A delay-aware resource reservation method based on spatiotemporal task demand prediction

Abstract

The present invention provides a delay-aware resource reservation method based on spatiotemporal task demand prediction. The present invention constructs a spatiotemporal deep learning model to predict regional task offloading demand by fusing a graph convolutional neural network and a gated recurrent neural network based on an attention mechanism, proposes a delay-aware edge server resource reservation strategy for spatial regions, and minimizes terminal task offloading delay by regional edge server resource reservation. Based on the spatial and time-varying characteristics of resource demand, the present invention effectively reduces the delay of terminal task offloading by predicting regional task offloading demand, proposing a delay-aware edge server resource reservation strategy for spatial regions, and designing a task offloading delay minimization model based on resource reservation, thereby improving the resource scheduling and operation efficiency of the edge network.

Description

Technical Field

[0001] The present invention relates to the field of edge computing and end-edge-cloud collaborative technology, and in particular to a delay-aware resource reservation method based on spatiotemporal task demand prediction. Background Art

[0002] With the rapid development of autonomous driving, smart transportation, smart cities and other fields, the demand for resources of various types of terminals has increased significantly. Due to the limitations of their own energy supply and computing power, terminal devices cannot effectively perform computing-intensive tasks and perform tasks stably for a long time. Computation offloading solves the limitations of terminal devices in the above factors. By offloading tasks to cloud servers or edge servers to assist terminal devices in completing computing-intensive tasks and diversified types of tasks, the energy consumption of terminal devices can be effectively saved and the long-term stable operation of terminal devices can be ensured. Compared with cloud computing, edge computing has the advantages of low task completion latency and low energy consumption, so offloading tasks to edge servers is a preferred option. Existing research on edge server resource reservation is based on historical data statistics, but resource requirements often have spatial and time-varying characteristics. Affected by this, how to flexibly reserve edge server resources is a direction worth studying, which will improve the resource scheduling and operation efficiency of city-level edge networks. Summary of the invention

[0003] The present invention aims at the problem of how to flexibly reserve edge server resources for spatial regions because the resource requirements of various types of existing terminals often have spatial and time-varying characteristics. A delay-aware resource reservation method based on spatiotemporal task demand prediction is proposed. First, a deep learning model is constructed to predict the regional spatiotemporal associated task offloading requirements. Then, based on the task offloading requirements, a delay-aware edge server resource reservation strategy for spatial regions is proposed. A terminal task offloading delay minimization model based on resource reservation is designed. The terminal task offloading delay is minimized by reserving regional edge server resources, thereby obtaining the minimum completion time of the terminal offloading task in the region.

[0004] The method of the present invention comprises the following steps:

[0005] Step 1: Divide the entire area into several sub-areas, and then predict the task offloading demand of each sub-area. Divide the area into several disjoint sub-areas, and use ν, ν∈S to represent the sub-areas, where S represents the set of sub-areas. Integrate the edge servers in each sub-area into one edge server to simplify the problem.

[0006] First, through the graph structure Indicates the spatial correlation between sub-regions. In the graph, each sub-region is represented as a vertex. In this case, the set S of sub-regions is the set of vertices in the graph. It is the adjacency matrix of the graph, which indicates the connectivity of the vertices in the graph. Four spatial correlation models between sub-regions are considered, including regional neighbor graph, regional connectivity graph, regional function correlation graph, and regional task type correlation graph. The nonlinear correlation is represented by the graph structure, and the feature extraction and further fusion of multiple graphs are realized through the multi-graph convolutional network.

[0007] At the same time, considering the correlation in time series, using O |S| (τ) represents the order of all sub-region observations at time slot τ. Assuming that the observations at time slot τ are affected by the previous ω time slots, the regional spatiotemporal correlation task offloading demand forecasting problem is expressed as a single-step spatiotemporal forecasting problem of a given time length. By learning the function Ψ: The historical data of terminal task unloading demand in the previous time slot ω is learned to obtain the regional terminal task unloading demand in time slot τ. k+1 The observed value of is affected by the previous ω time slots, O |S| (τ k+1 )=Ψ{O |S| (τ k-ω+1 ),…,O |S| (τ k )}, |S| represents the number of regions, and k represents the number of time slots.

