Edge computing system task unloading optimization method for multi-modal large model

Abstract

The invention discloses an edge computing system task unloading optimization method for a multi-modal large model. The method comprises the following steps: S1, constructing a relation topological graph by taking an edge server and user equipment as nodes; s2, extracting node features in the relation topological graph through a graph convolutional network to construct a multi-modal task model; s3, adjusting nodes in the multi-modal task model according to a dynamic heuristic algorithm to construct a task scheduling module; s4, the task scheduling module performs distributed coordination processing on the distributed tasks to generate integrated task data; and S5, processing the integrated task data through an optimization adjustment mode to output multi-modal task data, and according to the method, resources are reasonably allocated in a multi-node environment, and the overall operation efficiency of an edge computing system and the stability under a high load are ensured.

Description

Technical Field

[0001] The present invention relates to large model collaborative processing technology, and in particular to an edge computing system task offloading optimization method for multimodal large models. Background Art

[0002] In today's digital and intelligent era, as users' demand for diversified and high-quality services continues to grow, the traditional centralized computing model has gradually exposed the problems of computing resource bottlenecks and delayed service response. Especially in application scenarios involving multimodal data processing and generation, the existing computing architecture is difficult to meet the real-time and personalized needs of users in a distributed environment. Therefore, an innovative solution is urgently needed. Summary of the invention

[0003] In view of the existing technical difficulties, the present invention provides an edge computing system task offloading optimization method for multi-modal large models

[0004] In order to solve the problems of the prior art, the present invention adopts the following technical solutions:

[0005] A method for optimizing task offloading of edge computing systems for multimodal large models comprises the following steps:

[0006] S1, construct a relationship topology graph using edge servers and user devices as nodes;

[0007] S2. Build a multimodal task model by extracting node features from the relationship topology graph through a graph convolutional network;

[0008] S3, adjusting and constructing a task scheduling module for nodes in the multimodal task model according to a dynamic heuristic algorithm;

[0009] S4, the task scheduling module performs distributed coordination processing on the assigned tasks to generate integrated task data;

[0010] S5. Process the integrated task data to output multimodal task data through optimization and adjustment.

[0011] Furthermore, the relationship topology diagram in step S1 includes:

[0012] The graph structure is in: Represents all user devices and edge servers participating in the computation. It represents the connection between nodes, that is, the possible path for task scheduling;

[0013] The node feature is represented by the multidimensional feature of each node; the multidimensional feature is represented by the matrix It is represented by, where N is the number of nodes, F is the feature dimension of each node; the multidimensional features include the node ID, geographical location and computing performance of the server.

[0014] Furthermore, the process of adjusting and constructing a task scheduling module for nodes in the multimodal task model according to the dynamic heuristic algorithm in step S3 includes:

[0015] Adjust each assigned task to the corresponding node according to the real-time status of the node;

[0016] The tasks assigned to the nodes are updated according to inertia, local optimal gravity and global optimal gravity through the following formula;

[0017]

[0018] in:

[0019] Inertia part: Indicates the speed of the current task or node at time t It will be multiplied by an inertia weight ξ;

[0020] Local optimal gravity part: Indicates that the task or node is moving towards its historical best position Trends in mobility; is the best position found so far by the task or node; is the current position of the task or node; the acceleration constant g1 adjusts the strength of the local gravity, and r1 is a random factor;

[0021] Global optimal gravity part: It indicates guiding the task or node to move towards the global optimal position Δ(t); Δ(t) is the current global best solution found by the system among all tasks or nodes. is the position of the current task or node; the acceleration constant g2 controls the strength of the global gravity, and r2 is a random factor.

[0022] Beneficial Effects

[0023] Compared with the traditional technical solution, the beneficial effects brought by the present invention are:

[0024] 1. Scalability and flexibility: In a distributed environment, the present invention demonstrates stronger scalability than the traditional centralized computing model. Whether it is processing high-concurrency user requests or coping with diverse task requirements, the present invention can ensure the efficient operation of the system and adapt to complex computing environments of different scales.

[0025] Accurate modeling of high-dimensional feature representation: By introducing the graph convolutional network (GCN), this method can construct a high-dimensional feature representation of the complex relationships between users, nodes, and models in a distributed environment. This precise modeling greatly improves the accuracy of task allocation, allowing the system to effectively utilize the computing resources of edge nodes and achieve optimal resource allocation. Traditional methods often suffer from low resource utilization and high task delays when faced with complex and diverse task scenarios. The present invention uses advanced task scheduling algorithms to greatly improve resource utilization and reduce task execution delays, allowing the system to maintain smooth operation even under high load conditions.

