CN115720226A - Time delay perception resource reservation method based on space-time task demand prediction - Google Patents
Time delay perception resource reservation method based on space-time task demand prediction Download PDFInfo
- Publication number
- CN115720226A CN115720226A CN202211097503.6A CN202211097503A CN115720226A CN 115720226 A CN115720226 A CN 115720226A CN 202211097503 A CN202211097503 A CN 202211097503A CN 115720226 A CN115720226 A CN 115720226A
- Authority
- CN
- China
- Prior art keywords
- region
- task
- time
- sub
- graph
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a time delay perception resource reservation method based on space-time task demand prediction. According to the invention, based on the space and time-varying characteristics of resource requirements, by predicting the regional task unloading requirements, providing a time-delay perception edge server resource reservation strategy facing a space region and designing a task unloading time-delay minimization model based on resource reservation, the delay of terminal task unloading is effectively reduced, and the resource scheduling and operating efficiency of an edge network are improved.
Description
Technical Field
The invention relates to the technical field of edge computing and edge cloud cooperation, in particular to a time delay perception resource reservation method based on space-time task demand prediction.
Background
With the rapid development of the fields of automatic driving, smart transportation, smart cities, and the like, the demand of various types of terminals for resources has significantly increased. Since the terminal device is limited by its own energy supply, computing power, and the like, it is impossible to efficiently execute a computation-intensive task and a long-term stable execution task. The computing and unloading solve the limitation of the terminal equipment on the factors, and the tasks are unloaded to the cloud server or the edge server to assist the terminal equipment in completing computing-intensive tasks and diversified tasks, so that the energy consumption of the terminal equipment can be effectively saved, and the long-term stable operation of the terminal equipment is ensured. Compared with cloud computing, edge computing has the advantages of low task completion time delay and low energy consumption, so that offloading tasks to edge servers is a priority. The existing research on resource reservation of the edge server adopts a historical data statistics-based mode, but the resource demand often has space and time-varying characteristics, so that how to flexibly reserve the resource of the edge server is a direction worthy of research, and the resource scheduling and operating efficiency of an urban-scale edge network is improved.
Disclosure of Invention
The invention provides a time delay perception resource reservation method based on space-time task demand prediction, aiming at the problem that how to flexibly reserve edge server resources facing a space region because the existing various types of terminals have space and time-varying characteristics for resource demands. Firstly, a deep learning model is built to predict the unloading requirements of the space-time associated tasks in the region, then an edge server resource reservation strategy facing the time delay perception of the space region is provided based on the task unloading requirements, a terminal task unloading time delay minimization model based on resource reservation is designed, and the terminal task unloading time delay is minimized through the resource reservation of the edge server in the region, so that the minimum completion time of the terminal unloading tasks in the region is obtained.
The method comprises the following steps:
step 1, dividing the whole area into a plurality of sub-areas, and then predicting the task unloading requirement of each sub-area. The method comprises the steps of dividing a region into a plurality of mutually exclusive sub-regions, and representing the sub-regions by using v and v epsilon S, wherein S represents a set of the sub-regions. The edge servers in each sub-area are integrated into one edge server to simplify the problem.
First, by the graph structureRepresenting the spatial correlation between the sub-regions, each sub-region being represented in the graph as a vertex, when the set S of sub-regions is the set of vertices in the graph, the method is characterized in that the method is an adjacency matrix of a graph, represents the connectivity of vertexes in the graph, and considers four spatial correlation models among sub-regions, wherein the four spatial correlation models comprise a region neighbor graph, a region connectivity graph, a region function correlation graph and a region task type correlation graph. The nonlinear correlation is represented by a graph structure, and the feature extraction and further fusion of the multiple graphs are realized by a multiple graph convolution network.
Taking into account the correlation in time series, using O |S| (τ) represents the order of observations of all sub-regions at time slot τ. Assuming that the observed value of the time slot tau is influenced by the first omega time slots, the regional space-time correlation task unloading demand prediction problem is expressed as a single-step space-time prediction problem with a given time length. By learning the function ΨLearning terminal task unloading demand historical data of previous omega time slotsThereby acquiring the regional terminal task offloading requirement of the time slot tau, the time slot tau k+1 Is influenced by the first omega time slots, O |S| (τ k+1 )=Ψ{O |S| (τ k-ω+1 ),…,O |S| (τ k ) And | S | represents the number of areas, and k represents the number of time slots.