[0008] The global features of the region, that is, the spatial correlation features between regions, are extracted through the graph convolutional network. Aggregate the features of the current region and the associated region. Θ represents the aggregation function, which aggregates the feature information of the current region and the feature information of the spatial correlation region to form a regional feature description. represents multi-graph convolution operation, Represents the maximum degree of the Laplacian matrix of the graph.

[0009] Then, the global average pooling operation and Dimensionality reduction is used to obtain the regional description of the current sub-region and its related sub-regions, where represents the global average pooling operation, In order to aggregate the features of the current sub-region and the associated sub-regions through a multi-graph convolutional network, Represents the characteristics of different current regions and their associated regions.

[0010] Then pass The operation adaptively assigns corresponding weight coefficients to different observation sequences. s represents the weight vector of the time observation sequence, Indicates the operation of assigning weights to observation sequences at different times, W 1 and W 2is the weight parameter, σ and φ represent the activation functions ReLU and sigmoid, respectively.

[0011] pass Assign a weight coefficient to the original observation sequence to obtain a new observation sequence, k represents the number of time slots, O |S| (τ k ) represents the order of all sub-region observations at the kth time slot, s(τ k ) represents the weight vector of the time observation sequence of the kth time slot, Represents the operation of re-weighting the original observation sequence to obtain a new observation sequence.

[0012] The observations at different time slots are aggregated through a gated recurrent neural network, and the gated input sequence of each sub-region is aggregated into a single vector Ω through a shared RNN layer. i , m represents the maximum number of time slots k, i represents the number of sub-regions, the maximum is |S|, W 3 represents the weight parameter, It represents the new observation sequence obtained after assigning weight coefficients to different time slots. Finally, the extracted features are converted into prediction results through the fully connected layer, that is, the regional spatiotemporal correlation task offloading requirements.

[0013] Therefore, a SpatioTemporal Correlation Task Offloading Demand (STCTOD) prediction model based on regional spatial correlation and temporal correlation is constructed.

[0014] Step 2: According to the STCTOD model constructed in step 1, the task offloading demand of each sub-region is predicted. The cloud server sends resource reservation requests to the edge servers of the sub-regions according to the predicted values ​​to reserve the resources owned by the sub-regions. If the resource reservation request of the sub-region is not met, resources are scheduled between the sub-regions. The regions whose resource reservation requests are still not met will be received by the cloud.

[0015] Since long-distance transmission to cloud servers will lead to high latency and excessive consumption of resources such as bandwidth, cloud servers should be the last choice for resource reservation requests. The time cost of resource scheduling between sub-regions where resource reservation requests are not met and sub-regions with abundant resources is minimized, and the time cost includes task migration time and task calculation time.

[0016] The details are as follows:

[0017] Build an edge network regional topology map based on device-edge-cloud collaboration and define a time consumption model, mainly including the cost of migrating tasks between regions And the time cost of computing on the edge server The problem of minimizing the terminal task offloading delay based on regional resource reservation is defined; in order to minimize the terminal task offloading delay, the time cost of edge server resource reservation should be minimized.

[0018] In order to ensure the normal operation of the edge server, the actual workload of the edge server should be less than its maximum workload, and the threshold coefficient of resource usage in the edge server is set to ∈; a task offloading delay minimization model based on resource reservation is constructed:

[0019]

[0020]

[0021]

[0022]

[0023]

[0024] Indicates that in time slot τ k The total time cost of task offloading is m, m represents the maximum number of time slots k, and formula (2) indicates whether the current task needs to be offloaded from the current edge server v i Migration, V s It includes cloud servers and edge servers. f represents the task computing amount requested by the current task. F i Indicates the task computation amount of the edge server or cloud server and defines the decision variable It indicates that task i is migrated to other edge servers or cloud servers in time slot τ. Indicates migration to other edge servers. represents migration to the cloud server; Formula (3) shows that when the task needs to be migrated from the current edge server, the task is migrated to other edge servers or cloud servers, and the decision variable is defined Indicates the edge server v i Whether the tasks on the task need to be migrated. Indicates that the task is executed on the current edge server. It means that the task needs to be migrated from the current edge server to other edge servers or cloud servers for calculation; Formula (4) shows the computing power h requested by the current task i The computing power threshold H of the edge server or cloud server should not be exceeded i ; Formula (5) shows the final resource reservation value a of each area i (τ) should not exceed the threshold R of available resources in the current area i(τ)×∈.

[0025] Input the task offloading requirements of each regional terminal and output the resource reservation strategy that each sub-region edge server can accept. and the resource reservation policy received by the cloud

[0026] Compared with the prior art, the advantages and positive effects of the present invention are:

[0027] 1. From the analysis of spatial factors, there are complex spatial correlations between different regions. For example, the traffic flow and terminal task unloading demand in a certain area are often affected by the traffic flow in adjacent areas and the task unloading demand in associated areas. From the analysis of time factors, there is also a correlation between the observations of terminal task unloading demand in different time periods on a continuous time series. For example, the prediction of terminal task unloading demand is usually related to the historical data of its unloading task. It can be seen that there is a complex regional spatiotemporal correlation in the prediction of regional spatiotemporal associated task unloading demand. The present invention adopts a graph structure to construct four spatial correlation models between regions, a multi-graph convolutional network, and a global gated recurrent neural network based on an attention mechanism to accurately predict the terminal task unloading demand, effectively reduce the delay of terminal task unloading, and help city-level edge networks to efficiently schedule computing resources and improve urban operation efficiency.

[0028] 2. In order to solve the problem of over-scheduling of resources caused by real-time edge server resource allocation during task offloading, regional resource reservation is used to effectively reduce resource scheduling, improve resource scheduling efficiency, and save terminal task offloading costs. The present invention effectively responds to the outbreak of local time periods and regional task offloading demands through a reasonable regional resource reservation model, reduces resource waste caused by scheduling, and supports long-term and efficient operation of the system.

[0029] 3. Construct a task offloading delay minimization model based on resource reservation, which takes into account the migration time of tasks between edge server nodes and the calculation time of terminal offloading tasks. By reserving the resources of regional edge servers, the terminal offloading tasks can be executed in the local area as much as possible, reducing the waste of resources caused by the migration of tasks between edge servers, and ultimately minimizing the completion time of terminal offloading tasks in the region. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is the overall architecture diagram of the present invention.

[0031] Figure 2 The regional division diagram in the embodiment

[0032] Figure 3 Schematic diagram of the spatial correlation model.

[0033] Figure 4Schematic diagram of the time correlation model.

[0034] Figure 5 This is the topology diagram of the edge network area. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0036] Taking the Internet of Vehicles scenario as an example, due to the high mobility of vehicle terminal devices and the burst traffic in local time periods and regions, the resource scheduling will become longer, more costly, and less efficient. If resources can be reserved in advance for regional resource needs, it can effectively reduce resource scheduling, improve resource scheduling efficiency, and save task offloading costs. Therefore, reserving regional resources through effective regional task offloading demand prediction is of great practical application value.

[0037] like Figure 1 As shown, a delay-aware resource reservation method based on spatiotemporal task demand prediction is proposed, which decomposes the problem of delay-aware resource reservation based on spatiotemporal task demand prediction into task offloading demand prediction and delay-aware resource reservation.

[0038] First, for the prediction of spatiotemporal-related task offloading demand, a spatiotemporal deep learning model STCTOD was constructed to predict regional task offloading demand. The graph convolutional neural network and the gated recurrent neural network based on the attention mechanism were integrated to extract the spatial correlation and temporal correlation of highly complex task offloading demands between regions. In the second step, a latency-aware regional edge server resource reservation strategy was proposed, and an edge network regional topology map based on end-edge-cloud collaboration was constructed and a regional edge server resource reservation model was constructed based on this. Then, a time consumption model was defined, which mainly includes the cost of task migration between regions and the time cost of computing on the edge server. Thus, a task offloading delay minimization model ESRR based on resource reservation was constructed to minimize the terminal task offloading delay through regional edge server resource reservation.

[0039] The specific implementation process of this embodiment is as follows:

[0040] Step 1: A task offloading demand prediction model STCTOD based on regional spatial correlation and temporal correlation is constructed through a spatiotemporal multi-graph convolutional network and a gated recurrent neural network based on an attention mechanism to obtain regional task offloading demand.

