CN-121985379-A - Multi-edge computing task unloading and scheduling layering automatic optimization method based on large language model driving

CN121985379ACN 121985379 ACN121985379 ACN 121985379ACN-121985379-A

Abstract

The invention relates to the technical field of edge calculation and automatic optimization algorithms, and discloses a multi-edge calculation task unloading and scheduling layering automatic optimization method based on large language model driving, which comprises the following steps of S1, constructing a calculation system, acquiring task data and edge server data to form a data set, and dividing the data set into a training set, an optimizing set and a verification set; S2, constructing a combined task unloading and scheduling optimization model taking the total system completion time delay as an optimization target, taking the minimum total system completion time delay as an optimization target function, and taking the total system completion time delay obtained by the combined task unloading and scheduling optimization model as a fitness value. The invention not only realizes the efficient automatic optimization of task unloading and scheduling problems in a multi-edge computing environment, but also improves the intelligent level and generalization capability of the computing system through an algorithm evolution mechanism driven by a large language model, and has important theoretical significance and wide engineering application prospect.

Inventors

Lv Shenhuan
Cui Linkun
CHEN NING
HU SHIHONG
WANG YANYAN
TANG BIN

Assignees

河海大学

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (10)

1. The multi-edge computing task unloading and scheduling layering automatic optimizing method based on the large language model driving is characterized by comprising the following steps of: S1, constructing a computing system, acquiring task data and edge server data to form a data set, and dividing the data set into a training set, an optimizing set and a verification set; S2, constructing a combined task unloading and scheduling optimization model taking the total completion time delay of the system as an optimization target, taking the minimum total completion time delay of the system as an optimization target function, taking the total completion time delay of the system obtained by the combined task unloading and scheduling optimization model as a fitness value, and setting constraint conditions; s3, constructing a layered optimization framework, wherein the layered optimization framework comprises an unloading decision layer and a scheduling ordering layer; S4, performing iterative improvement on the heuristic rule, generating the heuristic rule based on the large language model, inputting a training set into a hierarchical optimization framework, performing iterative improvement on the heuristic rule by adopting a performance feedback mechanism, and generating a final heuristic rule; s5, performing iterative optimization on the hierarchical optimization framework, inputting an optimization set into the hierarchical optimization framework, and performing iterative optimization on the hierarchical optimization framework based on a hierarchical evolution search mechanism and a feasibility checking and repairing mechanism; S6, verifying the hierarchical optimization framework of the iterative optimization by using a verification set; s7, inputting the calculation tasks into the trained hierarchical optimization framework, and generating optimized task unloading and scheduling strategy contents.
2. The method for automatically optimizing task offloading and scheduling hierarchy of multi-edge computing based on large language model driving as recited in claim 1, wherein in step S1, the computing system comprises a task set and an edge server set, Each task comprises a computing resource requirement, a memory requirement and a data scale parameter, and each edge server comprises a computing capacity, a memory capacity and a communication bandwidth parameter; The data set is divided into a training set, an optimizing set and a verifying set according to the proportion of 4:4:2.
3. The method for automatically optimizing the unloading and scheduling of the multi-edge computing task based on the driving of the large language model according to claim 1, wherein in the step S2, the optimization objective function is: wherein Comprises a task transmission delay Waiting time delay in line Actual calculation execution time delay ; The constraint conditions include an edge server computing resource condition constraint, a communication bandwidth condition constraint, a task allocation condition constraint, and a task execution order condition constraint.
4. The method for automatically optimizing task offloading and scheduling layering of multi-edge computing based on large language model driving as claimed in claim 1, wherein in step S3, the offloading decision layer is used for generating candidate allocation schemes, the scheduling ordering layer is used for ordering and optimizing task sets inside each edge server based on the candidate allocation schemes generated by the offloading decision layer, and calculating the internal queuing waiting time and completion time of the edge servers.
5. The method for automatically optimizing task offloading and scheduling hierarchy of multi-edge computing based on driving of large language model according to claim 1, wherein in step S4, a large language model is constructed, the large language model automatically generates new heuristic rules by receiving description information of optimization problem, current optimal solution information and performance information of history function, and generates new heuristic rules to be embedded into hierarchy optimizing framework in executable function form.
6. The method for automatically optimizing the unloading and scheduling layering of the polygonal computing task based on the driving of the large language model according to claim 1, wherein in the step S4, the performance feedback mechanism is performed by adopting the following steps: And performing performance evaluation on the newly generated heuristic rule, and utilizing the total completion time delay of the joint task unloading and scheduling optimization model system to jointly form a performance evaluation index of the heuristic rule by using the average fitness value and the feasible solution proportion, and feeding back an evaluation result to the large language model as 'history performance' information in the next round of iterative optimization process to realize iterative optimization on the heuristic rule.
7. The method for automatically optimizing task offloading and scheduling layering based on a large language model driving is characterized in that in the step S5, the content of a hierarchical evolution search mechanism is that a task allocation scheme is subjected to population searching at an offloading layer, candidate content is generated through selection, intersection and mutation operations; The feasibility checking and repairing mechanism comprises the steps of performing constraint checking and repairing in each generation of iterative process, and ensuring that resource constraint is satisfied.
8. The method for automatically optimizing the unloading and scheduling of the polygonal computing task based on the driving of the large language model according to claim 1, wherein in the step S5, the step of the hierarchical evolution search mechanism is as follows: S11, initializing a population: randomly generating a task allocation matrix in an unloading decision layer to form an initial population, and enabling each task individual to meet the task unique allocation constraint; S12, selecting: sorting the task individuals according to the fitness value, and reserving the excellent task individuals by adopting a tournament selection or roulette selection mechanism; S13, crossing and mutation operation: The method comprises the steps of adopting a cross operator based on task subset exchange and adopting a mutation operator based on edge server reassignment in an unloading layer, wherein the mutation probability and the cross probability can be dynamically adjusted according to an iteration stage; and at the scheduling and sorting layer, reconstructing a task execution sequence according to heuristic rules generated by the large language model.
9. The method for automatically optimizing multi-edge computing task offloading and scheduling layering based on large language model driving as claimed in claim 1, wherein in step S5, feasibility checking and repairing mechanism adopts feasibility priority principle, according to computing resource amount of edge server, according to load size of task individual, reallocating to lowest load server, according to communication bandwidth traffic, preferentially adjusting task individual with maximum data amount, and applying punishment item to unrepairable task individual to reduce adaptability.
10. The method for automatically optimizing task offloading and scheduling hierarchy of multi-edge computing based on large language model driving as claimed in claim 1, wherein in step S7, an optimal task offloading and scheduling policy is outputted, and after reaching a preset iteration number or convergence condition, an optimal task allocation matrix and a corresponding internal execution sequence of the edge server are outputted, and a system performance index is computed.

