A cloud edge-end resource scheduling optimization method based on a double-layer graph neural network

Abstract

The application provides a cloud edge terminal resource scheduling optimization method based on a double-layer graph neural network, comprising the following steps: constructing a cloud edge terminal resource scheduler based on a double-layer graph neural network, a first layer of graph neural network is used to realize cloud edge terminal resource allocation, a second layer of graph neural network is used for industrial terminal layer scheduling, a heterogeneous graph is used to represent the state in the actual manufacturing environment, the complex relationship between workpieces, processes and available equipment is captured, a directed acyclic graph is constructed as the input of the second layer of graph neural network; a framework model for scheduling and unloading in stages is constructed, and the framework model for scheduling and unloading in stages adopts an optimization strategy to optimize task scheduling. The application can realize rational utilization of resources of the cloud, the edge terminal and the industrial terminal under the background of intelligent manufacturing, perform rapid preparation of resource unloading, improve the comprehensive scheduling capability of industrial production, improve the resource utilization rate and realize green scheduling.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent manufacturing, and is particularly related to a cloud-edge-end resource scheduling optimization method based on a double-layer graph neural network. BACKGROUND

[0002] In the process of upgrading and transformation of the manufacturing industry, the deployment and application of a large number of intelligent sensing devices, and the transformation of personalized production modes, have resulted in a large amount of real-time data generated by intelligent factories, forming a more complex intelligent production mode of small batches, multiple varieties, and personalization, which has put a huge pressure on the network resources, computing resources, and storage resources of the big data center. The processing and scheduling of industrial terminals cannot fully meet the current digital business computing power requirements, so in order to adapt to differentiated business needs, it is necessary to reasonably utilize the powerful computing capabilities of the cloud, and thus the "cloud-edge-end" collaborative concept is introduced into the field of intelligent manufacturing. Meanwhile, under the intelligent production mode, the products on the intelligent factory are constantly enriched, the order demand is constantly changing, and the probability of dynamic emergencies is greatly increased, which interferes with the production process of the enterprise and has a negative impact on product production and factory operation, so the study of cloud-edge-end collaborative resource scheduling needs to consider the dynamic state, and thus a graph neural network is introduced for combination.

[0003] Currently, to solve the problem of cloud-edge-end resource optimization, scholars have applied an improved differential artificial bee colony algorithm [1] to solve the problems of data processing not being timely due to edge backlog, uneven load, and waste of resources caused by scheduling data to the cloud computing center, thereby improving the overall performance of the resource scheduling system. The resource scheduling is optimized, but the dynamic environment resource optimization problem is not solved, and there is a lack of collaborative resource integration, resulting in a large amount of repeated construction of resources, which limits the green transformation of the intelligent manufacturing field. Wang Shuling et al. [2] For two typical cloud-edge collaborative scenarios, the scene is analyzed in sequence from splitting, scheduling target, and solution scheme, and a resource scheduling optimization reference scheme suitable for the characteristics of the scene is given. However, this scheme still does not maximize the utilization of the cloud, edge, and industrial terminal, does not solve the scheduling problem in the dynamic data update process, and the method is not suitable for real-time data update scheduling problems, and does not realize reasonable mapping to the actual production environment.

[0004] [1] Zhao Hongwei, Jing Xuehui, Zhang Shuai, Ruan Ying, Zhang Ziqi. Resource scheduling optimization strategy for cloud-edge collaboration [J]. Journal of Shenyang University (Natural Science Edition), 2021, 33(01): 41-46+74.

[0005] [2] Wang Shuling, Sun Jie, Wang Peng, Yang Aiding. Resource scheduling optimization in cloud-edge collaboration [J]. Telecommunication Science, 2023, 39(02): 163-170. SUMMARY

[0006] To at least solve one of the problems existing in the prior art, the present application provides a cloud edge end resource scheduling optimization method based on a double-layer graph neural network, which can realize rational utilization of resources of the cloud end, edge end and industrial terminal under the background of intelligent manufacturing, quickly prepare resource offloading, improve the comprehensive scheduling capability of industrial production, improve resource utilization, realize green scheduling, improve the real-time performance of data processing, improve the utilization rate of cloud edge end resources, and reduce overall energy consumption to cope with and handle uncertain events when the complex dynamic environment and demand change.