Extracting global characteristics of regions through a graph convolution network, namely, extracting spatial correlation characteristics of the regions and between the regions through a formulaThe features of the current region and the associated region are aggregated. And theta represents an aggregation function, and the feature information of the current region and the feature information of the spatial correlation region are aggregated to form region feature description.A multi-graph convolution operation is shown,the maximum value of the laplacian matrix degree of the graph is represented.
Then the global average pooling operation is carried outAndreducing dimensions to obtain the region description of the current sub-region and the related sub-region thereof, whereinA global average pooling operation is represented as,to aggregate the characteristics of the current sub-region and the associated sub-regions through the multi-graph convolutional network,representing different current areas andwhich correlates the characteristics of the region.
Then pass throughThe operation is to adaptively allocate corresponding weight coefficients for different observation sequences,s represents a weight vector of the temporal observation sequence,represents that the observation sequences at different times are endowed with weight operation W 1 And W 2 Are weight parameters, and σ and φ represent activation functions ReLU and sigmoid, respectively.
By passingGiving a weight coefficient to the original observation sequence to obtain a new observation sequence, k representing the number of time slots, O |S| (τ k ) Order, s (τ), representing all sub-region observations of the kth time slot k ) A weight vector representing the temporal observation sequence for the kth time slot,and representing the operation of obtaining a new observation sequence after the original observation sequence is endowed with weight again.
Aggregating observations of different time slots through a gated recurrent neural network, the gated input sequences of each sub-region being aggregated into a single vector Ω through a shared RNN layer i , m represents the maximum value of the time slot number k, i represents the number of sub-regions, and the maximum value is | S |, W 3 The weight parameter is represented by a weight value,and representing a new observation sequence obtained after weighting coefficients are given to different time slots. And finally, converting the extracted features into a prediction result through a full connection layer, namely the unloading requirements of the space-time associated tasks in the region.
Therefore, a Task Offloading Demand prediction model (STCTOD for short) based on the Correlation of the regional spatial Correlation and the temporal Correlation is constructed.
Step 2, predicting the task unloading demand of each sub-region according to the STCTOD model constructed in the step 1, and respectively sending resource reservation requests to sub-region edge servers by the cloud server according to the predicted values to reserve resources owned by the sub-regions; if the resource reservation request of the sub-areas is not satisfied, performing resource scheduling between the sub-areas; the areas where the resource reservation request has not yet been fulfilled will be received by the cloud.
The cloud server should be the last choice for the resource reservation request, since long distance transmission with the cloud server would result in high time delay and excessive consumption of resources such as bandwidth. And performing minimum calculation on the time cost for scheduling the resources between the sub-area which is not satisfied by the resource reservation request and the sub-area which is rich in the resources, wherein the time cost comprises task migration time and task calculation time.
The method comprises the following specific steps:
constructing an edge network area topological graph based on end edge cloud cooperation, and defining a time consumption model, wherein the time consumption model mainly comprises the cost of task migration between areasAnd time cost calculated at the edge serverDefining a terminal task unloading time delay minimization problem based on regional resource reservation; to minimize the terminal task offloading delay, the time cost of edge server resource reservation should be minimized.
In order to ensure the normal work of the edge server, the actual workload of the edge server is less than the maximum workload of the edge server, and a threshold coefficient of resource use in the edge server is set to be in the middle of the scope of E; constructing a task unloading delay minimization model based on resource reservation:
is indicated at time slot τ k M represents the maximum value of the time slot number k, and the formula (2) shows whether the current task needs to be unloaded from the current edge server v i Migration, V s The method comprises a cloud server and an edge server, wherein F represents the task calculation amount requested by the current task, and F i Representing task computation amount of edge server or cloud server, and defining decision variablesIndicating that task i migrates to other edge servers or cloud servers at time slot τ.The representation is migrated to the other edge server,representing migration to a cloud server; equation (3) shows that when a task needs to be migrated from a current edge server, the task is migrated to other edge servers or cloud servers, and decision variables are definedRepresenting edge servers v i Whether the task(s) on need to be migrated.Indicating that the task is executing on the current edge server,indicating that the task needs to be migrated from the current edge server to other edge servers or cloud servers for calculation; equation (4) indicates the computing power h requested by the current task i The computing capacity threshold H of the edge server or cloud server should not be exceeded i (ii) a Equation (5) indicates the final resource reservation value a of each region i (τ) should not exceed the threshold R of resources available for the current region i (τ)×∈。
Inputting the task unloading requirement of each area terminal, and outputting the resource reservation strategy receivable by each sub-area edge serverAnd cloud received resource reservation policies
Compared with the prior art, the invention has the advantages and positive effects that:
1. from the spatial factor analysis, there are complex spatial correlations between different regions. For example, the traffic flow and terminal task offloading needs of a certain area are often affected by the traffic flow and associated area task offloading needs of neighboring areas. From time factor analysis, correlation also exists between terminal task unloading requirement observed values of different time periods in a continuous time sequence. For example, terminal task offload demand forecasts are typically correlated with historical data of their offload tasks. Therefore, the prediction of the regional space-time correlation task unloading requirements has complex regional space-time correlation, the method adopts the graph structure to construct four spatial correlation models among regions, the multi-graph convolution network and the global gated recurrent neural network based on the attention mechanism to accurately predict the terminal task unloading requirements, effectively reduces the delay of terminal task unloading, helps the city-level scale edge network to efficiently schedule computing resources, and improves the city operation efficiency.