[0026] 2. Intelligent dynamic task scheduling: The present invention can dynamically adjust the task allocation strategy according to the personalized needs of users, greatly improving the diversity and matching accuracy of services. Compared with traditional methods, it can better meet the needs of users in different scenarios and ensure the efficient execution of personalized services. This method uses a dynamic heuristic algorithm to intelligently adjust the task scheduling strategy according to real-time user needs and system status. This dynamic adjustment not only improves the response speed of task processing, but also optimizes path selection and significantly reduces the delay during task execution, especially showing excellent performance when facing multimodal tasks and complex computing scenarios.

[0027] 3. Efficient implementation of distributed collaborative processing: This method innovatively implements distributed collaborative processing of multimodal generation tasks, and avoids the problem of single node overload by distributing tasks among multiple edge nodes. This processing method not only improves the scalability and fault tolerance of the system, but also ensures the efficiency and stability of the system under high concurrency.

[0028] 4. Globally optimized resource allocation and task scheduling: Compared with existing technologies, this method achieves global optimization in resource allocation and task scheduling, taking into full account multiple factors such as user needs, node performance, and network conditions. Through global optimization, the system can reasonably allocate resources in a multi-node environment, ensuring the overall operating efficiency of the system and stability under high load.

[0029] 5. Flexible framework to adapt to diverse needs: This method demonstrates high adaptability in complex network environments and diverse user needs. Whether in the case of uneven performance of edge nodes or when dealing with personalized needs, this method can flexibly respond, provide efficient and stable computing services, and quickly respond to user needs.

[0030] 6. System stability and extended node life: Through intelligent scheduling, the present invention can balance the load of each node when the performance of edge nodes varies greatly, preventing some nodes from being overloaded or idle for a long time. This not only improves the overall stability of the system, but also extends the service life of each node, reducing the loss of hardware resources and the frequency of replacement. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of a method for optimizing task offloading of edge computing system for multi-modal large model of the present invention;

[0032] Figure 2 This is a workflow diagram of the present invention involving GCN characterization of high-dimensional features;

[0033] Figure 3 The present invention relates to a workflow diagram of a dynamic heuristic task allocation algorithm for optimal path selection;

[0034] Figure 4 This is a comparison chart of the accuracy of the dynamic heuristic task allocation algorithm of the present invention and other algorithms when processing different task data volumes. Figure 5 This is a comparison chart of the efficiency of the dynamic heuristic task allocation algorithm of the present invention and other algorithms when processing different task data volumes. DETAILED DESCRIPTION

[0035] The following is combined with Figure 1 -Attached Figure 5 The present invention is described as follows:

[0036] like Figure 1 As shown, the method of the present invention is based on a graph convolutional network (GCN) and a dynamic heuristic task scheduling algorithm, aiming to optimize the processing flow of multimodal generation tasks in a distributed edge computing environment.

[0037] Step 1: In order to accurately describe the connections and relationships between nodes, the system first starts by building a graph structure for the entire distributed computing environment. The definition of the graph structure is the basis of the entire method and provides a framework support for subsequent high-dimensional feature representation and task scheduling.

[0038] 1. Definition of graph structure: In a distributed computing environment, the system first defines a graph

[0039] in: represents all nodes involved in the computation (including user devices and edge servers), It represents the connection between nodes (i.e. the possible path for task scheduling). Through this graph structure, the system can fully describe the relationship between different computing nodes.

[0040] 2. Representation of node features: Each node has multi-dimensional features, which are represented by matrices

[0041] It is represented as: N is the number of nodes, and F is the feature dimension of each node. These features include the node ID, geographic location (longitude and latitude), server computing performance (such as CPU and memory), and the type and number of tasks currently being processed. The representation of these features lays the foundation for subsequent feature modeling and task scheduling.

[0042] Next, we enter the modeling stage of feature representation.

[0043] Step 2: In order to deeply understand the roles and relationships of nodes in the network, after defining the graph structure and representing the node features, the system further uses the graph convolutional network (GCN) to model these features, such as Figure 2 This process aims to reveal and capture the complex relationships between nodes, thereby providing a basis for efficient task allocation.

