing power network task allocation and path optimization system according to claim 1 is characterized in that: in, The improved graph neural network model introduces identity mapping and residual connection, and updates node features through multi-layer iteration to finally obtain the task allocation prediction result of the node.

6. The neural network-based heterogeneous computing power network task allocation and path optimization system according to claim 5 is characterized in that: in, Node feature H (l) The update formula is as follows: Among them, H (l) : node feature matrix of the lth layer; H (0) =X: input initial node feature matrix; μ: identity mapping coefficient, used to balance the initial feature H (0) The influence of ρ: residual connection coefficient, used to introduce the previous layer feature H (l-1) the impact of; The normalized adjacency matrix of the graph, Where A: original adjacency matrix; I: identity matrix; D: degree matrix of the graph; σ: activation function, the expression is: where x' is an input variable; θ1∈(0,∞) and 7. The neural network-based heterogeneous computing power network task allocation and path optimization system according to claim 6 is characterized in that: in, The node task allocation module also includes a unit for defining a node task allocation objective function, which is used to measure node load balance and resource utilization. The objectives include maximizing overall computing efficiency, load balancing, and minimizing network communication overhead.

8. The neural network-based heterogeneous computing power network task allocation and path optimization system according to claim 1 is characterized in that: in, The path optimization module is used to set a path optimization objective function, which is composed of a transmission delay item from any node to other nodes and a load balancing weight item of any node, wherein the load balancing weight item of any node includes the task prediction result of any node, the current task load of any node and the computing power information of any node; minimize the transmission delay and balance the node computing power load, and finally select the optimal data transmission path.

9. The neural network-based heterogeneous computing power network task allocation and path optimization system according to claim 8, characterized in that: in, The path optimization module also performs path optimization based on the frequency characteristics of the graph Fourier transform domain to find the path with the best load balancing, including: Input node load balancing weight and transmission delay; Perform graph Fourier transform on the load balancing weights of nodes; Filter out high-frequency components and retain low-frequency components; Inverse transform back to the node domain to obtain the smoothed node load balancing weight; The optimal path is solved based on the updated node load balancing weights and transmission delays.

10. The neural network-based heterogeneous computing power network task allocation and path optimization system according to claim 9, characterized in that: The system optimization module constructs a loss function for task scheduling and system optimization, which includes a transmission delay loss term and a load balancing loss term, and is used to minimize the sum of delays on the task transmission path and the variance of node loads respectively; The system optimization module also includes: using a gradient descent method to iteratively update weight parameters to optimize the overall performance of the system.

Images