2. In order to solve the problem of resource over-scheduling caused by real-time edge server resource allocation in the task unloading process, the resource scheduling is effectively reduced, the resource scheduling efficiency is improved, and the terminal task unloading cost is saved through regional resource reservation. The invention effectively deals with the outbreak of the task unloading requirement in the local time period and the region through the reasonable regional resource reservation model, reduces the resource waste caused by scheduling, and supports the long-term and high-efficiency operation of the system.
3. The task unloading time delay minimization model based on resource reservation is constructed, migration time of tasks between edge server nodes and terminal unloading task calculation time are considered, the resources of the edge servers in the reserved area are reserved, so that the tasks unloaded by the terminals are executed in the area as far as possible, resource waste caused by migration of the tasks between the edge servers is reduced, and finally the completion time of the tasks unloaded by the terminals in the area is minimized.
Drawings
Fig. 1 is an overall architecture diagram of the present invention.
FIG. 2 is a diagram of region division in the embodiment
FIG. 3 is a schematic diagram of a spatial correlation model.
FIG. 4 is a schematic diagram of a time correlation model.
Fig. 5 is a topology diagram of an edge network area.
Detailed Description
The invention is described in further detail below with reference to the following figures and embodiments:
taking a car networking scenario as an example, due to high mobility of a vehicle terminal device and outburst traffic of a local period and area, scheduling of resources becomes longer in path, higher in cost and lower in efficiency. If the resource reservation can be carried out for the regional resource demand in advance, the resource scheduling can be effectively reduced, the resource scheduling efficiency is improved, and the task unloading cost is saved. Therefore, it is very practical to unload demand forecast through effective regional tasks so as to reserve resources of the region.
As shown in fig. 1, a time-delay-aware resource reservation method based on space-time task demand prediction decomposes the problem of time-delay-aware resource reservation based on space-time task demand prediction into task unloading demand prediction and time-delay-aware resource reservation.
Firstly, aiming at space-time correlation task unloading demand prediction, a space-time deep learning model STCTOD prediction region task unloading demand is constructed, and a graph convolution neural network and a gated cyclic neural network based on an attention mechanism are fused for extracting the space correlation and the time correlation of the highly complex task unloading demand between regions. And a second step of providing a resource reservation strategy of the edge server of the time delay perception area, constructing an edge network area topological graph based on end edge cloud cooperation, and constructing an area edge server resource reservation model based on the edge network area topological graph. Next, a time consumption model is defined, mainly including the cost of task migration between regions and the time cost computed at the edge server. Therefore, a task unloading time delay minimization model (ESRR) based on resource reservation is constructed, and the task unloading time delay of the terminal is minimized through resource reservation of the regional edge server.
The specific implementation process of this embodiment is as follows:
step 1: a task unloading demand prediction model STCTOD based on regional spatial correlation and temporal correlation is constructed through a spatio-temporal multi-graph convolution network and a gated cyclic neural network based on an attention mechanism, so that regional task unloading demands are obtained.
As shown in FIG. 2, in order to better predict the terminal task unloading demand, the whole area is divided into 9 sub-areas of 3 × 3, i.e., v i I =1,2, \8230; |, | S |, where | S |, isL =9. Different types of terminals are distributed in sub-regions, such as automobiles, various terminals have a series of tasks to be unloaded, the tasks are unloaded to an edge server through a road edge unit in a wireless transmission mode, the edge server in each sub-region is integrated into one edge server to simplify the problem, then the task unloading requirement of each sub-region is predicted, a deep learning model STCTOD for predicting the task unloading requirement of each sub-region is constructed, the space-time correlation of each sub-region task is considered, the spatial correlation of the task unloading requirement is respectively considered from four dimensions of region proximity, region connectivity, region function correlation and region task type correlation, the non-linear correlation is represented through a graph structure, and the feature extraction and further fusion of multiple graphs are realized through a multi-graph convolution network. Fig. 3 shows four graph structures of the spatial correlation model.
