CN-122019140-A - Cloud edge cooperative resource management and task scheduling optimization method crossing time scales

CN122019140ACN 122019140 ACN122019140 ACN 122019140ACN-122019140-A

Abstract

The invention belongs to the technical field of edge computing and cloud computing, and relates to a cloud edge collaborative resource management and task scheduling optimization method crossing time scales. The invention realizes the long-term optimization of service deployment and resource allocation under a large time scale and the real-time response of task scheduling under a small time scale through a layered reinforcement learning and two-way feedback mechanism, effectively balances the long-term planning and short-term load fluctuation, reduces the frequent reconfiguration overhead and scheduling delay, ensures the resource allocation to take account of performance and cost through explicit modeling budget and physical resource constraint, avoids overload, enhances the suitability of a system to sudden task and server state change, improves the resource utilization rate and task processing efficiency, and provides high-efficiency and stable support for intelligent edge application.

Inventors

LI RUI
ZHONG JING
GUO SHUAI
WU ZHENGZHONG
ZHANG XIAOPENG
ZHOU XINXING
CHENG RAN
LIU DERAN
ZHU SHANGQING

Assignees

中国人民解放军61618部队

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (10)

1. A cloud edge collaborative resource management and task scheduling optimization method crossing time scales is characterized by comprising the following steps of, Dividing the system run time into a large time scale period and a small time scale period, wherein the large time scale period comprises at least one small time scale period; Responding to a system instruction or a user deployment request from a large time scale period, generating a service deployment strategy and a resource allocation strategy by a cloud based on a depth deterministic strategy gradient algorithm according to collected slow-change state data, converting the deployment strategy and the resource allocation strategy into a feasible domain constraint and transmitting the feasible domain constraint to an edge node; and generating and executing a task scheduling strategy according to the collected fast-changing state data by the edge node under the constraint of a feasible domain and based on a near-end strategy optimization algorithm in a small time scale period.
2. The cloud computing resource management and task scheduling optimization method of claim 1, wherein said dividing system runtime into large time scale periods and small time scale periods comprises, When the system is started, initializing computing capacity, storage capacity and resource pricing of configuration edge nodes, and establishing an operation environment of the cloud edge cooperative system; according to the dynamic characteristics of the system, defining a large time scale as a minute-level period, and processing slow-change factors including long-term service demand trend, edge node hardware capacity and deployment cost constraint; The small time scale is defined as a millisecond to minute period, and the processing includes fast-varying factors of task real-time arrival rate, node instantaneous load and network bandwidth fluctuation.
3. The cloud computing resource management and task scheduling optimization method of claim 2, wherein generating a service deployment policy and a resource allocation policy based on the collected slow-varying state data comprises, Periodically collecting slowly-changing state data of the edge node, wherein the slowly-changing state data comprises computing resource budget, storage capacity and bandwidth load; generating a service deployment strategy and a resource allocation strategy by adopting a depth deterministic strategy gradient algorithm; performing hard resource constraint verification on the generated service deployment strategy and resource allocation strategy, and when the service deployment strategy causes the occupation amount of the resource c of any edge node n Satisfies the following conditions When the depth deterministic strategy gradient algorithm rewards functions, negative penalty terms are applied to optimize the strategy generation in the feasible domain, wherein, For a set of services deployed on edge node n, A physical upper limit of the resource c of the edge node n; and converting the service deployment strategy and the resource allocation strategy which pass the constraint verification into feasible domain constraints and transmitting the feasible domain constraints to the edge nodes.
4. The cloud edge collaborative resource management and task scheduling optimization method across time scales according to claim 3, wherein the decision model of the depth deterministic strategy gradient algorithm adopts a hybrid action space design, comprising, Softmax branch output discrete variable through Actor network Determining the mapping relation between the service and the node; Tanh branch output continuous variable through Actor network Representing the calculation resource allocation proportion of each service and ensuring that the sum of the resource proportion is less than or equal to 1; evaluating the long-term value of the decision by using a Critic network, and minimizing the difference between the value estimate and the target value to optimize the policy accuracy; Defining action space of depth deterministic strategy gradient algorithm The method comprises the following steps: ; Wherein, the For a discrete decision of service deployment, For a continuous decision of the allocation of resources, And (3) representing decision time of the cloud depth deterministic strategy gradient algorithm in the current large time scale period as a time step.