[0044] 1. Forward propagation of graph convolution: Through the graph convolution network, the system can map node features into a high-dimensional space so that the features of adjacent nodes can influence each other. Specifically, the forward propagation process of GCN is expressed by the following formula:

[0045] H (1) =σ(D -1 / 2 AD -1 / 2 XW (0) )

[0046] Among them, H (1) represents the node embedding representation of the first layer, W (1) is the weight matrix of layer 1, σ is the activation function, and D is the degree matrix.

[0047] 2. Modeling of node relationships: Through GCN, the system can effectively capture the complex relationships between nodes, including the interdependence between adjacent nodes and the potential influence between remote nodes. This high-dimensional feature representation can provide more accurate information support for subsequent task scheduling.

[0048] After completing high-dimensional feature modeling, the system enters the task scheduling stage to ensure the reasonable distribution of tasks among nodes.

[0049] Step 3: In order to efficiently utilize distributed resources to process complex tasks, after the system has established a high-dimensional feature representation, the first step in the task scheduling phase is to initialize the task and design a reasonable scheduling strategy. This strategy will formulate a preliminary plan for task allocation based on the information in the feature representation.

[0050] 1. Task initialization: In a distributed environment, tasks submitted by user devices are identified by the system as a series of pending computing tasks. These tasks may involve the generation of multimodal data, such as text, images, and videos. The system determines their computing requirements and priorities by analyzing the characteristics of these tasks.

[0051] 2. Scheduling strategy design: Based on the previous high-dimensional feature modeling, the system designed a set of preliminary task scheduling strategies. This strategy comprehensively considers the current load, computing power, network latency and user needs of the node to ensure that each task can be assigned to the most suitable node for processing. This strategy provides basic guidance for subsequent task scheduling.

[0052] Next, the system enters the actual task scheduling and execution phase.

[0053] Step 4: To ensure that each subtask can be executed by the optimal node, after completing the preliminary scheduling strategy design, the system performs task scheduling through a dynamic heuristic algorithm and optimizes the scheduling scheme in real time according to the actual operation situation; Figure 3 shown.

[0054] 1. Task scheduling and execution: The system assigns each task to the appropriate node based on the scheduling strategy. During this process, the system will consider the real-time status of the node and make corresponding dynamic adjustments. For example, if a node is overloaded, the system will reallocate some tasks to other nodes.

[0055] 2. Dynamic optimization: The system uses a dynamic heuristic algorithm to optimize the task scheduling process. Specifically, the algorithm continuously adjusts the task allocation scheme through the following speed update formula:

[0056]

[0057] This formula describes the speed update process in the dynamic heuristic task scheduling algorithm, which involves three key parts: inertia, local optimal gravity, and global optimal gravity. First, the inertia part Indicates the speed of the current task or node at time t It will be multiplied by an inertia weight ξ, which controls the impact of the previous speed on the current speed, ensuring that the task or node maintains a certain continuity in the solution space.

[0058] Locally optimal gravity part Indicates that the task or node is moving towards its historical best position The trend of movement. Here, is the current position of the task or node, is the optimal position found so far by the task or node. The acceleration constant g1 adjusts the strength of this gravitational force, while the random factor r1 (usually between 0 and 1) introduces a certain amount of randomness, making the search process diverse and avoiding premature convergence to the local optimum.

[0059] Finally, the global optimal gravity part Then the task or node is guided to move towards the global optimal position Δ(t). Δ(t) is the current global best solution found by the system among all tasks or nodes. is the position of the current task or node. Similarly, the acceleration constant g2 controls the strength of global gravity, and the random factor r2 provides some randomness to enhance the breadth and flexibility of the search.

[0060] Through these adjustments, the system can continuously optimize task allocation based on real-time operating data and improve the efficiency of the entire system.

[0061] After the tasks are successfully assigned and executed, the system needs to integrate and feedback the results.

[0062] Step 5: To ensure that tasks are completed efficiently according to dependencies, after completing task scheduling, each edge node starts to execute the assigned tasks and completes the entire multimodal generation task through collaborative work. The system then integrates the calculation results of each node and feeds them back to the user.

[0063] 1. Distributed collaborative processing: Different edge nodes collaboratively process the tasks assigned to them, and the system ensures that these nodes can communicate and coordinate with each other to avoid resource waste and computing conflicts. Through this collaboration, the system not only improves the processing efficiency of tasks, but also enhances the system's fault tolerance and scalability.