Use of graph structureRepresents the spatial correlation between sub-regions, where S is the set of vertices in the graph, represents the set of regions,is the adjacency matrix of the graph, representing the connectivity of the vertices in the graph.The degree matrix of the graph represents the degree of the nodes in the set S, and i is the number of the sub-regions and is 9 at most. The Laplacian matrix of the graph is obtained by an adjacency matrix A and a degree matrix D of the graph, and the symmetric normalized Laplacian matrix L = I-D -1/2 AD -1/2 Wherein I is an identity matrix.
Region neighbor mapRepresenting the correlation of spatial proximity between sub-regions by formulaIn a 3 × 3 gridThe heart region has a maximum of eight neighbor regions.
Having connectivity between two sub-areas means that the two sub-areas can be directly connected by a road (e.g. an expressway), by formulaIt is shown that, because the region neighborhood relationship in the region neighbor graph and the connection relationship in the region connection graph have a certain degree of overlap, the region connection graph is simplified in order to reduce complexity, the correlation relationship already shown in the region neighbor graph is excluded to make the region connection graph become a sparse graph, and the above formula can be expressed as
Different areas are positioned differently in the plan, resulting in different functions carried by the areas, and areas with the same functional positioning also have space-time correlation. Therefore, when predicting the task unloading demand of the terminal, the correlation relationship of areas with similar functions needs to be considered, and the correlation relationship is calculated through a formula Show, useRepresenting a region v i And v j In a region of functional dependency therebetween, whereinAndrespectively represent regions v i And v j The region feature vector of (1). When the functional characteristics between the two areas are more overlapped, the functional characteristics between the two areas are representedThe higher the degree of association.
The task types of different areas can also indicate the correlation among the areas, the correlation is judged by judging the overlapping degree of the task labels in the different areas, and the formula is used for judging the correlation And (4) showing. WhereinRepresenting a region v i And v j Regional task type dependencies between, whereinAndrespectively represent regions v i And v j A regional task type vector. When the task type labels between two areas are more overlapped, the higher the association degree between the two areas is.
The multi-graph feature extraction and further fusion are realized through a multi-graph convolution network, and the spatial correlation relationship between the regions is obtained. By the formula X I+1 =σ(∑ A∈Z f(A;θ i )X l W l ) And performing graph convolution operation.Andrespectively representing the regional feature vectors, P, of layer l +1 and layer l in the graph convolutional network l And P I+1 Indicating the l-th and l + 1-th layer graph convolution networks. σ denotes the activation function ReLU. f (A; theta) i ) The polynomial function of the K-th order of the laplacian matrix L of the diagram is represented.A feature transformation matrix is represented. A schematic diagram of the model for the multi-graph convolution is shown in fig. 3.
The correlation on the time series is modeled by a recurrent neural network, i.e. a globally gated recurrent neural network based on attention mechanism. First, the features of the current region and the associated region are aggregated by a multi-graph convolutional network, i.e., by a formulaObtaining a theta which represents an aggregation function, aggregating the feature information of the current region and the feature information of the spatial correlation region,representing the global characteristics of the region. Feature extraction of spatial correlation is achieved by a multi-graph convolution operation, in whichA multi-graph convolution operation is shown,the maximum value of the laplacian matrix degree of the graph is represented. Order toRepresents the regional observations at time slot τ, where P represents the dimension of the regional feature vector.
Next, the region observation information is further processed based on the attention mechanism, and weighting coefficients are given to the different time slot observations. And compressing the observation information of each region of the time slot tau through a global average pooling operation to extract global features of the regions. By the formula An average global pooling is performed. WhereinRepresenting a global average pooling operation, | S | representing the number of zones,representing characteristics of different current regions and their associated regions. And secondly, performing an attention-based mechanism, namely allocating weights to different observation sequences, adaptively giving higher weight coefficients to the key information, enabling the key information to have greater influence on the result, and discarding outdated information to a certain extent. By passingThe operation is to adaptively allocate corresponding weight coefficients for different observation sequences,s represents a weight vector of the temporal observation sequence,represents that the observation sequences at different times are endowed with weight operation W 1 And W 2 Is a weight parameter, σ and φ represent the activation functions ReLU and sigmoid, respectively.
By passingGiving a weight coefficient to the original observation sequence to obtain a new observation sequence, k representing the number of time slots, O |S| (τ k ) Order, s (τ), representing all sub-region observations of the kth time slot k ) A weight vector representing the temporal observation sequence for the kth time slot,and representing the operation of obtaining a new observation sequence after the original observation sequence is endowed with weight again.