5. The cloud computing resource management and task scheduling optimization method of claim 2, wherein generating and executing a task scheduling policy based on the collected fast-varying state data comprises, Receiving feasible domain constraints issued by a cloud, collecting quick change state data, including task arrival rate, node instantaneous load, task queue length and bandwidth state, and constructing state input; generating a task scheduling strategy by adopting a near-end strategy optimization algorithm, wherein the task scheduling strategy has an action space The definition is as follows: ; Wherein, the The probability matrix is M multiplied by N, M is the number of tasks, N is the number of edge nodes, and each row of elements represents the probability that the task M is allocated to the edge node N; through a constraint masking mechanism, nodes of the non-executable tasks are shielded according to the feasible region constraint, and the scheduling strategy is ensured to accord with the cloud constraint; and optimizing a scheduling strategy by on-line fine tuning according to task completion time, queuing delay and scheduling success rate feedback.
6. The cloud computing resource management and task scheduling optimization method across time scales as recited in claim 5, wherein said optimizing scheduling policy by online fine tuning comprises, Dynamically adjusting the exploration rate based on task completion time, queuing delay and scheduling success rate feedback, and adopting an epsilon-greedy strategy to balance exploration and utilization of strategies; when updating the model parameters, only updating the output layer and the attention layer of the strategy network, and freezing the feature extraction layer; And screening out the latest data reflecting the abnormal state or high-priority task from the task completion time, queuing delay and scheduling success rate data fed back by the edge node in real time as a key sample, and updating the parameters of an output layer and an attention layer by adopting a small-batch gradient descent method so as to optimize the adaptability of a scheduling strategy to the sudden load.
7. The cloud computing resource management and task scheduling optimization method across time scales of claim 1, further comprising, Based on the execution feedback of the edge nodes in a small time scale, the deployment strategy generated by the cloud or the scheduling strategy generated by the edge nodes is dynamically adjusted through a bidirectional feedback mechanism to optimize the system robustness and the resource utilization efficiency, and the method specifically comprises the following steps of, Deploying a lightweight monitoring agent at an edge node, sampling the time delay, the scheduling failure rate and the resource saturation of a task completion in real time, triggering a local feasible region breakthrough mechanism when a continuous abnormal state is detected, and temporarily adjusting the resource allocation or migration task; after the load is recovered to be normal or a preset time limit is reached, the edge node tightens up the feasible region, returns the resources and feeds back state recovery information to the cloud; The cloud collects execution feedback of all edge nodes, when the number of nodes triggering local breakthrough in a preset time window exceeds a threshold value, incremental update or retraining of a depth deterministic strategy gradient algorithm is triggered, feedback data related to the local breakthrough of multiple nodes are screened from an experience playback pool, and resource quota rationality and service stability are optimized by combining historical load characteristics; And adjusting the weight of the reward function in the updating process, generating a new strategy adapting to the current environment, and issuing the verification effect in stages.
8. The cloud edge collaborative resource management and task scheduling optimization system crossing the time scale is characterized by comprising a time scale dividing module, a resource management module, a task scheduling module and a bidirectional feedback module; the time scale division module is used for dividing the running time of the system into a large time scale period and a small time scale period, wherein the large time scale period comprises at least one small time scale period; The resource management module is used for responding to a system instruction or a user deployment request started in a large time scale period, generating a service deployment strategy and a resource allocation strategy by the cloud based on a depth deterministic strategy gradient algorithm according to the collected slow-change state data, converting the deployment strategy and the resource allocation strategy into a feasible domain constraint and transmitting the feasible domain constraint to the edge node; The task scheduling module is used for generating and executing a task scheduling strategy according to the collected fast-changing state data by the edge node under the constraint of a feasible domain and based on a near-end strategy optimization algorithm in a small time scale period.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the cross-time scale cloud edge collaborative resource management and task scheduling optimization method according to any one of claims 1-7 when executing the program.
10. A storage medium storing a computer program, wherein the computer program when executed by a processor implements the steps in the cross-time scale cloud-edge collaborative resource management and task scheduling optimization method according to any one of claims 1-7.