[0007] To achieve the purpose of the present application, the cloud edge end resource scheduling optimization method based on a double-layer graph neural network provided by the present application comprises the following steps:

[0008] A cloud edge end resource scheduler based on a double-layer graph neural network is constructed, the first layer of graph neural network is used to realize cloud edge end resource allocation, the cloud edge layer input graph represents the relationship between the servers of the cloud end and the edge end, the graph is established by connecting to each vertex related to one of the edges, the cloud edge network servers and the industrial terminals establish graphs with each other, a heterogeneous graph is constructed, the relationship among the cloud end, edge end and industrial terminal is captured, used as the input of the first layer of graph neural network, to realize the cloud edge end resource allocation problem, and realize the allocation of computing tasks; the second layer of graph neural network is used for industrial terminal layer scheduling, a heterogeneous graph is used to represent the state in the actual manufacturing environment, the complex relationship between workpieces, processes and available devices is captured, a directed acyclic graph is constructed, used as the input of the second layer of graph neural network, to perform task allocation and scheduling, to determine the order of allocating tasks, then the tasks are represented as node embedding graphs, and a reinforcement learning strategy is used to study the plan;

[0009] According to the cloud edge end resource scheduler based on a double-layer graph neural network, the computing resource allocation problem of the cloud end, edge end and industrial terminal is considered, a framework model for scheduling offloading in stages is constructed, tasks with large amount of computation and high task complexity are offloaded to the edge, the remaining tasks with small amount of computation and low task complexity are directly offloaded to the local terminal machine, and the graph neural network scheduler is used for unified scheduling.

[0010] The framework model for scheduling offloading in stages adopts an optimization strategy to optimize the scheduling of tasks, the optimization strategy is a three-stage adaptive hybrid heterogeneous resource allocation and task scheduling method in a dynamic multi-task mode, the first stage of the three-stage adaptive hybrid heterogeneous resource allocation and task scheduling method in a dynamic multi-task mode is task preordering, the second stage is resource preallocation, and the third stage is dynamic task scheduling, so as to coordinate the richness of cloud computing resources and the low transmission delay of local resources and edge resources, and optimize the scheduling result.

[0011] Further, in the cloud edge end resource scheduler based on the double-layer graph neural network, a state feature embedding module is constructed to obtain task state information such as workpieces, machines and processes from the input order information of the cloud end. For the occurrence of dynamic events, a directed acyclic graph (DAG) is used to describe the dynamic state. Specifically, multiple DAGs are used to represent the task set, the remaining minimum scheduling time is derived using EFT, the urgency and priority of each DAG are calculated, a two-dimensional sorting tuple is obtained, and the urgent DAG is preferentially processed during scheduling. When a dynamic change occurs, the new arriving task set does not interrupt the urgent task. For example, when a new DAG arrives, its priority is initialized to the average DAG priority and is added to the schedule after allocation, without preoccupying the resources of the almost completed DAG task. The task to be scheduled shares resources with all tasks from other DAGs. The tasks in all DAGs are scheduled one by one, and the sorting tuple is updated, and the scheduled DAG is deleted from the scheduling set in time. Then, the scheduling state described by the DAG is converted into a heterogeneous graph state (HetG), and then a double-layer heterogeneous graph neural network (2-GNN) model is used to represent the learning of the heterogeneous state graph, so as to obtain the node state feature embedding of the order information.

[0012] Further, the phased scheduling and unloading framework model comprises receiving order information as a computing task at the cloud edge, the edge comprises a plurality of edge servers, each server has a group of virtual machines (VMs), and the cloud server also comprises a group of virtual machines, both of which are rented on demand.

[0013] Further, an optimization strategy is adopted, and the computing task is first allocated to a heterogeneous platform. Multiple tasks are coupled by scheduling to adjust the execution sequence. If the resource allocation generates a good object library for scheduling, the uniformly distributed tasks will not accumulate in the platform. The overall effect of the scheduling can be modeled in the form of a directed acyclic graph, and the final scheduling scheme is determined by multiple directed acyclic graph scheduling. The scheduling is performed by inserting the idle state, and only when the allocation decision and the scheduling decision are generated at the same time, the task is dispatched to different virtual machines (VMs).

[0014] Further, a three-stage adaptive hybrid heterogeneous resource allocation and task scheduling method in a dynamic multi-task mode coordinates the richness of cloud computing resources and the low transmission delay of local resources and edge resources, and optimizes the scheduling result. The specific three-stage description is as follows:

[0015] The first stage is task pre-scheduling. EDF is used to group and schedule tasks. The tasks with the earliest deadline are iteratively assigned to the first virtual machine that can meet their requirements. EDF is better than FF, FFD, BF and EFTF in terms of device SLA optimization because EDF prioritizes the requirements of tasks with tighter deadlines, thus postponing tasks with more slack time. After determining the tasks to be assigned, an initial task scheduling group is generated. The order of task execution needs to be adjusted by the scheduling after resource allocation. The scheduling scheme is adjusted according to the constraints, objectives, resources and system environment. Compared with other methods, more tasks with tighter deadlines can be completed.