Aggregating observations of different time slots through a gated recurrent neural network, the gated input sequences of each sub-region being aggregated into a single vector Ω through a shared RNN layer i , m represents the maximum value of the number k of time slots, i represents the number of subregions, and the maximum value is | S |, W 3 The weight parameter is represented by a weight value,representing new observation sequences obtained after weighting coefficients are given to different time slots, wherein m represents the maximum value of the number k of the time slots, i is the number of subregions, and the maximum value is | S |. And finally, converting the extracted features into a prediction result through a full connection layer, namely, the unloading requirements of the regional space-time associated tasks.
The modeling of the time correlation is realized through the steps, the influence of historical data of different time slots on the prediction can be better captured, and the model of the time correlation is shown in FIG. 4. First, ω observation sequences { O are inputted |S| (τ k-ω+1 ),…,O |S| (τ k ) Get all the average pooling operationAndreducing dimensions to obtain region description of current region and its correlation region, and further processingThe operation is to self-adaptively allocate corresponding weight coefficients for different observation sequences, and finally, to passAnd giving a weight coefficient to the original observation sequence to obtain a new observation sequence, wherein the weight coefficient of the observation sequence represents the importance degree of the observation sequence, namely the higher the weight is, the greater the influence on the prediction result of the terminal task unloading demand is.
Step 2: and performing resource scheduling between the regions based on the predicted values to meet the resource reservation request of the regions, and constructing a time delay perception edge server resource reservation model facing the spatial regions.
Firstly, an edge network area topological graph based on edge cloud cooperation is constructed and divided into four layers, namely a terminal equipment layer, a roadside unit layer, an edge layer and a cloud layer. The terminal equipment layer consists of terminals, and the terminals have a series of tasks to be unloaded. The roadside unit layer is used for connecting various terminals and providing relay communication service for task unloading of the terminals. The edge layer is composed of edge servers widely distributed in cities and used for bearing computation tasks unloaded by various terminals of the terminal equipment layer, and edge nodes can be connected with each other within a communication range. The cloud layer is composed of a powerful server cluster and is used for bearing tasks which cannot be executed by the edge server and tasks with high calculation amount, and the calculation pressure of the edge server is relieved. The edge network topology is shown in fig. 5.
Constructing a network graph G (V, E) represents an edge network area topology graph, where V represents a set of nodes, i.e., V = { V = { (V) } i I =1, \8230;, M }, M representing the maximum value of the number of nodes. E represents a set of edges, i.e. network communication links,and | E | is the maximum value N of the number of edges. The node set V is composed of clouds, edge servers, roadside units and terminal equipment. The invention uses v without confusion 0 In particular cloud servers, v s In particular edge servers, v r In particular roadside units, v m In particular to terminal equipment. The node set V is divided into three subsets, and V is used respectively s 、V r And V m Is shown in which V s Including cloud servers and edge servers, V r Comprising roadside units, V m Involving a variety of terminal equipment, i.e.e jk Representing a slave node v j To v k A communication link of (a). Considering time-varying factors in the process of task unloading, the invention takes timeDiscrete by n time slots of equal spacing, using τ k K =1,2, \8230n. Time slot tau k Using t at the start time of k K =0,1,2, \ 8230;, n denotes. Each time t k Determining time slot τ k The value of (d) is determined. The time slot lying between two moments, e.g. time slot τ k At time t k-1 And t k In between.