Description

Multi-edge computing task unloading and scheduling layering automatic optimization method based on large language model driving Technical Field The invention relates to the technical field of edge calculation and automatic optimization algorithms, in particular to a multi-edge calculation task unloading and scheduling layering automatic optimization method based on large language model driving. Background With the rapid development of augmented reality, industrial internet of things, internet of vehicles and low-latency intelligent application, the traditional architecture relying on central cloud computing gradually exposes the problems of serious network congestion, high response delay and large bandwidth consumption. Multi-access edge computing (MEC) has become an important technology path to solve the above problems by sinking computing power to the network edge, providing low latency, highly reliable computing services for terminal devices. In computing systems, task offloading and task scheduling are two core issues that determine the overall performance of a computing system. Task offloading is used to determine the execution location of each computing task, i.e., which edge server is selected for processing, and task scheduling is used to determine the order of execution of multiple tasks within the server. Both affect the end-to-end latency of the computing system, the resource utilization, and the computing system stability. Especially in complex environments where multiple tasks, multiple servers and multiple resources are constrained to coexist, task offloading and scheduling problems often appear as strongly coupled optimization problems. The prior art mainly adopts heuristic rules designed based on manual experience, such as minimum delay priority allocation, minimum operation priority scheduling and the like, search methods based on a group intelligent optimization algorithm (such as a particle swarm algorithm, a genetic algorithm and the like), and in recent years, a deep reinforcement learning method is also used for constructing a state-action mapping model to realize self-adaptive decision. However, the above method still has the following drawbacks and disadvantages in practical applications: 1. The generalization capability is limited due to the fact that the traditional heuristic algorithm is highly dependent on a scoring function or a priority rule of the manual design, and parameter adjustment optimization is often carried out aiming at a specific scale or a specific load scene. When the scale of the computing system, the task characteristics or the resource distribution change, the original rule is difficult to maintain stability, and the cross-instance self-adaption capability is lacked. 2. Under the condition of multi-resource hard constraint, the partial optimization algorithm processes the problem of resource overrun through a punishment function, but an infeasible solution is easy to generate in a high-dimensional discrete space, and the feasibility rate is obviously reduced in a large-scale scene. The presence of an infeasible solution not only affects the stability of the computing system, but also reduces the actual engineering usability of the algorithm. 3. The method for achieving the collaborative optimization of the upper and lower layer decision-making is lack of computational system level collaborative optimization, wherein the existing method is used for processing task unloading and server internal scheduling separately, a fixed scheduling strategy or a simple combination mode is adopted for joint application, a unified modeling and collaborative evolution mechanism is lack, and end-to-end optimization in a real sense is difficult to achieve. 4. The method based on deep reinforcement learning generally requires a large amount of training data and long-time iterative training, and when environmental parameters change, the model is often required to be retrained, the training cost is high, the adaptation period is long, and quick deployment and online adjustment in a dynamic edge environment are not facilitated. Therefore, a new optimization method is needed that can automatically generate high-quality unloading and scheduling strategies under strict resource constraint conditions, has feasibility guarantee capability, supports cross-instance generalization and does not need large-scale pre-training, so as to improve the overall performance and engineering practical value of the computing system. Disclosure of Invention In order to solve the technical problems in the prior art, the invention provides a multi-edge computing task unloading and scheduling layering automatic optimization method based on large language model driving. The invention adopts the following technical scheme that the multi-edge computing task unloading and scheduling layering automatic optimizing method based on large language model driving comprises the following steps: S1, constructing a computi