[0016] The second stage is resource pre-allocation. The cloud-edge-end resource pre-allocation mechanism is used. Emergency tasks are assigned to the virtual machine with the earliest completion time using EFTF. Only the completion time is considered, and non-emergency tasks are assigned to resources with lower cost and slower speed. Therefore, in the resource pre-allocation stage, data is cached to the edge server in advance, especially in the cloud, which can improve the performance of task execution. For the caching mechanism, considering the priority restrictions after the first stage of task allocation, multiple idle time periods will be formed on the heterogeneous platform. The earliest completion time of the previous sub-task of the same processing task and the earliest start time of the same processing task group have an idle time period, which can be used. Therefore, the idle time period can be used to cache data in advance. Therefore, the idle time period allocation method is used in resource pre-allocation to reduce idle time slots to achieve full resource utilization. First, all tasks that can be completed by cloud resources are pre-allocated to the cloud. For each task, a virtual machine with the best cost performance is pre-rented. Then, if there are no tasks that can be completed in the cloud, the remaining tasks are pre-allocated to the edge server using LSTF and EDF. Pre-allocation can improve by re-allocating some tasks from local devices or edges to the cloud, leaving some resources idle to complete the remaining unallocated tasks. Assigning tasks that can be completed by the cloud and local devices or edges in the first time to the cloud generates more available local and edge resources for processing tasks that can only be met by local devices or edges. In this way, the resource utilization is improved. (The main task is to fully utilize local resources, release part of the edge resources, re-schedule multiple tasks from the edge to the corresponding device, and use multiple user-shared edge resources to complete more tasks.)

[0017] Third stage task scheduling. Implement all the task allocations described above in the second stage on each device, edge and cloud using a task scheduling algorithm, which decides which core processes the task. Check each task allocated to the cloud and if the task can be done by the corresponding device, re-allocate the task to the device, otherwise, check if the requirements of the task can be met by the edge server and if so, re-allocate the task to the edge server. After these re-allocations, re-allocate the tasks allocated to the next computing virtual machine to the previous computing virtual machine by using the task scheduling algorithm and re-rent the idle virtual machines. Reduces the idle time of virtual machines and improves cost efficiency.

[0018] Implement all the task allocations described above on each device, edge and cloud using a task scheduling algorithm, which decides which core processes the task.

[0019] Compared with the prior art, the application can achieve the beneficial effects at least as follows:

[0020] (1) The double-layer graph neural network scheduler proposed by the method uses a directed acyclic graph to represent dynamic state information, which can better realize the interaction process between the cloud edge end environment and the actual environment, complete physical mapping, and reduce the difference between the model expression and the actual problem. At the same time, when dynamic environment changes and demand changes and other dynamic events come, it can be adaptively and quickly adjusted to assist in improving the cloud edge end resource scheduling efficiency.

[0021] (2) The three-stage resource allocation method provided by the application has the following advantages: a task classification is performed before pre-allocation, the data cache idea is introduced in pre-allocation, and the subsequent resource allocation is more clear for the entire scheduling task group, that is, whether to allocate to the cloud or the edge, and all data does not need to be uniformly and centrally processed. The task grouping and sorting in the first stage is more advantageous for urgent tasks, which does not need to wait for the previous tasks to be processed before processing, thereby speeding up the processing of more urgent tasks. The purpose of the second stage pre-allocation corresponds to the overall resource optimization goal, which ensures high resource utilization and reduces the total processing time. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 It is a three-stage adaptive hybrid heterogeneous resource allocation and task scheduling architecture diagram for the cloud edge end dynamic multi-task mode in the embodiment of the application.

[0023] Figure 2 It is a double-layer graph neural network in the embodiment of the application.

[0024] Figure 3 It is a GNN graph of K embedding layers in the embodiment of the application. DETAILED DESCRIPTION

[0025] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts are within the protection scope of the present application.

[0026] Referring to Figures 1-2 The present application provides a cloud edge end resource scheduling optimization method based on a double-layer graph neural network, comprising the following steps:

[0027] Step 1: a cloud edge end resource scheduler based on a double-layer graph neural network is proposed, a state feature embedding module is constructed to obtain task state information such as workpieces, machines and processes from input cloud order information, and directed acyclic graphs (DAGs) are used to describe dynamic states in response to occurrence of dynamic events.