[0064] 2. Result integration and feedback: After all tasks are completed, the system integrates the calculation results of each node to form the final output result. After the integration is completed, the system will feedback the results to the user and record the relevant data of task execution to further optimize future task scheduling strategies.

[0065] Step 6: Provide users with consistent and high-quality multimodal outputs. Finally, the system continuously learns and optimizes itself by analyzing the data collected during task execution. This step ensures that the system can schedule tasks more accurately and efficiently over time.

[0066] 1. Data analysis and learning: The system analyzes the execution data of each task, including the task execution time, resource utilization, user feedback, etc. Through these data, the system can identify deficiencies in the scheduling strategy and perform targeted optimization.

[0067] 2. Policy update: Based on the results of data analysis, the system automatically updates the scheduling strategy to ensure that future tasks can be processed more efficiently. This self-optimization mechanism enables the system to continuously evolve with changes in the external environment and maintain leading computing performance.

[0068] Through the above six steps, the method of the present invention can comprehensively optimize the processing flow of multimodal generation tasks in distributed edge computing, ensuring the efficiency, stability and adaptability of the system. Each step is closely linked to form a complete technical solution, which effectively solves the problems of computing resource management and task scheduling in a distributed environment.

[0069] Server configuration and actual effect: In the experiment, the present invention was deployed on a server equipped with an Intel(R) Xeon(R) Platinum 8255C CPU with 12 vCPUs, 128GB RAM, and 1TB NVMe SSD storage. The experimental setup includes 31 nodes representing edge servers to simulate the interaction mode in a complex network environment. Under this configuration, the present invention effectively improves the resource utilization of each server, achieving an average task delay reduction of about 30%, and under high load conditions, the task execution time is shortened by about 20% compared to traditional methods.

[0070] For example, in a task for multimodal data processing, the present invention makes full use of the computing resources of each server through reasonable task allocation. Experimental results show that even when some servers are under high load, the system can still maintain a stable response time, avoiding the problem of server overload or idleness, thereby greatly improving the stability and operation efficiency of the entire distributed system.

[0071] Although the present invention has been described above, the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many modifications without departing from the purpose of the present invention, which are all within the protection of the present invention.

Claims

1. A method for optimizing task offloading in edge computing systems for multimodal large models, characterized in that: The steps include: S1, construct a relationship topology graph using edge servers and user devices as nodes; S2. Build a multimodal task model by extracting node features from the relationship topology graph through a graph convolutional network; S3, adjusting and constructing a task scheduling module for nodes in the multimodal task model according to a dynamic heuristic algorithm; S4, the task scheduling module performs distributed coordination processing on the assigned tasks to generate integrated task data; S5. Process the integrated task data to output multimodal task data through optimization and adjustment.

2. According to claim 1, a method for optimizing task offloading of edge computing systems for multimodal large models is characterized in that: The relationship topology diagram in step S1 includes: The graph structure is in: Represents all user devices and edge servers participating in the computation. It represents the connection between nodes, that is, the possible path for task scheduling; The node feature is represented by the multidimensional feature of each node; the multidimensional feature is represented by the matrix It is represented by, where N is the number of nodes, F is the feature dimension of each node; the multidimensional features include the node ID, geographical location and computing performance of the server.

3. The method for optimizing task offloading of edge computing systems for multimodal large models according to claim 1, characterized in that: The process of adjusting and constructing a task scheduling module for nodes in the multimodal task model according to the dynamic heuristic algorithm in step S3 includes: Adjust each assigned task to the corresponding node according to the real-time status of the node; The tasks assigned to the nodes are updated according to inertia, local optimal gravity and global optimal gravity through the following formula; in: Inertia part: Indicates the speed of the current task or node at time t It will be multiplied by an inertia weight ξ; Local optimal gravity part: Indicates that the task or node is moving towards its historical best position Trends in mobility; is the best position found so far by the task or node; is the current position of the task or node; the acceleration constant g1 adjusts the strength of the local gravity, and r1 is a random factor; Global optimal gravity part: It indicates guiding the task or node to move towards the global optimal position Δ(t); Δ(t) is the current global best solution found by the system among all tasks or nodes. is the position of the current task or node; the acceleration constant g2 controls the strength of the global gravity, and r2 is a random factor.

Images