The resource reservation of the regional edge server means that the task unloading requirement of the regional terminal is obtained through the model STCTOD, so that sufficient computing resources of the edge server are reserved for the region. Let ρ be i (τ) denotes a terminal device v i The task capacity at time slot τ is unloaded. Let R i (τ) represents the amount of terminal task offload for region i, R without confusion i (τ) also denotes a resource reservation request value, R, for the region i i (τ) is affected by the amount of terminal offload tasks for zone i, byv j At v i Within the range, this formula is calculated. The total resource amount of the edge server in the area i in the time slot tau is Q i (τ). When the edge server resource in area i is not enough to satisfy the resource reservation request of the current area, i.e. R i (τ)>Q i (τ), a resource reservation request migration will be sent to region j. If the region j exists, the resource reservation request of the region i is made to satisfy R i (τ)≤Q j (τ), which is then referred to as feasible region resource reservation request migration from region i to region j, using p ij (τ) indicates a resource reservation request value of | p ij (τ) |. When a given region i, for any region j, the condition R is satisfied i (τ)>Q j (τ), at this point the resource reservation request is split into two parts for processing separately. The edge server of the first part searching area j reserves resources with the resource value of | p ij (τ) |. Another part of resource reservation request is sent to the cloud, and the resource reservation request quantity is defined as b i ,b i =R i (τ)-|p ij (τ) |. Defining a variable a i Final resource Prep representing regional edge Server iAnd (5) reserving the value. Defining two set variablesAndrespectively containing the edge server resource amount required to be reserved in each area and the cloud server resource reservation amount required by each area, namely
The time cost for unloading the terminal task mainly comprises four parts, namely the time cost for transmitting the task to the roadside unit by the terminalTime cost of roadside unit offloading tasks to edge serversTime cost of task migration between edge servers or from edge server to cloudTime cost of an edge server or cloud to complete a task
Order toIs v is m Node v i Transmitting tasks to v at time slots τ r Node v j Of the speed of (c). Then at time slot τ, the time cost of the terminal's offloaded task transfer to the roadside unitBy the formula Calculation of rho i (τ) denotes a terminal device v i The task capacity at time slot τ is unloaded. Order toIs v is r Node v i Offloading tasks to v at time slot τ s Node v j Speed of (2), time cost of offloading of tasks from roadside units to edge serversTime cost of (2) is expressed by Calculation of d (v) i ,v j ) Denotes v r Node v i And v s Node v j By the formulaAnd (4) calculating.
And finally, designing a task unloading delay minimization model (ESRR) based on resource reservation. Since offloading tasks to cloud servers can result in high latency and under-utilization of edge servers, an edge server resource reservation based approach is employed to address this problem, i.e., the regional edge servers try to receive resource reservation requests. The time consumption for a task to migrate between edge servers is mainly determined by the migration distance, i.e. finding the path from area i to j is shortest. The Dijkstra algorithm can effectively solve the problem of the shortest path of a single source. Additionally, the time for the edge server to perform the task is also considered. Formula (II)Representing edge servers v i Whether or not the task onMigration is required, wherein the time cost influenced by the edge server resource reservation policy is task migration timeAnd task computation timeThus, simplification to the formula when considering edge server resource reservation
By the formulaCalculating, defining decision variablesIndicating that task i migrates to other edge servers or cloud servers at time slot τ.The representation is migrated to the other edge server,representing migration to a cloud server. Order toIndicating task slave edge server node v at time slot τ i Migrating to other edge server nodes v j Of the speed of (c). Order toIndicating task slave edge server node v at time slot τ i Migration to v 0 The rate of the node. d (v) i ,v j ) Is v is r Node v i And v s Node v j The euclidean distance of (c).Is calculated by the formula Wherein F i (τ) denotes edge Server v i Or cloud server v 0 Task computation amount of h i (τ) denotes edge Server v i Or cloud server v 0 The computing power of (a).
And finally, selecting a region edge server with the minimum sum of the migration time and the execution time to receive the resource reservation request. The input of the ESRR model comprises region terminal task unloading historical observation data O |S| (τ 1 ),O |S| (τ 2 ),…,O |S| (τ k-1 ) Average unload workload ρ of terminal, edge server in time slot τ k Available resource Q (τ) k ) Edge server workload threshold coefficient ∈ and at time slot τ k Is transmitted at a rate of transmission alpha (tau) k )、β(τ k )、γ(τ k )、η(τ k ). The output of the model comprises a resource reservation strategy of the regional edge serverResource reservation policy for cloudThe specific implementation steps of the model are as follows. Firstly, inputting terminal task unloading historical observation data, predicting through a model STCTOD, and acquiring a time slot tau k Regional terminal task offload requirement O |S| (τ k ). Next, the following steps are carried outAnd a set of regions that do not satisfy the resource reservation requestSet as an empty set. For each region i, v i E is S, and the working load threshold coefficient E of the edge server is used for updating the time slot tau of the edge server k Actual available resource Q i (τ k ). Offloading demand according to predicted terminal task O |S| (τ k ) And resource reservation request R of terminal unloading task amount rho calculation area i i (τ k ). The resource reservation request is received if the edge server resources of the current area satisfy the request. Otherwise, marking the current area and putting the current area into the setFor each region And finding the nearest region j to the request region by using a Dijkstra algorithm. If j ≠ -1, it indicates that an area satisfying the condition exists, which will receive the resource reservation request. If j = -1, indicating that no region satisfies the resource reservation request, the resource reservation request is uploaded to the cloud. Final output resource reservation strategy receivable for each areaAnd cloud received resource reservation policy