[0028] The overall description is as follows: a plurality of DAGs are used to represent a task set, the earliest finish time priority (EFT) is used to derive a minimum remaining scheduling time, the urgency and priority of each DAG are calculated, a two-dimensional sorting tuple is obtained, and the urgent DAG is preferentially processed during scheduling. When a dynamic change occurs, it is reflected on the DAG, and a newly arrived task set does not interrupt an urgent task. For example, when a new DAG arrives, its priority is initialized to the average DAG priority and is added to the schedule after allocation, and it does not preempt the resources of a nearly completed DAG task. The task to be scheduled shares resources with all tasks from other DAGs. The tasks in all DAGs are scheduled one by one, and the sorting tuple is updated, and the scheduled DAG is deleted from the scheduling set in time. Then, the scheduling state expressed by the DAG is converted into a heterogeneous graph state (HetG), and a double-layer heterogeneous graph neural network (2-GNN) model is used to represent the learning of the heterogeneous state graph, so as to obtain the node state feature embedding of the order information.

[0029] The first layer graph neural network is used for cloud edge layer resource allocation. Referring to Figure 2 The cloud edge layer input graph represents the relationship between the servers of the cloud, edge and industrial terminals, the graph is established by connecting to each vertex related to one of the edges, the graphs between the cloud edge network servers and the industrial terminals are established, the heterogeneous graph is constructed, the relationship between the cloud, edge and industrial terminals is captured, thereby serving as the input of the first layer graph neural network, to realize the cloud edge end resource allocation problem and realize the allocation of computing tasks.

[0030] The second layer graph neural network is used for industrial terminal layer scheduling, and the state in the actual manufacturing environment is represented by a heterogeneous graph, so as to capture the complex relationship between workpieces and available equipment. The solution of the job shop scheduling problem is modeled as a heterogeneous graph model, which contains rich arc information between multiple types of nodes and unstructured content related to each node (the unstructured content refers to the heterogeneity of different node contents, and the correlation between node attributes is established by encoding). The heterogeneous graph (HetG) node types include machines, workpieces and processes, and the relationship types include machine-production-workpiece, workpiece-contains-process and machine-processing-process. A workpiece contains various processes, the same type of workpiece can be placed on multiple machines of the same type for processing, and machines, workpieces and processes are related to each other. Therefore, a new state feature graph structure G t =(V,M,C,E,O V ,C E ) is defined, V is a node of multiple operation types (here, the operation type refers to a processing process, in some embodiments of the present application, such as feeding, discharging, covering, discharging, carving, marking, packaging, etc.), which includes all operations and two virtual operations representing the start and end of production (the processing time is zero), the machine node is M, each node corresponds to a machine Mk, C is a set of conjunctive arcs, which are directed arcs forming n paths from the start to the end, representing the various processing processes of the job. The disjunctive arc set E is an undirected arc, which is a link connecting the operation node and the compatible machine node. O V represents a set of operation object types, C E represents a set of relationship types, and each node is associated with heterogeneous content. The process correlation constraint is also added, the set of conjunctive arcs C is enriched, and the tuple form of the arc weight is obtained by considering the transmission time between processes, machine processing time and delivery period, so as to represent the state information of the graph node.

[0031] In an actual intelligent factory, the size of the scheduling instance is inconsistent, which causes the size of the state graph to change. In order to facilitate the use of deep reinforcement learning to obtain the actual scheduling strategy, the neural structure must be able to operate on state graphs of different sizes. In order to represent the learning of the heterogeneous graph, the present application proposes a double-layer graph neural network model to obtain the state feature embedding of the nodes in the heterogeneous graph. In some embodiments of the present application, the double-layer graph neural network model is as follows Figure 2As shown, first, the nodes in the heterogeneous graph HetG are sampled, grouped according to node types, and then aggregated using a graph neural network structure. The network structure fully considers the topological information and numerical information (original features) of the graph, effectively encodes the disjunctive graph. The steps include feature extraction, feature conversion using hot encoding, and finally, feature pooling using a pool function L to calculate the global embedding graph from the input graph for feature pooling, which facilitates subsequent feature aggregation.

[0032] The feature information of the neighbor nodes sampled in the previous step includes machine node embedding and operation node embedding. The network structure fully considers the topological information and numerical information (original features) of the graph, effectively encodes the disjunctive graph. The graph node embedding method considers both node and edge features, preserving different types of relationship features in the graph. Meanwhile, node embedding can be considered as a feature vector, and these information can perform various final tasks. For a graph G = (V, E), node embedding is calculated using iterative application of embedding. The p-dimensional embedding of node v ∈ V is calculated by applying k embedding layers, each of which represents the relationship between different nodes in the graph. The node v at iteration k is represented as The calculation is as follows: where is a multi-layer perceptron, is the previous node embedding, used to distinguish a particular node v in the graph from other nodes, is the initial node embedding, used to consider the initial input features of the target node. ∈ is an arbitrary number, and N(v) is the neighborhood of V, represents the node u at iteration k-1. A GNN is constructed by stacking k embedding layers, and the GNN has K embedding layers, as shown in Figure 3 As shown, the circles represent operations, the straight edges and the dashed edges represent connection edges and disjunctive edges respectively, and the source nodes of the gray arc arrows in embedding layers 1, 2 and 3 are 1, 2 and 3-hop neighbors of the target node v respectively. After k times of embedding iteration, a pool function L can be used to calculate the p-dimensional vector of the global embedding graph G from the input graph Since solving the dynamic scheduling problem is equivalent to selecting a disjunctive arc for each node and fixing the direction, that is, the disjunctive graph G t associated with each state s t is a mixed graph with directed arcs, which describe key features such as priority constraints and operation sequences on machines. The original GNN structure is used for undirected graphs, and the structure of the GNN needs to be adjusted for processing. In the initial state, the undirected disjunctive arcs are ignored, and then an approximately complete directed arc GD t = (V, C∪D u ) is added. udenotes the set of disjunctive arcs, as the dispatching process proceeds, D u increases, meaning that more directed arcs are added to the set of approximate directed arcs, the graph does not become too dense. At this point, the neighborhood of a node v is defined as: N(v) = {u | (u, v) E(GD t}, where E is the set of arcs of the graph. Finally, the state s t The original feature of each node v E V is defined as The node embedding obtained after k iterations is denoted as The embedding of the graph is denoted as h G (s t For each node, two binary indicators I(v, s t ) and an integer C(v, s t ) are used to represent the node feature, the integer represents the lower bound of the estimated completion time of the node at state s t .

[0033] Step 2: Build a phased scheduling and offloading framework model, consider the computational complexity and task complexity, offload tasks with large computational complexity and high task complexity to the edge, and uniformly schedule through the graph neural network scheduler; the remaining tasks with small computational complexity and low task complexity are directly offloaded to the local terminal machine.

[0034] Please refer to Figure 1 , the phased scheduling and offloading framework model includes receiving order information as a computing task at the cloud edge, the edge includes multiple edge servers, each edge server has a group of virtual machines (VM), and the cloud server also includes a group of virtual machines, the virtual machines on the cloud server and the edge server are rented on demand. An optimization strategy is adopted, first, the computing task is allocated to a heterogeneous platform, and multiple tasks are adjusted by scheduling as coupling (multiple tasks are allocated to different servers for execution, and the association between multiple tasks, i.e., coupling, is established to dynamically adjust the execution of the remaining tasks) to adjust the execution sequence. If the resource allocation generates a good object library for scheduling, the uniformly distributed tasks will not accumulate in the platform. The overall effect of the scheduling can be modeled in the form of a directed acyclic graph, and the final scheduling scheme is determined by multiple directed acyclic graph schedules, and the idle state insertion method is used for scheduling. Only when the allocation decision and the scheduling decision are generated at the same time, the task is dispatched to different virtual machines VM.

[0035] Step 3: Build the three-stage adaptive hybrid heterogeneous resource allocation and task scheduling method of the dynamic multi-task mode, which coordinates the richness of cloud computing resources and the low transmission delay of local resources and edge resources, and optimizes the scheduling result. Please refer to Figure 1 In some embodiments of the present application, the specific three stages are described as follows:

[0036] The first stage task pre-scheduling. The earliest deadline first (EDF) method is used for task grouping scheduling, and the task with the earliest deadline is iteratively assigned to the first virtual machine that meets its requirements; the reason why EDF is superior to first task first (FF), largest task first (FFD), iterative assignment first task (BF) and earliest finish time first (EFTF) in device service (SLA) optimization is that EDF prioritizes the requirements of tasks with tighter deadlines, thereby postponing tasks with more slack time; after determining the tasks to be assigned, an initial task scheduling group is generated, and the task execution order is adjusted through scheduling after resource allocation; the scheduling scheme is adjusted according to the constraints, targets, resources and system environment (the constraints include deadline time constraints and resource constraints, and the target is to minimize the maximum completion time and maximum resource utilization according to the constraints; the scheduling scheme is adjusted according to the resource limit and the dynamic changes in the environment, including dynamic insertion, machine failure, processing order adjustment and other events; specifically, the state of the environment is modeled into a graph structure as the dynamic state input of the neural network, thereby realizing the process of dynamic scheduling); compared with other methods, more tasks with tighter deadlines can be completed.

[0037] The second stage resource pre-allocation. The cloud edge resource pre-allocation mechanism is performed, and the emergency task adopts EFTF (earliest finish time first), and the first task is iteratively assigned to the virtual machine with the earliest finish time; only the finish time is considered, and the non-emergency task is assigned to the resource with lower cost and slower speed. Existing research considers that data is transmitted to the computing node only when a decision is made on the allocation of each task, which greatly affects the execution performance of dynamic tasks. In the resource pre-allocation stage, the data is cached in the edge server in advance, especially in the cloud, which can improve the performance of task execution. For the caching mechanism, considering the priority limitation after the task allocation in the first stage, multiple idle time periods will be formed on the heterogeneous platform, and there is an idle time period between the earliest finish time of the previous sub-task in the same working task and the earliest start time of the same processing task group, which can be used. Therefore, the idle time period can be used for data caching in advance. Therefore, the idle time period allocation method is considered in the resource pre-allocation, and it is crucial to reduce the idle time period to realize sufficient resource utilization.

[0038] The idle time period allocation method is as follows: first, all tasks that can be completed by cloud resources are pre-allocated to the cloud, for each task, a best value for money virtual machine VM is pre-leased, then when there is no task that can be completed in the cloud, the remaining tasks are pre-allocated to the edge server using LSTF (iteratively assigning the first task to the provider with the least idle time) and EDF. The pre-allocation work can be improved by re-allocating some tasks from local devices or edges to the cloud, leaving some resources idle to complete the remaining unallocated tasks. Assigning tasks that can be completed by the cloud and local devices or edges at the first time to the cloud to generate more available local and edge resources for processing tasks that can only be met by local devices or edges, so that resource utilization is improved. (The main task is to fully utilize local resources, release part of the edge resources, re-schedule multiple tasks from the edge to the corresponding device, and use multiple user shared edge resources to complete more tasks.)

[0039] Third stage task scheduling. Dynamic task scheduling adjustment strategy is performed, and a task scheduling algorithm is used to implement all the task allocations described in the second stage on each device, edge and cloud, which determines which core processes the task. The task scheduling algorithm is as follows: check each task allocated to the cloud, if the task can be completed by the corresponding device, then re-allocate the task to the device, otherwise, check if the requirements of the task can be met by the edge server, if so, then re-allocate the task to the edge server. After these re-allocations, the tasks allocated to the next computing virtual machine are re-allocated to the previous computing virtual machine by using the task scheduling algorithm, and the idle virtual machines are re-leased. The idle time of the virtual machines is reduced, and the cost efficiency is improved.

[0040] As shown in Figure 1 The task pre-sorting stage (TPS) groups the initial tasks and the dynamic tasks added later according to the task urgency, and each task group is further divided into different task nodes according to the task type; then the resource pre-allocation stage (RPA) caches the tasks according to the time delay requirements, and the tasks with high time delay requirements are preferentially allocated to the industrial terminal task set for processing, the data with low time delay requirements and large scale are cached to the cloud, the cloud allocates the tasks according to the task priority, preferentially allocates to the edge task set, and the remaining tasks are allocated to the industrial terminal task set for processing; finally, the task scheduling stage (TD) uses a scheduling strategy to allocate each task set to the cloud device group, the edge device group and the industrial terminal device group for task and resource matching, and obtains the execution order.

[0041] All the task allocations are implemented on each device, edge and cloud by using a task scheduling algorithm, which determines which core processes the task.

[0042] The cloud edge end resource scheduling optimization method based on the double-layer graph neural network provided by the foregoing embodiments of the application can better realize the interaction process between the cloud edge end environment and the actual environment, complete physical mapping, and reduce the difference between the model expression and the actual problem by constructing a graph neural network scheduler, using a directed acyclic graph to represent dynamic state information. The method is suitable for adapting to dynamic environment changes, assisting in improving the cloud edge end resource scheduling efficiency, and setting a three-stage resource allocation method to accelerate the processing speed of emergency tasks and improve the resource utilization rate.

[0043] The method provided by the foregoing embodiments of the application can obtain total production tasks by obtaining order information from the cloud, distribute the tasks to the edge end or the industrial terminal according to a task intelligent distribution algorithm, and perform real-time statistics on the equipment load condition according to the task distribution. In the face of an emergency, task migration can be performed to reduce the equipment load rate. According to the foregoing task distribution and resource management, the task distribution and scheduling are performed by using a graph neural network to determine the order of the distributed tasks, and then the tasks are represented as node embedding graphs. Considering the calculation amount and task complexity, the tasks with large calculation amount and high task complexity are offloaded to the edge, and the remaining tasks with small calculation amount and low task complexity are directly offloaded to the local terminal. The production tasks are scheduled by the graph neural network scheduler when the tasks are distributed to the edge end. A three-stage distributed cloud edge end resource allocation algorithm is designed for the task offloading strategy problem. The allocation strategy is obtained under the optimal resource allocation, and then the optimal resource allocation is performed under the given allocation strategy. The two processes are iterated with each other until convergence, so that the system utility is maximized. The method can realize the rational use of resources of the cloud, the edge end and the industrial terminal in the intelligent manufacturing background, perform fast preparation of resource offloading, improve the comprehensive scheduling capability of industrial production, improve the resource utilization rate, and realize green scheduling.

[0044] The step numbers in the foregoing embodiments are only set for facilitating the description, and the order between the steps is not limited in any way.

[0045] The foregoing description of the disclosed embodiments enables a person skilled in the art to implement or use the application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cloud edge resource scheduling optimization method based on a double-layer graph neural network, characterized in that, The method comprises the following steps: A cloud-edge-end resource scheduler based on a double-layer graph neural network is constructed, a first layer of graph neural network is used to realize cloud-edge-end resource allocation, a cloud-edge layer input graph represents the relationship between cloud and edge servers, a graph is established by connecting to each vertex related to one of the edges through two edges, a graph is established between cloud-edge network servers and industrial terminals, a heterogeneous graph is constructed to capture the relationship between the cloud, edge and industrial terminals as the input of the first layer of graph neural network to realize cloud-edge-end resource allocation and realize the allocation of computing tasks; a second layer of graph neural network is used for industrial terminal layer scheduling, a heterogeneous graph is used to represent the state in the actual manufacturing environment to capture the complex relationship between workpieces, processes and available devices, a directed acyclic graph is constructed as the input of the second layer of graph neural network to perform task allocation and scheduling to determine the order of allocating tasks, and then the tasks are represented as node embedding graphs; According to the cloud-edge-end resource scheduler based on the double-layer graph neural network, the computing resource allocation problem of the cloud, edge and industrial terminal is considered, a framework model for scheduling and offloading in stages is constructed, part of the tasks are offloaded to the edge, and the remaining tasks are directly offloaded to the local terminal machine through the graph neural network scheduler for unified scheduling. The framework model for scheduling and offloading in stages adopts an optimization strategy to optimize the scheduling of tasks, the optimization strategy is a three-stage adaptive hybrid heterogeneous resource allocation and task scheduling method in a dynamic multi-task mode, the first stage of the three-stage adaptive hybrid heterogeneous resource allocation and task scheduling method in the dynamic multi-task mode is task pre-sorting, the second stage is resource pre-allocation, and the third stage is dynamic task scheduling, so as to coordinate the richness of cloud computing resources and the low transmission delay of local resources and edge resources, and optimize the scheduling result.

2. The cloud edge resource scheduling optimization method based on the double-layer graph neural network according to claim 1, characterized in that, In the cloud-edge-end resource scheduler based on the double-layer graph neural network, a state feature embedding module is constructed to obtain task state information from the input order information of the cloud, and a directed acyclic graph DAG is used to describe the dynamic state in response to the occurrence of dynamic events.

3. The cloud edge resource scheduling optimization method based on the double-layer graph neural network according to claim 2, characterized in that, A plurality of directed acyclic graphs DAGs are used to represent a task set, the EFT is used to derive the minimum scheduling time, the urgency and priority of each directed acyclic graph DAG are calculated, a two-dimensional sorting tuple is obtained, the urgent directed acyclic graph DAG is preferentially processed during scheduling, the dynamic change is reflected on the directed acyclic graph DAG, and the newly arrived task set does not interrupt the urgent task; all tasks in the directed acyclic graph DAG are scheduled one by one, and the sorting tuple is updated, the scheduled directed acyclic graph DAG is deleted from the scheduling set in time; the scheduling state described by the directed acyclic graph DAG is converted into a heterogeneous graph state, and then a double-layer heterogeneous graph neural network is used to represent the learning of the heterogeneous state graph, so as to obtain the node state feature embedding of the order information.

4. The cloud edge resource scheduling optimization method based on the double-layer graph neural network according to claim 1, characterized in that, In the second layer graph neural network, the solution of the job shop scheduling problem is modeled as a heterogeneous graph model, and the state feature graph structure G t =(V,M,C,E,O V ,C E ) is defined, V is the node of multiple operation types, M is the machine node, each node corresponds to a machine Mk, C is the set of conjunctive arcs, which represents each processing procedure of the job, the disjunctive arc set E is a undirected arc, which is the link connecting the operation node and the compatible machine node, O V represents the set of operation object types, C E represents the set of relationship types, each node is associated with heterogeneous content; increase the procedure correlation constraint, enrich the conjunctive arc set C, and get the tuple form of the arc weight, so as to represent the state information of the graph node.

5. The cloud-edge-end resource scheduling optimization method based on the double-layer graph neural network according to claim 1, characterized in that, The framework model for scheduling and offloading in stages comprises receiving order information as a computing task at the cloud-edge-end, the edge comprises a plurality of edge servers, each edge server has a group of virtual machines, the cloud server also comprises a group of virtual machines, and the virtual machines on the cloud server and the edge server are rented on demand.

6. The cloud edge resource scheduling optimization method based on the double-layer graph neural network according to any one of claims 1-5, characterized in that, The optimization strategy first allocates the computing tasks to the heterogeneous platform, and adjusts the execution sequence by scheduling as coupling, if the resource allocation generates a good object library for the scheduling, the uniformly distributed tasks will not accumulate in the platform, the overall effect of the scheduling is modeled in the form of a directed acyclic graph, and finally the scheduling scheme is determined by multiple directed acyclic graph scheduling, the scheduling is performed in the form of idle state insertion, and when the allocation decision and the scheduling decision are generated at the same time, the tasks are dispatched to different virtual machines.

7. The cloud-edge-end resource scheduling optimization method based on the double-layer graph neural network according to claim 6, characterized in that, In the first stage of task preordering, the earliest deadline first method is used for task grouping and ordering, and the task with the earliest deadline is iteratively allocated to the first virtual machine that meets its requirements; after determining the task to be allocated, an initial task scheduling group is generated, and the task execution order is adjusted by the scheduling after resource allocation, and the scheduling scheme is adjusted according to the constraints, targets, resources and system environment.

8. The cloud-edge-end resource scheduling optimization method based on the double-layer graph neural network according to claim 7, characterized in that, In the adjustment of the scheduling scheme according to the constraints, targets, resources and system environment, the constraints include deadline time constraints and resource constraints; the target is to minimize the maximum completion time and the maximum resource utilization according to the constraints; the scheduling scheme is adjusted according to the resource limit and the dynamic change in the environment, specifically, the state of the environment is modeled into a graph structure as the dynamic state input of the neural network, thereby realizing the process of dynamic scheduling.

9. The cloud edge resource scheduling optimization method based on the double-layer graph neural network according to claim 6, characterized in that, In the second stage of resource pre-allocation, the first task is iteratively allocated to the virtual machine with the earliest completion time for urgent tasks; only the completion time is considered, and non-urgent tasks are allocated to resources with lower cost and slower speed; the idle time period allocation method is used to cache data to the edge server or cloud in advance, the idle time period allocation method is, first, all tasks that can be completed by cloud resources are pre-allocated to the cloud, for each task, a best value for money virtual machine VM is pre-rented, then no task can be completed in the cloud, the remaining tasks are pre-allocated to the edge server using the first task allocation method; the pre-allocation work is improved by re-allocating some tasks from the local device or edge to the cloud, so that some resources are idle to complete the remaining unallocated tasks.

10. The cloud edge resource scheduling optimization method based on the double-layer graph neural network according to claim 6, characterized in that, In the third stage of task scheduling, a task scheduling algorithm is used to implement all task allocations in the second stage on each device, edge and cloud, which determines which core processes the task, the task scheduling algorithm checks each task allocated to the cloud, if the task can be completed by the corresponding device, the task is re-allocated to the device, otherwise, it is checked whether the requirements of the task can be met by the edge server, if so, the task is re-allocated to the edge server, after these re-allocations, the tasks allocated to the next computing virtual machine are re-allocated to the previous computing virtual machine, and the idle virtual machine is re-rented.

Images