[0041] like Figure 2 As shown in Figure 2, in order to better predict the terminal task offloading demand, the entire area is divided into 9 sub-areas of 3×3, namely, v i,i=1,2,…,|S|, where |S|=9. Different types of terminals, such as cars, are distributed in the sub-areas. Each terminal has a series of tasks that need to be unloaded. The tasks are unloaded to the edge server through wireless transmission via the roadside unit. The edge servers in each sub-area are integrated into one edge server to simplify the problem. Then, the task unloading demand of each sub-area is predicted, and a regional task unloading demand prediction deep learning model STCTOD is constructed. The spatiotemporal correlation of tasks in each sub-area is considered. The spatial correlation of task unloading demand is considered from four dimensions: regional proximity, regional connectivity, regional functional correlation, and regional task type correlation. The nonlinear correlation is represented by the graph structure, and the feature extraction and further fusion of multiple graphs are achieved through a multi-graph convolutional network. Figure 3 Shown are four graph structures of the spatial correlation model.

[0042] Using graph structure represents the spatial correlation between sub-regions, where S is the set of vertices in the graph, represents the set of regions, It is the adjacency matrix of the graph, which represents the connectivity of the vertices in the graph. is the degree matrix of the graph, which represents the degree of the nodes in the set S. i is the number of sub-regions, with a maximum of 9. The Laplacian matrix of the graph is obtained through the adjacency matrix A and the degree matrix D of the graph. The symmetric normalized Laplacian matrix L = ID -1 / 2 AD -1 / 2 , where I is the identity matrix.

[0043] Regional Neighborhood Graph Represents the correlation of spatial proximity between sub-regions, through the formula Indicates that the central region of the 3×3 grid has a maximum of eight neighbor regions.

[0044] The connectivity between two sub-areas means that the two sub-areas can be directly connected through roads (such as highways). Indicates that, since the regional proximity relationship in the regional neighbor graph and the connectivity relationship in the regional connectivity graph will overlap to a certain extent, in order to reduce the complexity, the regional connectivity graph is simplified and the correlation relationship that has been represented in the regional neighbor graph is excluded to make it a sparse graph. The above formula can be expressed as

[0045] Different areas have different positioning in planning, resulting in different functions. Areas with the same functional positioning also have temporal and spatial correlation. Therefore, when predicting the terminal task offloading demand, it is necessary to consider the correlation relationship between areas with similar functions. Indicates that use Represents region v iand v j The regional functional correlation between and Respectively represent the region v i and v j The more overlapping the functional features between two regions are, the higher the correlation between them is.

[0046] The task types in different regions can also indicate the correlation between regions. This correlation can be determined by judging the degree of overlap of task labels in different regions. Indicates. Represents region v i and v j The regional task type correlation between and Respectively represent the region v i and v j The more overlaps in task type labels between two regions, the higher the correlation between them.

[0047] The feature extraction and further fusion of multiple images are realized through multi-image convolutional networks to obtain the spatial correlation relationship between regions. I+1 =σ(∑ A∈Z f(A;θ i )X l W l ) to perform graph convolution operations. and Respectively represent the regional feature vectors of the l+1th layer and the lth layer in the graph convolutional network, P l and P I+1 represents the lth and l+1th layers of the graph convolutional network. σ represents the activation function ReLU. f(A; θ i ) represents the K-order polynomial function of the Laplacian matrix L of the graph. Represents the feature conversion matrix. The model diagram of multi-graph convolution is as follows Figure 3 shown.

[0048] The correlation in time series is modeled through a recurrent neural network, that is, a global gated recurrent neural network based on the attention mechanism. First, the features of the current region and the associated region are aggregated through a multi-graph convolutional network, that is, through the formula Get, where Θ represents the aggregation function, which aggregates the feature information of the current region and the feature information of the spatial correlation region. Represents the global features of the region. The feature extraction of spatial correlation is achieved through multi-graph convolution operations, where represents multi-graph convolution operation, represents the maximum degree of the Laplacian matrix of the graph. Let represents the regional observation value at time slot τ, where P represents the dimension of the regional feature vector.

[0049] Next, the regional observation information is further processed based on the attention mechanism, and weight coefficients are assigned to observation values ​​in different time slots. First, the observation information of each region in the time slot τ is compressed through the global average pooling operation to extract the global features of the region. Perform average global pooling. represents the global average pooling operation, |S| represents the number of regions, Indicates the characteristics of different current regions and their associated regions. The second step is to perform operations based on the attention mechanism, that is, to assign weights to different observation sequences, adaptively give higher weight coefficients to key information so that it has a greater impact on the results, and discard outdated information to a certain extent. The operation adaptively assigns corresponding weight coefficients to different observation sequences. s represents the weight vector of the time observation sequence, Indicates the operation of assigning weights to observation sequences at different times, W 1 and W 2 is the weight parameter, σ and φ represent the activation functions ReLU and sigmoid, respectively.

[0050] pass Assign a weight coefficient to the original observation sequence to obtain a new observation sequence, k represents the number of time slots, O |S| (τ k ) represents the order of all sub-region observations at the kth time slot, s(τ k ) represents the weight vector of the time observation sequence of the kth time slot, Represents the operation of re-weighting the original observation sequence to obtain a new observation sequence.

[0051] The observations at different time slots are aggregated through a gated recurrent neural network, and the gated input sequence of each sub-region is aggregated into a single vector Ω through a shared RNN layer. i , m represents the maximum number of time slots k, i represents the number of sub-regions, the maximum is |S|, W 3 represents the weight parameter, represents the new observation sequence obtained after assigning weight coefficients to different time slots, m represents the maximum number of time slots k, i represents the number of sub-regions, and the maximum is |S|. Finally, the extracted features are converted into prediction results through the fully connected layer, that is, the regional spatiotemporal correlation task offloading requirements.

[0052] The modeling of time correlation is achieved through the above steps, which can better capture the impact of historical data in different time slots on prediction. The model of time correlation is as follows Figure 4 As shown. First, input ω observation sequences {O |S| (τ k-ω+1 ),…,O |S| (τ k )}, after global average pooling operation and Dimensionality reduction is used to obtain the regional description of the current region and its related regions, and further The operation adaptively assigns corresponding weight coefficients to different observation sequences. Finally, The operation assigns a weight coefficient to the original observation sequence to obtain a new observation sequence. The weight coefficient of the observation sequence represents its importance, that is, the higher the weight, the greater the impact on the terminal task offloading demand prediction result.

[0053] Step 2: Based on the predicted values, resources are scheduled between regions to meet the resource reservation requests of the regions, and a latency-aware edge server resource reservation model for spatial regions is constructed.

[0054] First, we build an edge network regional topology based on end-edge-cloud collaboration, which is divided into four layers: terminal device layer, roadside unit layer, edge layer, and cloud layer. The terminal device layer is composed of terminals, and the terminals have a series of tasks to be offloaded. The roadside unit layer is used to connect various terminals and provide relay communication services for terminals to offload tasks. The edge layer is composed of edge servers widely distributed in the city, which are used to carry the computing tasks offloaded by various terminals in the terminal device layer. The edge nodes can be connected to each other within the communication range. The cloud layer is composed of powerful server clusters, which are used to carry tasks that edge servers cannot perform and tasks with high computing volume, thereby alleviating the computing pressure on edge servers. The edge network topology diagram is shown below: Figure 5 shown.

[0055] Construct a network graph G(V,E) to represent the edge network area topology graph, where V represents the node set, that is, V = {v i ,i=1,…,M}, M represents the maximum number of nodes. E represents the set of edges, i.e., the network communication links. and |E| is the maximum number of edges N. The node set V consists of the cloud, edge servers, roadside units, and terminal devices. The present invention uses v without confusion. 0 Specifically refers to cloud servers, v s Specifically refers to the edge server, v r Specifically refers to the roadside unit, v m Specifically refers to terminal devices. The node set V is divided into three subsets, using V s 、V r and Vm Indicates that V s Including cloud servers and edge servers, V r Including roadside units, V m Includes various types of terminal equipment, namely e jk Represents the slave node v j to v k Considering the time-varying factors in the task offloading process, the present invention discretizes time into n time slots with equal intervals and uses τ k , k=1,2,…n represents. Time slot τ k The starting time is t k , k=0,1,2,…,n represents. At each time t k Determine the time slot τ k The time slot is between two moments, such as the time slot τ k At time t k-1 and t k between.

[0056] Regional edge server resource reservation refers to obtaining the regional terminal task offloading demand through the STCTOD model, thereby reserving sufficient edge server computing resources for the region. Let ρ i (τ) represents the terminal device v i The amount of task offloading at time slot τ. Let R i (τ) represents the terminal task offloading amount of region i. In the absence of confusion, R i (τ) also represents the resource reservation request value of region i, R i (τ) is affected by the amount of terminal offloading tasks in area i, through v j In v i Within the range, this formula is calculated. The total amount of resources of the edge server located in region i in time slot τ is Q i (τ). When the edge server resources in region i are insufficient to satisfy the resource reservation request of the current region, that is, R i (τ)>Q i (τ), a resource reservation request migration will be sent to region j. If there is a region j such that the resource reservation request of region i satisfies R i (τ)≤Q j (τ), then it is called a feasible regional resource reservation request migration from region i to region j, using p ij (τ) indicates that the resource reservation request value is |p ij (τ)|. When a given region i, for any region j, satisfies the condition R i (τ)>Q j(τ), the resource reservation request is divided into two parts and processed separately. The first part searches for the edge server reserved resources in area j, and the resource value is |p ij (τ)|. The other part of the resource reservation request is sent to the cloud, and the resource reservation request amount is defined as b i , b i =R i (τ)-|p ij (τ)|. Define variable a i Indicates the final resource reservation value of regional edge server i. Define two collection variables and They respectively include the edge server resources that need to be reserved in each region and the cloud server resource reservation requested by each region, that is,

[0057] The time cost of terminal task unloading mainly includes four parts: the time cost of terminal transmission task to roadside unit The time cost of unloading tasks from roadside units to edge servers The time cost of migrating tasks between edge servers or from edge servers to the cloud The time cost of completing tasks on edge servers or in the cloud

[0058] make v m Node v i Transfer the task to v at time slot τ r Node v j The time cost of transferring the task unloaded from the terminal to the roadside unit in the time slot τ is By formula Calculate, ρ i (τ) represents the terminal device v i The amount of task offloading at time slot τ. Let v r Node v i Offload the task to v at time slot τ s Node v j The time cost of unloading tasks from roadside units to edge servers is The time cost is given by the formula Calculate, d(v i ,v j ) indicates v r Node v i and v s Node v j The Euclidean distance of calculate.

[0059] Finally, a task offloading delay minimization model ESRR based on resource reservation is designed. Since offloading tasks to cloud servers will lead to high latency and underutilization of edge servers, a method based on edge server resource reservation is adopted to solve this problem, that is, the regional edge server tries its best to receive resource reservation requests. The time consumed in task migration between edge servers is mainly determined by the migration distance, that is, finding the shortest path from region i to j. The Dijkstra algorithm can effectively solve the single-source shortest path problem. In addition, the time it takes for the edge server to execute the task is also taken into consideration. Formula Indicates the edge server v i Whether the tasks on the edge server need to be migrated, where the time cost affected by the edge server resource reservation strategy is the task migration time and task calculation time Therefore, when considering edge server resource reservation, it simplifies to the formula

[0060] By formula Calculate and define decision variables It indicates that task i is migrated to other edge servers or cloud servers in time slot τ. Indicates migration to other edge servers. Indicates migration to the cloud server. Indicates that in time slot τ, the task is sent from the edge server node v i Migrate to other edge server nodes v j The rate of Indicates that in time slot τ, the task is sent from the edge server node v i Migrate to v 0 The rate of the node. d(v i ,v j ) is v r Node v i and v s Node v j The Euclidean distance. The calculation formula is Among them, F i (τ) represents the edge server v i or cloud server v 0 The task computation amount, h i (τ) represents the edge server v i or cloud server v 0 computing power.

[0061] Finally, a regional edge server with the smallest sum of migration time and execution time is selected to receive the resource reservation request. The input of the ESRR model includes the historical observation data of regional terminal task offloading O |S| (τ 1 ),O |S| (τ 2 ),…,O |S| (τ k-1 ), the average offload task volume of the terminal ρ, the edge server in time slot τ k Available resources Q(τ k ), the edge server workload threshold coefficient ∈ and the time slot τ k The transmission rate α(τ k ), β(τ k ),γ(τ k ),η(τ k ). The output of the model includes the resource reservation policy for the regional edge servers Cloud resource reservation strategy The specific execution steps of the model are as follows. First, input the historical observation data of terminal task unloading, and predict it through the model STCTOD to obtain the time slot τ k Regional terminal task offloading demand O |S| (τ k ). Next, and the set of regions that do not satisfy resource reservation requests Set to an empty set. For each region i,v i ∈S, updates the edge server in time slot τ according to the edge server workload threshold coefficient ∈ k The actual available resources Q i (τ k ). According to the predicted terminal task offloading demand O |S| (τ k ) and the terminal offload task volume ρ calculates the resource reservation request R of area i i (τ k ). If the edge server resources in the current region meet the resource reservation request, the request is accepted. Otherwise, the current region is marked and put into the collection For each area The Dijkstra algorithm is used to find the region j that is closest to the requested region. If j≠-1, it indicates that a region that meets the conditions exists and the region will receive the resource reservation request. If j=-1, it indicates that no region meets the resource reservation request and the resource reservation request is uploaded to the cloud. The final output is the resource reservation policy that each region can accept. and the resource reservation policy received by the cloud

Claims

1. A delay-aware resource reservation method based on spatiotemporal task demand prediction, characterized in that: The specific steps include: Step 1: Divide the entire area into several sub-areas, and then predict the task offloading demand of each sub-area; divide the area into several disjoint sub-areas, and use ν, ν∈S to represent the sub-areas, where S represents the set of sub-areas; integrate the edge servers in each sub-area into one edge server to simplify the problem; First, through the graph structure Indicates the spatial correlation between sub-regions. In the graph, each sub-region is represented as a vertex. In this case, the set S of sub-regions is the set of vertices in the graph. It is the adjacency matrix of the graph, which represents the connectivity of the vertices in the graph. It considers four spatial correlation models between sub-regions, including regional neighbor graph, regional connectivity graph, regional function correlation graph, and regional task type correlation graph. It uses the graph structure to represent its nonlinear correlation, and the feature extraction and further fusion of multiple graphs are realized through the multi-graph convolutional network. At the same time, considering the correlation in time series, using O |S| (τ) represents the order of all sub-region observations at time slot τ; assuming that the observations at time slot τ are affected by the previous ω time slots, the regional spatiotemporal correlation task offloading demand prediction problem is expressed as a single-step spatiotemporal prediction problem of a given time length; by learning function Ψ: The historical data of terminal task unloading demand in the previous time slot ω is learned to obtain the regional terminal task unloading demand in time slot τ. k+1 The observed value of is affected by the previous ω time slots, O |S| (τ k+1 )=ψ{O |S| (τ k-ω+1 ),…,O |S| (τ k )}, |S| represents the number of regions, and k represents the number of time slots; The global features of the region, that is, the spatial correlation features between regions, are extracted through the graph convolutional network. Aggregate the features of the current region and associated regions; Where Θ represents an aggregation function, which aggregates the feature information of the current region and the feature information of the spatial correlation region to form a regional feature description; represents multi-graph convolution operation, Represents the maximum value of the Laplacian matrix degree of the graph; Then, the global average pooling operation and Dimensionality reduction is used to obtain the regional description of the current sub-region and its related sub-regions, where represents the global average pooling operation, In order to aggregate the features of the current sub-region and the associated sub-regions through a multi-graph convolutional network, Represents the characteristics of different current regions and their associated regions; Then pass The operation adaptively assigns corresponding weight coefficients to different observation sequences. s represents the weight vector of the time observation sequence, It means to assign weights to observation sequences at different times, W1 and W2 are weight parameters, σ and Represent the activation functions ReLU and sigmoid respectively; pass Assign a weight coefficient to the original observation sequence to obtain a new observation sequence, k represents the number of time slots, O |S| (τ k ) represents the order of all sub-region observations at the kth time slot, s(τ k ) represents the weight vector of the time observation sequence of the kth time slot, It represents the operation of re-weighting the original observation sequence to obtain a new observation sequence; The observations at different time slots are aggregated through a gated recurrent neural network, and the gated input sequence of each sub-region is aggregated into a single vector Ω through a shared RNN layer. i , m represents the maximum number of time slots k, i represents the number of sub-regions, the maximum is |S|, W3 represents the weight parameter, Represents the new observation sequence obtained after assigning weight coefficients to different time slots; finally, the extracted features are converted into prediction results through the fully connected layer, that is, the regional spatiotemporal correlation task offloading requirements; Thus, a task offloading demand prediction model STCTOD based on regional spatial correlation and temporal correlation is constructed; Step 2: According to the STCTOD model constructed in step 1, the task offloading demand of each sub-region is predicted. The cloud server sends resource reservation requests to the edge servers of the sub-regions according to the predicted values ​​to reserve the resources owned by the sub-regions. If the resource reservation request of the sub-region is not met, resource scheduling is performed between the sub-regions. The region whose resource reservation request is still not met will be received by the cloud. The details are as follows: Discretize time into n equally spaced time slots, using τ k ,k=1,2,…n represents; time slot τ k The starting time is t k ,k=0,1,2,…,n represents; at each moment t k Determine the time slot τ k The variable value of the time slot is between two moments, and the time slot τ k At time t k-1 and t k between; Build an edge network regional topology map based on device-edge-cloud collaboration and define a time consumption model, including the cost of migrating tasks between regions. And the time cost of computing on the edge server Define the problem of minimizing the delay of terminal task offloading based on regional resource reservation; in order to minimize the delay of terminal task offloading, the time cost of edge server resource reservation should be minimized; In order to ensure the normal operation of the edge server, the actual workload of the edge server should be less than its maximum workload, and the threshold coefficient of resource usage in the edge server is set to ; Construct a task offloading delay minimization model based on resource reservation: Indicates that in time slot τ k The total time cost of task offloading is m, which represents the maximum number of time slots k. Formula (2) indicates whether the current task needs to be offloaded from the current edge server v i Migration, V s It includes cloud servers and edge servers, f represents the task computing amount requested by the current task, F i Indicates the task computation amount of the edge server or cloud server and defines the decision variable It indicates that task i is migrated to other edge servers or cloud servers at time slot τ; Indicates migration to other edge servers. Indicates migration to cloud server; Formula (3) shows that when a task needs to be migrated from the current edge server, the task is migrated to other edge servers or cloud servers, and the decision variable is defined as Indicates the edge server v i Whether the tasks on the system need to be migrated; Indicates that the task is executed on the current edge server. Indicates that the task needs to be migrated from the current edge server to other edge servers or cloud servers for computing; Formula (4) shows the computing power h requested by the current task i The computing power threshold H of the edge server or cloud server should not be exceeded i ; Formula (5) shows the final resource reservation value a of each area i (τ) should not exceed the threshold R of available resources in the current area i (τ)× ; make v m Node v i Transfer the task to v at time slot τ r Node v j The rate at which the task unloaded by the terminal is transmitted to the roadside unit at the time slot τ. By formula Calculate, ρ i (τ) represents the terminal device v i The amount of task offloading at time slot τ; let v r Node v i Offload the task to v at time slot τ s Node v j The time cost of unloading tasks from roadside units to edge servers is The time cost is given by the formula Calculate, d(v i ,v j ) indicates v r Node v i and v s Node v j The Euclidean distance of v i ∈V r ,v j ∈V s calculate; Input the task offloading requirements of each regional terminal and output the resource reservation strategy that each sub-region edge server can accept. and the resource reservation policy received by the cloud Specifically: First, input the historical observation data of terminal task offloading, and predict it through the model STCTOD to obtain the time slot τ k Regional terminal task offloading demand O |S| (τ k ); Next, and the set of regions that do not satisfy resource reservation requests Set to an empty set; for each region i, ν i ∈S, based on the edge server workload threshold coefficient Update edge server in time slot τ k The actual available resources Q i (τ k ); According to the predicted terminal task offloading demand O |S| (τ k ) and the terminal offload task volume ρ calculates the resource reservation request R of area i i (τ k ); if the edge server resources in the current area meet the resource reservation request, the request is accepted; otherwise, the current area is marked and put into the collection For each region i, The Dijkstra algorithm is used to find the region j that is closest to the requested region; if j≠-1, it indicates that a region that meets the conditions exists and the region will receive the resource reservation request; if j=-1, it indicates that no region meets the resource reservation request, and the resource reservation request is uploaded to the cloud; the final output is the resource reservation policy that each region can accept and the resource reservation policy received by the cloud

Images