Description

Cloud edge cooperative resource management and task scheduling optimization method crossing time scales Technical Field The invention belongs to the technical field of edge computing and cloud computing, and particularly relates to a cloud edge collaborative resource management and task scheduling optimization method crossing time scales. Background With the rapid development of information technology, the proliferation of computation-intensive and time-delay-sensitive applications in the scenes of internet of things, internet of vehicles, industrial internet and the like has put higher demands on computing power and instantaneity. Because the terminal equipment has limited computing and storage resources, complex tasks are difficult to independently process, and the remote cloud server is limited by network delay and bandwidth bottlenecks, so that the low-delay requirement is difficult to meet. The edge calculation is widely applied to scenes with high real-time requirements by sinking the calculation capability to the network edge, effectively reducing delay and relieving bandwidth pressure. However, the edge node has limited resources and strong isomerism, and the task arrival mode is highly dynamic, and the traditional method faces the problems of low resource utilization rate, low deployment efficiency and increased task delay in service deployment, resource allocation and task scheduling. In recent years, some researches propose cross-time scale optimization methods, such as service deployment and resource allocation by using gibbs sampling and an alternating minimization algorithm, and optimizing task scheduling proportion by combining a sub-gradient descent method so as to minimize delay and meet queue stability constraint. However, these methods rely on heuristic algorithms, have low state exploration efficiency, are prone to being locally optimal, have limited response capability to sudden task or server state change, and are difficult to combine long-term planning and short-term real-time requirements. In addition, the existing scheme often ignores budget constraint of the edge server, so that resource allocation is disjointed from actual cost, and problems such as storage overload or computing resource overdropping are caused. Therefore, an efficient cloud edge collaborative optimization method is needed, the problems of insufficient cross-time scale collaboration, lack of budget constraint modeling and poor dynamic environment suitability are solved, and the system performance and the service quality are improved. Disclosure of Invention The invention aims to provide a cloud edge cooperative resource management and task scheduling optimization method across time scales, which aims to solve the problems of insufficient cooperation across time scales, lack of budget constraint modeling and poor dynamic environment suitability in the existing cloud edge cooperative resource management and task scheduling method. In order to achieve one of the above objects, an embodiment of the present invention provides a cloud-edge collaborative resource management and task scheduling optimization method across a time scale, the method comprising, Dividing the system run time into a large time scale period and a small time scale period, wherein the large time scale period comprises at least one small time scale period; Responding to a system instruction or a user deployment request from a large time scale period, generating a service deployment strategy and a resource allocation strategy by a cloud based on a depth deterministic strategy gradient algorithm according to collected slow-change state data, converting the deployment strategy and the resource allocation strategy into a feasible domain constraint and transmitting the feasible domain constraint to an edge node; and generating and executing a task scheduling strategy according to the collected fast-changing state data by the edge node under the constraint of a feasible domain and based on a near-end strategy optimization algorithm in a small time scale period. As a further refinement of an embodiment of the invention, the method further comprises, said dividing the system run time into a large time scale period and a small time scale period comprises, When the system is started, initializing computing capacity, storage capacity and resource pricing of configuration edge nodes, and establishing an operation environment of the cloud edge cooperative system; according to the dynamic characteristics of the system, defining a large time scale as a minute-level period, and processing slow-change factors including long-term service demand trend, edge node hardware capacity and deployment cost constraint; The small time scale is defined as a millisecond to minute period, and the processing includes fast-varying factors of task real-time arrival rate, node instantaneous load and network bandwidth fluctuation. As a further improvement of an embodiment of the prese