Claims (1)
1. A time delay perception resource reservation method based on space-time task demand prediction is characterized in that: the method specifically comprises the following steps:
step 1, dividing the whole area into a plurality of sub-areas, and predicting the task unloading requirement of each sub-area; dividing the region into a plurality of mutually disjoint sub-regions, and using v, v belongs to S to represent the sub-regions, wherein S represents a set of the sub-regions; integrating the edge servers in each sub-area into one edge server to simplify the problem;
first, by the graph structureRepresenting the spatial correlation between the sub-regions, each sub-region being represented in the graph as a vertex, when the set S of sub-regions is the set of vertices in the graph,the method is characterized in that the method is an adjacency matrix of a graph, represents the connectivity of vertexes in the graph, and considers four spatial correlation models among sub-regions, wherein the four spatial correlation models comprise a region neighbor graph, a region connectivity graph, a region function correlation graph and a region task type correlation graph; the nonlinear correlation is represented by a graph structure, and the feature extraction and further fusion of the multiple graphs are realized by a multiple graph convolution network;
taking into account the correlation in time series, using O |S| (τ) represents the order of observations of all sub-regions at time slot τ; assuming that an observed value in a time slot tau is influenced by the first omega time slots, expressing a regional space-time correlation task unloading demand prediction problem as a single-step space-time prediction problem with a given time length; by learning functionsLearning historical data of terminal task unloading demands of the first omega time slots so as to acquire regional terminal task unloading demands of the time slot tau, wherein the time slot tau k+1 Is influenced by the first omega time slots, O |S| (τ k+1 )=Ψ{O |S| (τ k-ω+1 ),…,O |S| (τ k ) H, S represents the number of areas, and k represents the number of time slots;
extracting global characteristics of regions through a graph volume network, namely extracting spatial correlation characteristics of the regions and between the regions through a formulaAggregating features of the current region and the associated region; wherein Θ represents an aggregation function, and aggregates the feature information of the current region and the feature information of the spatial correlation region to form region feature description;a multi-graph convolution operation is shown,represents the maximum value of the laplacian matrix degree of the graph;
then through the global average pooling operationAndreducing dimensions to obtain the region description of the current sub-region and the related sub-region thereof, whereinA global average pooling operation is represented as, to aggregate the characteristics of the current sub-region and the associated sub-regions through the multi-graph convolutional network,features representing different current regions and their associated regions;
then pass throughOperate to adaptively assign corresponding weights to different observation sequencesThe weight coefficient of the mixture is as follows,s represents a weight vector of the temporal observation sequence,showing that the observation sequences at different times are given weight operations, W 1 And W 2 Is a weight parameter, sigma and phi respectively represent an activation function ReLU and sigmoid;
by passingGiving a weight coefficient to the original observation sequence to obtain a new observation sequence, k representing the number of time slots, O |S| (τ k ) Order, s (τ), representing all sub-region observations of the kth time slot k ) A weight vector representing the temporal observation sequence for the kth time slot,representing the operation of obtaining a new observation sequence after the original observation sequence is endowed with the weight again;
aggregating observations of different time slots through a gated recurrent neural network, the gated input sequences of each sub-region being aggregated into a single vector Ω through a shared RNN layer i , m represents the maximum value of the time slot number k, i represents the number of sub-regions, and the maximum value is | S |, W 3 The weight parameter is represented by a weight value,representing new observation sequences obtained after weighting coefficients are given to different time slots; finally, the extracted features are converted into a prediction result through a full connection layer,namely, unloading requirements of the space-time associated tasks in the region;
thereby constructing a task unloading demand prediction model STCTOD based on the correlation of the regional spatial correlation and the temporal correlation;
step 2, predicting the task unloading requirement of each sub-region according to the STCTOD model constructed in the step 1, and respectively sending resource reservation requests to a sub-region edge server by a cloud server according to the predicted values to reserve resources owned by the sub-regions; if the resource reservation request of the subareas is not met, resource scheduling is carried out among the subareas; areas where the resource reservation request has not yet been satisfied will be received by the cloud;
the method comprises the following specific steps:
constructing an edge network area topological graph based on end edge cloud cooperation, and defining a time consumption model, wherein the time consumption model mainly comprises the cost of task migration between areasAnd time cost calculated at the edge serverDefining a terminal task unloading time delay minimization problem based on regional resource reservation; in order to minimize the time delay of terminal task unloading, the time cost of edge server resource reservation should be minimized;
in order to ensure the normal work of the edge server, the actual workload of the edge server is less than the maximum workload of the edge server, and a threshold coefficient of resource use in the edge server is set to be in the middle of the scope of E; constructing a task unloading delay minimization model based on resource reservation:
is indicated in a time slot tau k M represents the maximum value of the time slot number k, and the formula (2) shows whether the current task needs to be unloaded from the current edge server v i Migration, V s The method comprises a cloud server and an edge server, wherein F represents the task calculation amount requested by the current task, and F i Representing task computation amount of edge server or cloud server, and defining decision variablesRepresenting that the task i is migrated to other edge servers or cloud servers in the time slot tau;the representation is migrated to the other edge server,representing migration to a cloud server; equation (3) shows that when a task needs to be migrated from a current edge server, the task is migrated to other edge servers or cloud servers, and decision variables are definedRepresenting edge servers v i OnWhether the task needs to be migrated;indicating that the task is executing on the current edge server,indicating that the task needs to be migrated from the current edge server to other edge servers or cloud servers for calculation; equation (4) indicates the computing power h requested by the current task i The computing capacity threshold H of the edge server or cloud server should not be exceeded i (ii) a Equation (5) indicates the final resource reservation value a of each region i (τ) should not exceed the threshold R of resources available for the current region i (τ)×∈;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211097503.6A CN115720226A (en) | 2022-09-08 | 2022-09-08 | Time delay perception resource reservation method based on space-time task demand prediction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211097503.6A CN115720226A (en) | 2022-09-08 | 2022-09-08 | Time delay perception resource reservation method based on space-time task demand prediction |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115720226A true CN115720226A (en) | 2023-02-28 |
Family
ID=85254048
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211097503.6A Pending CN115720226A (en) | 2022-09-08 | 2022-09-08 | Time delay perception resource reservation method based on space-time task demand prediction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115720226A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116366133A (en) * | 2023-04-06 | 2023-06-30 | 广州爱浦路网络技术有限公司 | Unloading method and device based on low-orbit satellite edge calculation |
-
2022
- 2022-09-08 CN CN202211097503.6A patent/CN115720226A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116366133A (en) * | 2023-04-06 | 2023-06-30 | 广州爱浦路网络技术有限公司 | Unloading method and device based on low-orbit satellite edge calculation |
CN116366133B (en) * | 2023-04-06 | 2023-10-27 | 广州爱浦路网络技术有限公司 | Unloading method and device based on low-orbit satellite edge calculation |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Kishor et al. | Task offloading in fog computing for using smart ant colony optimization | |
Tran-Dang et al. | Reinforcement learning based resource management for fog computing environment: Literature review, challenges, and open issues | |
WO2022171066A1 (en) | Task allocation method and apparatus based on internet-of-things device, and network training method and apparatus | |
US11172035B2 (en) | Data management for edge computing environment | |
CN113703984B (en) | Cloud task optimization strategy method based on SOA (service oriented architecture) under 5G cloud edge cooperative scene | |
CN116541106B (en) | Computing task unloading method, computing device and storage medium | |
Wang et al. | A reinforcement learning approach for online service tree placement in edge computing | |
Ibn-Khedher et al. | Edge computing assisted autonomous driving using artificial intelligence | |
KR20210063990A (en) | Method of machine learning based unmanned aerial vehicle mobile edge server collabrative task matching and offloading | |
CN115720226A (en) | Time delay perception resource reservation method based on space-time task demand prediction | |
Santos et al. | Resource provisioning in fog computing through deep reinforcement learning | |
Nguyen et al. | Flexible computation offloading in a fuzzy-based mobile edge orchestrator for IoT applications | |
Li et al. | Deep reinforcement learning for load balancing of edge servers in iov | |
Hu et al. | Joint load balancing and offloading optimization in multiple parked vehicle‐assisted edge computing | |
Gong et al. | Dynamic resource allocation scheme for mobile edge computing | |
Hussain et al. | Resource allocation of industry 4.0 micro-service applications across serverless fog federation | |
Zhang et al. | An Efficient Caching and Offloading Resource Allocation Strategy in Vehicular Social Networks | |
Zhang et al. | Offloading demand prediction-driven latency-aware resource reservation in edge networks | |
Chen et al. | Knowledge distillation for mobile edge computation offloading | |
CN115473838A (en) | Network request processing method and device, computer readable medium and electronic equipment | |
Zhang et al. | A hybrid fast inference approach with distributed neural networks for edge computing enabled UAV swarm | |
CN116109058A (en) | Substation inspection management method and device based on deep reinforcement learning | |
Midya et al. | Pso based optimized resource allocation in three tier cloud architecture for vanet | |
CN114741191A (en) | Multi-resource allocation method for compute-intensive task relevance | |
Hsieh et al. | Deep reinforcement learning-based task assignment for cooperative mobile edge computing |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |