Clustering-based directed acyclic graph cluster task unloading method and system

Abstract

The invention provides a directed acyclic graph cluster task unloading method and system based on clustering, and belongs to the field of information energy system optimizing.The method comprises the steps that a basic model is established; constructing an energy consumption minimization problem of minimizing the terminal equipment; quantizing energy requirements of different directed acyclic graphs by adopting a CCR value, and performing clustering grouping on clusters of the directed acyclic graphs by adopting an iterative self-organizing data analysis technology algorithm based on the CCR value; an incremental control scheme is adopted for a cluster of the compute-intensive directed acyclic graph; aiming at the cluster of the communication-intensive directed acyclic graph, adopting a swarm intelligence task unloading scheme to obtain a calculation unloading destination of each real-time unloading task in the cluster of the directed acyclic graph; according to the method, the problem of heterogeneity of terminal equipment in the task unloading process is solved, and the reliability of large-scale information-energy system optimization is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the field of information energy system optimization, and in particular relates to a directed acyclic graph cluster task unloading method and system based on clustering. Background Art

[0002] In recent years, Cyber-Physical Systems (CPS) have been widely used in industrial manufacturing and our daily life. End-to-end collaboration, combined with mobile edge computing and device-to-device communication, is a promising computing approach that can meet the stringent real-time requirements of large-scale CPS applications. However, due to the limited energy supply and the inherent openness of terminal devices, energy and reliability issues should be carefully addressed in large-scale CPS with end-to-end collaboration.

[0003] CPS is a tight integration of hardware and software components achieved by leveraging advanced networking, sensing, computing, control, and communication technologies. Nowadays, the spatiotemporal dimensions of CPS are increasing due to the popularity of interconnected terminal devices. In recent years, an attractive paradigm of coupling mobile edge computing (MEC) and device-to-device (D2D) communication technologies has attracted widespread attention. In the general MEC approach, real-time offloaded tasks on a single terminal device can complete their execution locally or offload computing instructions to the MEC server. However, this computing approach cannot handle the heterogeneity of terminal devices during task offloading well, resulting in unbalanced resource utilization among terminal devices in the local network. System reliability cannot be guaranteed when dealing with energy minimization problems. Therefore, it is urgent to explore reliability optimization methods for large-scale information-energy systems with end-to-end collaboration, which has very broad application prospects.

[0004] Existing research on large-scale information-energy systems usually has the following problems:

[0005] 1. System reliability cannot be guaranteed when dealing with energy minimization issues;

[0006] 2. There is a lack of effective clustering methods to analyze heterogeneous energy consumption components to quantify different energy requirements. Summary of the invention

[0007] In view of the deficiencies in the prior art, the present application proposes a directed acyclic graph cluster task offloading method and system based on clustering.

[0008] In a first aspect, the present application proposes a directed acyclic graph cluster task offloading method based on clustering, comprising:

[0009] Establishing a basic model, wherein the basic model includes: hardware architecture, application model, reliability model and energy model;

[0010] According to the basic model, a problem of minimizing the energy consumption of terminal equipment is constructed;

[0011] The CCR value is used to quantify the energy requirements of different directed acyclic graphs, wherein the directed acyclic graphs represent the dependency between the real-time offload tasks of different applications;

[0012] Based on the CCR value, an iterative self-organizing data analysis technology algorithm is used to cluster and group the clusters of the directed acyclic graph to obtain a clustering result, wherein the clustering result includes: a cluster of a computationally intensive directed acyclic graph and a cluster of a communication-intensive directed acyclic graph, wherein the cluster of the directed acyclic graph is a plurality of different directed acyclic graphs, and the directed acyclic graph includes a plurality of different applications;

[0013] For a cluster of computationally intensive directed acyclic graphs, an incremental control scheme is adopted according to the energy consumption optimal solution of the energy consumption minimization problem to obtain the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph, wherein the destination includes: a local terminal device, a non-local terminal device, and a virtualized container in an edge server S;

[0014] For the cluster of communication-intensive directed acyclic graph, according to the energy-optimal solution of the energy-minimization problem, a swarm intelligence task offloading scheme is adopted to obtain the computational offloading destination of each real-time offloading task in the cluster of directed acyclic graph.

[0015] The hardware architecture is an end-to-end collaborative information-physical system architecture consisting of M terminal devices, a base station B, and an edge server S. The edge server S provides multiple virtualized containers, and different virtualized containers have different computing capabilities.

[0016] The application model includes: associating each terminal device in the information-physical system architecture in a directed acyclic graph, the directed acyclic graph includes: a set of real-time offloading tasks with constrained priorities in the application and a communication data set from one real-time offloading task to another real-time offloading task, the execution modes of the real-time offloading tasks include: local mode, D2D mode and remote execution mode, the local mode is a mode in which the real-time offloading task is completed on the local terminal device without the help of an edge server S or other terminal devices, the D2D mode is a mode in which the calculation of the real-time offloading task is offloaded to another terminal device through a D2D communication link and executed on the other terminal device, and the remote execution mode is a mode in which the real-time offloading task is sent to the edge server S and the real-time offloading task is executed on the edge server S.

[0017] The reliability model is the reliability of the real-time offloading task, and the calculation formula is as follows:

[0018]

[0019] in, To offload the task τ in real time m,i The reliability of To offload the task τ in real time m,i For the reliability of communication in local mode, To offload the task τ in real time m,i The constant bit error rate of the communication link, To offload the task τ in real time m,i The ability to tolerate soft errors at time t, To offload the task τ in real time m,i Constant soft error rate, The correct communication data of all direct nodes is successfully delivered to the real-time offload task τ m,i probability.

[0020] The energy model includes: static energy consumption of the terminal device, dynamic energy consumption of the terminal device when performing real-time offloading tasks in local mode, and D2D mode of the terminal device when performing real-time offloading tasks. m The communication energy of the D2D mode, the energy dissipation of delivering the real-time offload task in the D2D mode, and the energy dissipation of the other terminal device D when processing the real-time offload task in the D2D mode. π The energy consumption of the remote execution mode and the energy consumption of sending real-time offload tasks.

[0021] According to the basic model, the energy consumption minimization problem of the terminal device is constructed, and the calculation formula is as follows:

[0022] The objective function of the problem of minimizing the energy consumption of the terminal device is:

[0023]

[0024] in, Energy for terminal device D m energy consumption, is the total energy consumption, M is the total number of clusters in the directed acyclic graph, and m is the cluster of the mth directed acyclic graph;

[0025] Among them, the terminal device D m The energy consumption is calculated as follows:

[0026]

[0027] in,

[0028]

[0029] in, For terminal device D m The static energy consumption, Terminal device D in local mode m Execute real-time offloading taskτ m,i The dynamic energy consumption when When performing real-time offloading tasks in D2D mode, the terminal device D m The communication energy, When processing real-time offloading tasks in D2D mode, another terminal device D π energy consumption, Delivering real-time offloading tasks in D2D mode m,i The energy dissipation Send real-time offload tasks for remote execution modeτ m,i Energy consumption, τ n,j is a cluster G of a directed acyclic graph n The jth real-time offloading task in (1≤n≤M,1≤j≤ξ n ), M is the total number of clusters in the directed acyclic graph, ξ n G is a cluster of a directed acyclic graph n The total number of real-time offload tasks in τ m,p G is a cluster of a directed acyclic graph m The pth real-time offloading task in m G is a cluster of a directed acyclic graph m The total number of real-time offload tasks in , i is the cluster G of the i-th directed acyclic graph m In the real-time offloading task, h is the h-th container, H is the total number of containers on the edge server, and α m,i is the first and first variable, β n,j,m is the 201st variable, γ m,i,h is a zero-one variable, δ i,j A zero-one variable.

[0030] The constraints of the problem of minimizing the energy consumption of the terminal device are as follows:

[0031] (1) Each real-time offloading task meets the time constraint, and the calculation formula is as follows:

[0032]

[0033] Among them, To offload the task τ in real time m,i The start time, W m,i To offload the task τ in real time m,i The number of CPU instruction cycles at terminal m, f mFor terminal device D m The operating frequency, β m,i,π To offload the task τ in real time m,i In terminal device D m The relationship factor with the cluster π of the directed acyclic graph, π is the cluster type of the directed acyclic graph, T m For terminal device D m The deadline for executing the task;

[0034] (2) Each real-time offloading task can only execute one execution mode, and the calculation formula is as follows:

[0035]

[0036] (3) Each real-time offloading task satisfies the reliability constraint, which is calculated as follows:

[0037]

[0038] in, For the real-time offloading task τ in the reliability model m,i The reliability, R m is the reliability target of the m-th DAG cluster;

[0039] (4) The energy consumption of each terminal device has an upper limit, which can be calculated as follows:

[0040]

[0041] in, For terminal device D m Energy consumption threshold;

[0042] (5) The real-time offloading tasks should satisfy the priority constraint, which is calculated as follows:

[0043]

[0044] in, To offload the task τ in real time m,i The completion time, To offload the task τ in real time m,i The starting time, e m,i,p Real-time offloading taskτ m,i In terminal device D m communication data;

[0045] (6) All real-time offloading tasks are executed within their respective durations without overlap. The calculation formula is as follows:

[0046]

[0047] Among them, ν nis the voltage at point n, ν m is the power supply voltage, Z is a constant, τ n,j G is a cluster of a directed acyclic graph n The jth real-time offloading task in m,i G is a cluster of a directed acyclic graph m The i-th real-time offloading task in To offload the task τ in real time n,j The start time of

[0048] and τ m,i ≠τ n,j , we get the following inequality:

[0049]

[0050] same, π∈[1,M], h∈[1,H], we get the following inequality:

[0051]

[0052] in, To offload the task τ in real time n,j The completion time, To offload the task τ in real time n,j The starting time, α n,j It is the 501st variable.

[0053] The method for obtaining a computing offloading destination of each real-time offloading task in the cluster of the directed acyclic graph of computation-intensive type adopts an incremental control scheme according to the optimal solution of energy consumption of the energy consumption minimization problem, including:

[0054] For the cluster of computationally intensive directed acyclic graphs, calculate the optimal solution to the energy consumption minimization problem of the cluster of directed acyclic graphs;

[0055] Expressing the optimal solution of energy consumption as a function of the dependency relationship between processor characteristics and tasks;

[0056] Under the premise of maintaining the topological order of the directed acyclic graph, all real-time offload tasks in the cluster of the directed acyclic graph are sorted in descending order according to the processor characteristics and the dependencies between the tasks;

[0057] Among the sorted real-time offloading tasks, PID control technology is used to iteratively search for the optimal offloading solution, and the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph is obtained.

[0058] The cluster of the communication-intensive directed acyclic graph adopts a group intelligent task offloading scheme according to the energy consumption optimal solution of the energy consumption minimization problem to obtain the calculation offloading destination of each real-time offloading task in the cluster of the directed acyclic graph, including:

[0059] The PSOSR method is used to transform the cluster of communication-intensive directed acyclic graph into the cluster of computation-intensive directed acyclic graph;

[0060] For the cluster of computationally intensive directed acyclic graphs, calculate the optimal solution to the energy consumption minimization problem of the cluster of directed acyclic graphs;

[0061] Expressing the optimal solution of energy consumption as a function of the dependency relationship between processor characteristics and tasks;

[0062] Under the premise of maintaining the topological order of the directed acyclic graph, all real-time offload tasks in the cluster of the directed acyclic graph are sorted in descending order according to the processor characteristics and the dependencies between the tasks;

[0063] Among the sorted real-time offloading tasks, PID control technology is used to iteratively search for the optimal offloading solution, and the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph is obtained.

[0064] The method of using the PSOSR method to convert a cluster of a communication-intensive directed acyclic graph into a cluster of a computation-intensive directed acyclic graph includes:

[0065] Step S6.1.1: taking all real-time offloading tasks in the cluster of the communication-intensive directed acyclic graph as particles in the particle swarm algorithm;

[0066] Step S6.1.2: Divide the position vector of each particle into a segment encoding continuous variables and a segment encoding discrete variables;

[0067] Step S6.1.3: using the first replication scheme to create a first offspring position vector connected to the encoding continuous variable segment and the discrete variable segment, using the second replication scheme to create a second offspring position vector connected to the encoding continuous variable segment and the discrete variable segment, connecting the first offspring position vector with the second offspring position vector to obtain a complete offspring particle;

[0068] Step S6.1.3: According to step S6.1.2, multiple complete offspring particles are obtained to form initial feasible particles, and the fitness of each initial feasible particle is calculated, where the fitness is the difference between the CCR value of the particle and a predefined first CCR threshold;

[0069] Step S6.1.4: using particle swarm algorithm to iterate particles;

[0070] Step S6.1.5: Determine whether the termination condition is met after iteration;

[0071] Step S6.1.6: If the termination condition is not met, all particles are sorted in descending order according to their fitness. For each particle after sorting, a binary variable is randomly selected to relax the value range of the particle's fitness. Steps S6.1.2 to S6.1.3 are re-executed to regenerate complete offspring particles. Based on multiple complete offspring particles, feasible particles are re-formed.

[0072] Step S6.1.7: Apply feasible particles as real-time offloading tasks to the cluster of the directed acyclic graph;

[0073] Step S6.1.8: Determine whether the CCR value of the cluster of the directed acyclic graph is lower than a predefined second CCR threshold;

[0074] Step S6.1.9: If the CCR value of the cluster of the directed acyclic graph is lower than a predefined second CCR threshold, mark the current cluster of the directed acyclic graph as true;

[0075] Step S6.1.10: Select the particle with the largest fitness in the cluster of the directed acyclic graph marked as true as the optimal solution, and take the optimal solution as the real-time offloading task in the cluster of the converted dense directed acyclic graph;

[0076] Step S6.1.11: If the CCR value of a cluster in the directed acyclic graph is higher than or equal to a predefined second CCR threshold, part of the load is diverted to other clusters in the directed acyclic graph.

[0077] In a second aspect, the present application proposes a directed acyclic graph cluster task offloading system based on clustering, comprising: a model building module, a problem optimization module, a value quantization module, a cluster grouping module, an incremental control module and an intelligent offloading module, wherein the model building module is connected to the problem optimization module, the value quantization module is connected to the cluster grouping module, the problem optimization module and the cluster grouping module are respectively connected to the incremental control module, and the incremental control module is connected to the intelligent offloading module;

[0078] A model building module, used to build a basic model, wherein the basic model includes: a hardware architecture, an application model, a reliability model, and an energy model;

[0079] A problem optimization module, used for constructing a problem of minimizing the energy consumption of a terminal device according to the basic model;

[0080] A value quantification module, used for quantifying energy requirements of different directed acyclic graphs using CCR values, wherein the directed acyclic graphs represent the degree of dependency between real-time offloading tasks of different applications;

[0081] A clustering grouping module is used to cluster and group the clusters of the directed acyclic graph based on the CCR value by using an iterative self-organizing data analysis technology algorithm to obtain clustering results, wherein the clustering results include: clusters of computationally intensive directed acyclic graphs and clusters of communication-intensive directed acyclic graphs, wherein the clusters of the directed acyclic graphs are multiple different directed acyclic graphs, and the directed acyclic graphs include multiple different applications;

[0082] The incremental control module is used for obtaining the computing offloading destination of each real-time offloading task in the cluster of the directed acyclic graph according to the optimal solution of energy consumption of the energy consumption minimization problem, and adopting the incremental control scheme for the cluster of the directed acyclic graph, wherein the destination includes: a local terminal device, a non-local terminal device, and a virtualized container in the edge server S;

[0083] The intelligent offloading module is used for clusters of communication-intensive directed acyclic graphs. According to the energy-optimal solution of the energy-minimization problem, a group intelligent task offloading scheme is adopted to obtain the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph.

[0084] Beneficial effects:

[0085] This application proposes a cluster-based directed acyclic graph cluster task offloading method and system, which uses clustering methods to distinguish the different energy requirements of directed acyclic graph clusters, and can well handle the heterogeneity of terminal devices during task offloading. Different solutions are proposed for clusters of computationally intensive directed acyclic graphs and clusters of communication-intensive directed acyclic graphs, ensuring the optimization reliability of large-scale information-energy systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 Flow chart of the directed acyclic graph cluster task offloading method based on clustering according to an embodiment of the present application;

[0087] Figure 2 A schematic diagram of a cluster-based directed acyclic graph cluster task offloading method flow chart of an embodiment of the present application;

[0088] Figure 3 Comparison diagram of simulation results of the embodiment of the present application;

[0089] Figure 4 A principle block diagram of a directed acyclic graph cluster task offloading device based on clustering according to an embodiment of the present application. DETAILED DESCRIPTION

[0090] The specific implementation of the present application is further described in detail below in conjunction with the drawings and examples.

[0091] In view of the shortcomings of the prior art, the present application proposes a cluster-based directed acyclic graph cluster task offloading method and system, the core ideas of which include: first, construct a basic model, including system architecture, application model, reliability model, and energy model. Then, based on the basic model, a problem of minimizing the energy consumption of terminal devices is formed. Secondly, the communication-computation-ratio (CCR) metric is used to quantify different DAG energy requirements. The DAG clusters are grouped based on the CCR value using the iterative self-organizing data analysis technology algorithm (ISODATA), and a single DAG cluster is divided into two categories: computationally intensive and communication intensive. Thirdly, based on the obtained clustering results, an incremental control scheme for realizing task offloading is designed for computationally intensive DAG clusters. Finally, a group intelligent task offloading scheme for communication-intensive DAG clusters is designed. The present application ensures the reliability of the system by optimizing the reliability of large-scale information-energy systems with end-to-end collaboration.

[0092] Embodiment 1:

[0093] This embodiment proposes a directed acyclic graph cluster task offloading method based on clustering. Figure 1 , Figure 2 As shown, including:

[0094] Step S1: Establish a basic model, which includes: hardware architecture, application model, reliability model and energy model;

[0095] In this embodiment, the hardware architecture is an end-to-end collaborative cyber-physical system architecture consisting of M terminal devices, a base station B, and an edge server S. The edge server S provides multiple virtualized containers, and different virtualized containers have different computing capabilities. Specifically, consider a system consisting of M terminal devices D = {D1, D2, ..., D M}, the end-to-end collaborative large-scale CPS architecture composed of base station B and edge server S, for each terminal device D m (1≤m≤M), whose computing power is provided by a processor Θ m Indicates that its power supply voltage is v m , the operating frequency is f m The edge server S is powered by container-based virtualization technology, which supports virtual packaging and isolation of different applications. Let C = {C1, C2, ..., C H} is the set of total available containers H created by edge server S, then each container C H (1≤h≤H) can be used with tuple C H :{b h ,f h} means that b h is the communication bandwidth, f his the working frequency. Base station B is usually deployed near the edge server S, which acts as a global controller and grasps the entire system information.

[0096] In this embodiment, the application model includes: associating each terminal device in the information-physical system architecture in a directed acyclic graph, the directed acyclic graph includes: a set of real-time offloading tasks with priority constraints in the application and a communication data set from one real-time offloading task to another real-time offloading task, the execution mode of the real-time offloading task includes: local mode, D2D mode and remote execution mode, the local mode is a mode in which the real-time offloading task is completed on the local terminal device without the help of the edge server S or other terminal devices, the D2D mode is a mode in which the calculation of the real-time offloading task is offloaded to another terminal device through a D2D communication link and executed on the other terminal device, and the remote execution mode is a mode in which the real-time offloading task is sent to the edge server S and the real-time offloading task is executed on the edge server S. Specifically: Assume that each terminal device D m With DAG G m =(V m ,E m ) described in the application. is a set of priority constraints in the application with a total number of ξ m Real-time tasks. m ={(e m,i,p ,η m,i,p )1≤i,p≤ξ m ,i≠p} is used to capture task dependencies and m,i To Task E m,i,p The collection of communication data. If the task τ m,p is τ m,i The direct successor of m,i,p Therefore, it is set to 1, that is, the task τ m,p In the task τ m,i Receive a total of η m,i,p The execution of a task cannot start before the communication data. m,i , whose characteristics are described by the tuple τ m,i : where μ m,i ∈[0,1] is the task activity factor, W m,i is the number of CPU instruction cycles, γ m,i is the CPU instruction cycle data volume, T m is the public deadline, R m is the reliability target. In the end-end-edge collaborative system, task τ m,i It is equipped with local, D2D and remote execution modes. In the local execution mode, the task τ m,iIt should be completely completed on the terminal device D without the help of the edge server S or other terminal devices. m In the D2D execution mode, the task τ m,i Able to offload its computation to another terminal device D via a D2D link π (1≤π≤M, π≠m), then request the terminal device D π Complete task execution. Similar to the D2D mode, the remote execution mode allows the task τ m,i Send its calculation to the edge server S, and then request the edge server S to perform the task execution. Since base station B is the global manager, it needs to decide the task execution mode based on system information.

[0097] In this embodiment, the reliability model is the reliability of the real-time offloading task, taking into account the occurrence of bit errors and soft errors. Specifically, bit errors mainly occur on the communication link due to environmental interference or bit synchronization errors. For the task τ m,i The fault tolerance of task τ is m,i Execution mode. In task τ m,i In the case of local execution mode (i.e., ρ = m), the communication reliability is considered to be 1. In addition, when the task τ m,i Unload to terminal device D at time t (i.e., ρ = π or ρ = h) π or container C h hour, as follows:

[0098]

[0099] in, To offload the task τ in real time m,i A constant bit error rate for a communication link. Unlike bit errors, soft errors can occur during the execution of a task on a processor. represents the real-time offloading task τ m,i The ability to tolerate soft errors at time t, the execution reliability is expressed as:

[0100]

[0101] in, To offload the task τ in real time m,i Constant soft error rate. The correct communication data of all direct nodes is successfully delivered to the real-time offload task τ m,i Combining (1) and (2), the real-time offloading task τ m,i Reliability:

[0102]

[0103] in, To offload the task τ in real time m,i The reliability of To offload the task τ in real time m,i For the reliability of communication in local mode, To offload the task τ in real time m,i The ability to tolerate soft errors at time t.

[0104] In this embodiment, the energy model includes: static energy consumption of the terminal device, dynamic energy consumption of the terminal device when performing real-time offloading tasks in local mode, and D2D mode of the terminal device when performing real-time offloading tasks. m The communication energy of the D2D mode, the energy dissipation of delivering the real-time offload task in the D2D mode, and the energy dissipation of the other terminal device D when processing the real-time offload task in the D2D mode. π The energy consumption of the remote execution mode and the energy consumption of sending real-time offload tasks.

[0105] The energy consumption of terminal equipment can be decomposed into static and dynamic components. is the static power D of the terminal equipment m , then the static energy consumption in a scheduling time period T (i.e., the static energy consumption of the terminal device) is obtained by the following formula:

[0106]

[0107] Terminal device D m The dynamic energy consumption depends on the task τ m,i In local execution mode, the processor Θ m In running task τ m,i The dynamic energy consumed when the terminal device performs the real-time offloading task in local mode is expressed as:

[0108]

[0109] In the formula, ψ m For processor Θ m At the same time, the terminal device D m The task τ m,i The output of is passed to the direct successor. Represents the task τ m,i The set of direct successors of , where each element Must be from the task τ m,i Receive a total of η m,i,p Therefore, assuming that in the terminal device D k(1≤k≤M) to perform task τ m,i To task τ m,p When the output result is m The communication energy (i.e., the terminal device D2D performs real-time offloading tasks in D2D mode) m The communication energy) is as follows:

[0110]

[0111] In the formula, For terminal device D m D2D transmission power. m,k For terminal device D m With D k The D2D communication rate between b m,k is the communication bandwidth, g m,k is the channel gain, and w is the background interference. m,p The target (i.e. terminal device D k ) is determined, the above data transmission process will be triggered. Task τ m,i In addition to the local execution mode, you can also choose to offload its computation to the terminal device D π D2D execution mode. Then, the delivery task τ m,i The energy dissipation of (i.e., the energy dissipation of delivering real-time offload tasks in D2D mode) is calculated as

[0112]

[0113] in To handle the offload task τ m,i When the terminal device D π The energy consumption of another terminal device D when processing real-time offloading tasks in D2D mode π Energy consumption):

[0114]

[0115] Among them, T m,i,π For the task τ m,i Edge device D m and D π The D2D communication time between m,i,π =γ m,i / υ m,π Given. The target edge device D π Similarly, the edge device D m To container C h Send task τ m,iThe energy consumption when sending real-time offload tasks in remote execution mode is calculated as:

[0116] in For terminal device D m Transmit power in remote execute mode.

[0117] Step S2: Based on the basic model, construct the energy consumption minimization problem of the terminal device, and the calculation formula is as follows:

[0118] By introducing the binary variable α m,i , β m,i,π , γ m,i,h and δ i,j , a problem calculation formula based on integer linear programming (ILP) is given:

[0119]

[0120] where τ n,j DAGG n The jth task in (1≤n≤M,1≤j≤ξ n ). The ILP objective function is expressed as:

[0121]

[0122] In the objective function, Energy for terminal device D m energy consumption, is the total energy consumption, M is the total number of clusters in the directed acyclic graph; among them, the terminal device D m The energy consumption is calculated as follows:

[0123]

[0124] in, For terminal device D m The static energy consumption, Terminal device D in local mode m Execute real-time offloading taskτ m,i The dynamic energy consumption when When performing real-time offloading tasks in D2D mode, the terminal device D m The communication energy When processing real-time offloading tasks in D2D mode, another terminal device D π energy consumption, Delivering real-time offloading tasks in D2D mode m,i The energy dissipation Send real-time offload tasks for remote execution modeτ m,i Energy consumption, τ n,j G is a cluster of a directed acyclic graph n The jth real-time offloading task in (1≤n≤M,1≤j≤ξ n ), M is the total number of clusters in the directed acyclic graph, ξ n is a cluster G of a directed acyclic graph n The total number of real-time offload tasks in τ m,p G is a cluster of a directed acyclic graph m The pth real-time offloading task in m G is a cluster of a directed acyclic graph m The total number of real-time offload tasks in , i is the cluster G of the i-th directed acyclic graph m In the real-time offloading task, h is the h-th container, H is the total number of containers on the edge server, and α m,i is the first coefficient, β n,j,m is the second coefficient, γ m,i,h is the third coefficient, δ i,j is the fourth coefficient.

[0125] At the same time, in order to generate a feasible task offloading solution, the following linear constraints must not be violated:

[0126] (1) Each task must satisfy its time constraint. For the task τ m,i The start time, then we have:

[0127]

[0128] Among them, To offload the task τ in real time m,i The start time, W m,i To offload the task τ in real time m,i The number of CPU instruction cycles at terminal m, f m For terminal device D m The operating frequency, β m,i,π To offload the task τ in real time m,i In terminal device D m The relationship factor with the cluster π of the directed acyclic graph, π is the cluster type of the directed acyclic graph, T m For terminal device D m The deadline for executing the task.

[0129] (2) Each task can only be executed in one execution mode:

[0130]

[0131] (3) Each task should satisfy the reliability constraints:

[0132]

[0133] in, For the real-time offloading task τ in the reliability model m,i The reliability, R m is the reliability target of the m-th DAG cluster;

[0134] (4) The energy consumption of each terminal device has an upper limit. Let E m The above is the terminal device D m The energy consumption threshold is:

[0135]

[0136] in, For terminal device D m Energy consumption threshold;

[0137] (5) The priority constraints between tasks should be satisfied. For the task τ m,i The completion time is

[0138]

[0139] in, To offload the task τ in real time m,i The completion time, To offload the task τ in real time m,i The starting time, e m,i,p Real-time offloading taskτ m,i Communication data at terminal m;

[0140] (6) All tasks are executed within their durations without overlap. Given two tasks τ m,i and τ n,j , and a sufficiently large constant Z, so the following conditions hold:

[0141]

[0142] Among them, ν n is the voltage at point n, ν m is the power supply voltage, Z is a sufficiently large constant, τ n,j G is a cluster of a directed acyclic graph n The jth real-time offloading task in m,i G is a cluster of a directed acyclic graph m The i-th real-time offloading task in To offload the task τ in real time n,j The start time of

[0143] also, and τ m,i ≠τ n,j , we can easily get the following inequality:

[0144]

[0145] same, π∈[1,M], h∈[1,H], we can infer that:

[0146]

[0147] in, To offload the task τ in real time n,j The completion time, To offload the task τ in real time n,j The starting time, α n,j is the 501st variable,

[0148] Step S3: using CCR values ​​to quantify energy requirements of different directed acyclic graphs, where the directed acyclic graphs represent the degree of dependency between real-time offload tasks of different applications;

[0149] For each application G m , its CCR value CCR m It is estimated by the following formula:

[0150]

[0151] Where a1 and a2 are the energy consumption coefficients of the target system’s computation and communication, respectively, and W m,i To offload the task τ in real time m,i In terminal device D m The number of CPU instruction cycles, η m,i,p To offload the task τ in real time m,i With real-time offloading task τ m,p In terminal device D m communication.

[0152] Step S4: Based on the CCR value, the directed acyclic graph clusters are clustered and grouped using an iterative self-organizing data analysis technology algorithm to obtain clustering results, wherein the clustering results include: a cluster of a computationally intensive directed acyclic graph and a cluster of a communication-intensive directed acyclic graph, wherein the directed acyclic graph clusters are multiple different directed acyclic graphs, and the directed acyclic graphs include multiple different applications;

[0153] For two different applications G m Sets X and W, each with N and M observations:

[0154] X={x1,x2,...,xN} (30)

[0155] W={w1,w2,...,w M} (31)

[0156] (1) The first stage: initial clustering:

[0157] First, the sets in X and W are clustered into K and U homogeneous sets by minimizing the following equation.

[0158]

[0159] Satisfies: K≤N,

[0160]

[0161] Satisfies: U≤M,

[0162] Here, x y and w g are the y-th and g-th sets of observations X and W, respectively, belonging to the k-th cluster and the uth cluster Therefore, μ k and μ u are the initial clustering points of set X and set W respectively. Equations (34) and (35) represent the solutions of equations (32) and (33) respectively.

[0163]

[0164] in, arrive and arrive Represent the homogeneous clusters selected by sets X and W respectively.

[0165] (2) Second stage: hierarchical clustering based on linkage criteria

[0166] The homogeneous clusters in (30) and (31) are further clustered by determining the individual heterogeneous features of the cluster elements and pairing them hierarchically one after another. This is achieved by solving the optimization problem given in (36).

[0167]

[0168] in and From C x and C w The cluster selected in is the solution of (36). In addition, w k,u is based on

[0169] The selection metric derived from the properties of the observation sets X and W. In order to solve the optimization problem, Equation (36) is simplified to Equation (37) to calculate the weighted link ω * , and iteratively solve (36) and (37) until (34) and (10) are simplified to a single cluster.

[0170]

[0171] The entire algorithm finally outputs the final clustering result Φ.

[0172] Step S5: for the cluster of the computationally intensive directed acyclic graph, according to the energy consumption optimal solution of the energy consumption minimization problem, an incremental control scheme is adopted to obtain the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph, wherein the destination includes: a local terminal device, a non-local terminal device, and a virtualized container in the edge server S;

[0173] The method for obtaining a computing offloading destination of each real-time offloading task in the cluster of the directed acyclic graph of computation-intensive type adopts an incremental control scheme according to the optimal solution of energy consumption of the energy consumption minimization problem, including:

[0174] Step S5.1: for a cluster of a computationally intensive directed acyclic graph, calculating an optimal solution to the energy consumption minimization problem of the cluster of the directed acyclic graph;

[0175] Step S5.2: Expressing the optimal solution for energy consumption as a function of the dependency relationship between processor characteristics and tasks;

[0176] Step S5.3: Under the premise of maintaining the topological order of the directed acyclic graph, all real-time offload tasks in the cluster of the directed acyclic graph are sorted in descending order according to the processor characteristics and the dependencies between the tasks;

[0177] Step S5.4: In the sorted real-time offloading tasks, the PID control technology is used to iteratively search for the optimal offloading solution to obtain the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph.

[0178] In this embodiment, according to the clustering results obtained in step S4, an incremental control scheme for implementing task offloading is designed for computationally intensive DAG clusters. According to the average CCR value generated by each DAG cluster, it is divided into a computationally intensive category or a communication intensive category, that is,

[0179]

[0180] if Below a threshold, the DAG cluster is naturally classified as computationally intensive; otherwise, it is classified as communication intensive. is a computationally intensive DAG cluster, and It is a DAG cluster The set of all tasks in . It is a group of terminal devices whose tasks are in the DAG cluster Then, let As a set of tasks that choose local or D2D execution mode. Let τ ε (1≤ε≤φ′) represents the task set The εth element in . Task subset Energy calculation:

[0181]

[0182] in, Represents a set of real-time tasks that are executed locally or offloaded and executed on the terminal device D m In addition, A vector representing storage processor related parameters, where Represents processor Θ m Similarly, let Represents a vector storing task-dependent parameters, where y m Denotes the task subset Ω′ m The power coefficient of . Obviously, Can be rewritten as the product of the processor-dependent parameters and the task-dependent parameters.

[0183] According to the task power coefficient in the set Sort all tasks in descending order while maintaining the topological order. Sort all processors in the processor set Θ from low to high power coefficients. Enter the process of iteratively searching for the optimal task offloading solution using proportional-integral-derivative (PID) control technology. Then determine whether the termination condition is met. If not, set the PID controller to limit the communication energy, that is, any communication energy Cannot exceed the threshold After the PID controller is set up, a suitable container C needs to be selected h As each task τ i Update the task offloading solution obtained by the current iteration Task Set Output task offloading solution Terminate the entire algorithm.

[0184] Step S6: For the cluster of the communication-intensive directed acyclic graph, according to the energy consumption optimal solution of the energy consumption minimization problem, a swarm intelligence task offloading scheme is adopted to obtain the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph.

[0185] The cluster of the communication-intensive directed acyclic graph adopts a group intelligent task offloading scheme according to the energy consumption optimal solution of the energy consumption minimization problem to obtain the calculation offloading destination of each real-time offloading task in the cluster of the directed acyclic graph, including:

[0186] Step S6.1: using the PSOSR method to convert the clusters of the communication-intensive directed acyclic graph into the clusters of the computation-intensive directed acyclic graph;

[0187] Step S6.2: for a cluster of a computationally intensive directed acyclic graph, calculating an optimal solution to the energy consumption minimization problem of the cluster of the directed acyclic graph;

[0188] Step S6.3: Expressing the optimal energy consumption solution as a product of the processor-dependent parameters and the task-dependent parameters;

[0189] Step S6.4: Under the premise of maintaining the topological order of the directed acyclic graph, all real-time offloading tasks in the cluster of the directed acyclic graph are sorted in descending order according to the parameters on which the tasks depend;

[0190] Step S6.5: In the sorted real-time offloading tasks, the PID control technology is used to iteratively search for the optimal offloading solution to obtain the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph.

[0191] The method of using the PSOSR method to convert a cluster of a communication-intensive directed acyclic graph into a cluster of a computation-intensive directed acyclic graph includes:

[0192] Step S6.1.1: taking all real-time offloading tasks in the cluster of the communication-intensive directed acyclic graph as particles in the particle swarm algorithm;

[0193] Step S6.1.2: Divide the position vector of each particle into a segment encoding continuous variables and a segment encoding discrete variables;

[0194] Step S6.1.3: using the first replication scheme to create a first offspring position vector connected to the encoding continuous variable segment and the discrete variable segment, using the second replication scheme to create a second offspring position vector connected to the encoding continuous variable segment and the discrete variable segment, connecting the first offspring position vector with the second offspring position vector to obtain a complete offspring particle;

[0195] Step S6.1.3: According to step S6.1.2, multiple complete offspring particles are obtained to form initial feasible particles, and the fitness of each initial feasible particle is calculated, where the fitness is the difference between the CCR value of the particle and a predefined first CCR threshold;

[0196] Step S6.1.4: using particle swarm algorithm to iterate particles;

[0197] Step S6.1.5: Determine whether the termination condition is met after iteration;

[0198] Step S6.1.6: If the termination condition is not met, all particles are sorted in descending order according to their fitness. For each particle after sorting, a binary variable is randomly selected to relax the value range of the particle's fitness. Steps S6.1.2 to S6.1.3 are re-executed to regenerate complete offspring particles. Based on multiple complete offspring particles, feasible particles are re-formed.

[0199] Step S6.1.7: Apply feasible particles as real-time offloading tasks to the cluster of the directed acyclic graph;

[0200] Step S6.1.8: Determine whether the CCR value of the cluster of the directed acyclic graph is lower than a predefined second CCR threshold;

[0201] Step S6.1.9: If the CCR value of the cluster of the directed acyclic graph is lower than a predefined second CCR threshold, mark the current cluster of the directed acyclic graph as true;

[0202] Step S6.1.10: Select the particle with the largest fitness in the cluster of the directed acyclic graph marked as true as the optimal solution, and take the optimal solution as the real-time offloading task in the cluster of the converted dense directed acyclic graph;

[0203] Step S6.1.11: If the CCR value of the cluster of the directed acyclic graph is higher than or equal to the predefined second CCR threshold, part of the load is diverted to other clusters of the directed acyclic graph. By diverting part of the load to other resources or systems, the pressure of the current system is reduced, thereby reducing the CCR value of the current cluster of the directed acyclic graph.

[0204] In this embodiment, according to the clustering result obtained in step S4, a swarm intelligence task offloading solution for communication-intensive DAG clusters is designed. A new DAG type conversion method PSOSR (Particle Swarm Optimization Segment Routing) is designed to convert communication-intensive DAG clusters into computation-intensive DAG clusters.

[0205] let Represents a communication-intensive DAG cluster In addition, is a set of binary merge points where only when task τ l When the variable X is selected as the merging point l Set to 1.

[0206] The PSOSR method relaxes l merging points, taking any real number from [0,1]. Then, a PSO variant is adopted to solve the mixed variable task merging problem. Unlike other PSO variants, this latest PSO technique features three stages: mixed variable encoding, mixed offspring reproduction, and adaptive parameter adjustment. In the first stage, the particle position vector (i.e., the transformation solution) is divided into two different segments, encoding continuous variables and discrete variables respectively. Therefore, different evolution operators can be used to evolve continuous variables and discrete variables separately. In the second stage, two replication schemes are designed to create a part of the offspring position vector connected to continuous and discrete variables in parallel. Then, a complete offspring particle can be easily constructed by putting the two position vector segments together. In order to accurately determine the merging point, the continuous variable is rounded to 0 or 1 by comparing with a predefined threshold in combination with the SR technique. After the above stages, the third stage focuses on optimizing the key evolution parameters for the next iteration.

[0207] First, generate the initial feasible particles X = {S1, S2, ..., S J}. Each initial feasible particle is calculated according to the fitness, which is calculated as the difference between the particle's CCR value and the predefined CCR threshold. Enter the iterative process of finding a task offloading solution. If the termination condition is not met, all particles are sorted in descending order. For each particle S j , randomly select a binary variable l j The value range is relaxed. j Divided into S j,1 and S j,2 There are two parts, storing continuous variables and discrete variables respectively. Then, by calling the function offesprin continuous variables For segment S j,1 Construct a sub-generation (S j,1 ) descendants. Similarly, by calling the function's descendant discrete variables For segment S j,1 The discrete variables in the generation of a child (S j,2 ) variables. Therefore, by combining and Generate a complete offspring Using the SR function Input Segment j,1When the task is combined into a solution When the task merging sub-item is applied to the task set Γ, it checks whether the current CCR value of the task set Γ is lower than the predefined CCR threshold. If so, the flag Δ j Set to true. By using the comparison function To compare and S j ..if Better than S j , then use Instead of S j , and adjust the particle S j After checking all particles, we select the particle S with the best fitness from the swimming X by using the selection function Select(X) opt Since the DAG type conversion is now complete, call the step 4 algorithm to set the task set Generate uninstall solution When the termination condition is met, the solution is output.

[0208] Experimental results:

[0209] In this embodiment, a total of 15 containers are installed on the edge server, and the operating frequency of a container is randomly selected from [2000, 4500] MHz. According to the computing power of each processor and container, it is assumed that their average fault arrival rate is [4×10 -6 ,7×10 -5 ]MHz. The bit error rate is from [2×10 -6 ,5×10 -5 ] The bit error rate of one communication link selected in the example above is the bit error rate of one communication link selected in the example above. The D2D transmission and reception powers of a single-end device are [200,1000] and [100,800] mW, respectively. The D2D communication bandwidth between any two end devices varies from 20 to 100 MHz. The transmission power of any end device in remote execution mode is selected from [600,1500] mW, and its communication bandwidth to the target container is selected from [100,400] MHz. The TGFF tool is used to generate different synthetic DAG applications, whose number of tasks varies in [20,300]. Therefore, it is easy to generate a set of heterogeneous DAG applications with different CCR values. The reliability target of each DAG application is randomly selected from [0.70,0.9999]. The proportional coefficient, integral coefficient, and derivative coefficient of the PID controller are set to 0.5, 0.005, and 0.1, respectively. .

[0210] A total of 100 experiments were conducted to obtain average evaluation data. As shown in Table 1, the energy consumption achieved by the method of this embodiment when running the synthetic DAG application is shown. It is observed that the method of this embodiment significantly reduces the overall energy consumption of the terminal device, especially when more terminal devices participate in the task offloading process.

[0211] Table 1. Energy consumption achieved when running synthetic DAG applications

[0212] Number of terminal devices Energy consumption 100 1853.57 150 2430.20 200 3303.82 250 4450.59 300 5789.71 350 7762.89 400 10171.62

[0213] like Figure 3 As shown in the figure, the distribution of energy consumption in the computing and communication processes of five DAG applications (cluster quantization, cluster detection, cluster computing, task offloading, and cluster identification) is shown. The horizontal axis represents energy consumption (in joules), and the vertical axis lists the names of the applications. Each bar in the chart represents the energy consumption of each application in computing and communication, marked with different proportions. From the figure, we can see that, for example, the computing energy consumption of cluster quantization accounts for 78%, and communication accounts for only 22%; while the energy consumption of cluster detection in communication accounts for 71%, and computing accounts for only 29%. This difference in energy distribution helps us understand the differences in the efficiency of energy use of different applications in computing and communication, thereby providing a reference for optimizing the energy efficiency of these applications.

[0214] A clustering-based directed acyclic graph cluster task offloading method proposed in this embodiment first designs a clustering method to quantify different energy requirements by analyzing the energy consumption composition of heterogeneous applications. Then, using incremental control and swarm intelligence-based technologies, a task offloading solution with energy-saving and reliable guarantees is obtained for different application clusters. Secondly, the dependencies between tasks described by a directed acyclic graph (DAG) and the concurrent task offloading of multiple DAG applications are considered. By analyzing the energy dissipation composition of heterogeneous DAG applications, a clustering method is designed to quantify differentiated DAG energy requirements. By analyzing the optimality of computational energy, a computationally intensive DAG cluster task offloading solution based on incremental control is designed. Combined with a new DAG type conversion technology, a communication-intensive DAG cluster task offloading solution based on swarm intelligence is proposed. The method of this embodiment can well handle the heterogeneity of terminal devices during task offloading and ensure the reliability of large-scale information-energy system optimization.

[0215] Embodiment 2:

[0216] This embodiment proposes a directed acyclic graph cluster task offloading system based on clustering. Figure 4As shown, it includes: a model building module, a problem optimization module, a value quantization module, a clustering grouping module, an incremental control module and an intelligent unloading module, wherein the model building module is connected to the problem optimization module, the value quantization module is connected to the clustering grouping module, the problem optimization module and the clustering grouping module are respectively connected to the incremental control module, and the incremental control module is connected to the intelligent unloading module;

[0217] A model building module, used to build a basic model, wherein the basic model includes: a hardware architecture, an application model, a reliability model, and an energy model;

[0218] A problem optimization module, used for constructing a problem of minimizing the energy consumption of a terminal device according to the basic model;

[0219] A value quantification module, used for quantifying energy requirements of different directed acyclic graphs using CCR values, wherein the directed acyclic graphs represent the degree of dependency between real-time offloading tasks of different applications;

[0220] A clustering grouping module is used to cluster and group the clusters of the directed acyclic graph based on the CCR value by using an iterative self-organizing data analysis technology algorithm to obtain clustering results, wherein the clustering results include: clusters of computationally intensive directed acyclic graphs and clusters of communication-intensive directed acyclic graphs, wherein the clusters of the directed acyclic graphs are multiple different directed acyclic graphs, and the directed acyclic graphs include multiple different applications;

[0221] The incremental control module is used for obtaining the computing offloading destination of each real-time offloading task in the cluster of the directed acyclic graph according to the optimal solution of energy consumption of the energy consumption minimization problem, and adopting the incremental control scheme for the cluster of the directed acyclic graph, wherein the destination includes: a local terminal device, a non-local terminal device, and a virtualized container in the edge server S;

[0222] The intelligent offloading module is used for clusters of communication-intensive directed acyclic graphs. According to the energy-optimal solution of the energy-minimization problem, a group intelligent task offloading scheme is adopted to obtain the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph.

[0223] The various embodiments in the present application are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0224] The protection scope of the present application is not limited to the above-mentioned embodiments. Obviously, those skilled in the art can make various changes and modifications to the present disclosure without departing from the scope and spirit of the present disclosure. If these changes and modifications fall within the scope of the claims of the present disclosure and their equivalents, the intention of the present disclosure also includes these changes and modifications.

Claims

1. A directed acyclic graph cluster task offloading method based on clustering, characterized in that: include: Establishing a basic model, wherein the basic model includes: hardware architecture, application model, reliability model and energy model; According to the basic model, a problem of minimizing the energy consumption of terminal equipment is constructed; The CCR value is used to quantify the energy requirements of different directed acyclic graphs, wherein the directed acyclic graphs represent the dependency between the real-time offload tasks of different applications; Based on the CCR value, an iterative self-organizing data analysis technology algorithm is used to cluster and group the clusters of the directed acyclic graph to obtain a clustering result, wherein the clustering result includes: a cluster of a computationally intensive directed acyclic graph and a cluster of a communication-intensive directed acyclic graph, wherein the cluster of the directed acyclic graph is a plurality of different directed acyclic graphs, and the directed acyclic graph includes a plurality of different applications; For a cluster of computationally intensive directed acyclic graphs, an incremental control scheme is adopted according to the energy consumption optimal solution of the energy consumption minimization problem to obtain the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph, wherein the destination includes: a local terminal device, a non-local terminal device, and a virtualized container in an edge server S; For the cluster of communication-intensive directed acyclic graph, according to the energy-optimal solution of the energy-minimization problem, a swarm intelligence task offloading scheme is adopted to obtain the computational offloading destination of each real-time offloading task in the cluster of directed acyclic graph.

2. The clustering-based directed acyclic graph cluster task offloading method according to claim 1, characterized in that: The hardware architecture is an end-to-end collaborative information-physical system architecture consisting of M terminal devices, a base station B, and an edge server S. The edge server S provides multiple virtualized containers, and different virtualized containers have different computing capabilities.

3. The clustering-based directed acyclic graph cluster task offloading method according to claim 2, characterized in that: The application model includes: associating each terminal device in the information-physical system architecture in a directed acyclic graph, the directed acyclic graph includes: a set of real-time offloading tasks with constrained priorities in the application and a communication data set from one real-time offloading task to another real-time offloading task, the execution modes of the real-time offloading tasks include: local mode, D2D mode and remote execution mode, the local mode is a mode in which the real-time offloading task is completed on the local terminal device without the help of an edge server S or other terminal devices, the D2D mode is a mode in which the calculation of the real-time offloading task is offloaded to another terminal device through a D2D communication link and executed on the other terminal device, and the remote execution mode is a mode in which the real-time offloading task is sent to the edge server S and the real-time offloading task is executed on the edge server S.

4. The clustering-based directed acyclic graph cluster task offloading method according to claim 3, characterized in that: The reliability model is the reliability of the real-time offloading task, and the calculation formula is as follows: in, To offload the task τ in real time m,i The reliability of To offload the task τ in real time m,i For the reliability of communication in local mode, To offload the task τ in real time m,i The constant bit error rate of the communication link, To offload the task τ in real time m,i The ability to tolerate soft errors at time t, To offload the task τ in real time m,i Constant soft error rate, The correct communication data of all direct nodes is successfully delivered to the real-time offload task τ m,i probability.

5. The cluster-based directed acyclic graph cluster task offloading method according to claim 4, characterized in that: The energy model includes: static energy consumption of the terminal device, dynamic energy consumption of the terminal device when performing real-time offloading tasks in local mode, and D2D mode of the terminal device when performing real-time offloading tasks. m The communication energy of the D2D mode, the energy dissipation of delivering the real-time offloading task in the D2D mode, and the energy dissipation of the other terminal device D when processing the real-time offloading task in the D2D mode. π The energy consumption of the remote execution mode and the energy consumption of sending real-time offload tasks.

6. The clustering-based directed acyclic graph cluster task offloading method according to claim 5, characterized in that: According to the basic model, the energy consumption minimization problem of the terminal device is constructed, and the calculation formula is as follows: The objective function of the problem of minimizing the energy consumption of the terminal device is: in, Energy for terminal device D m energy consumption, is the total energy consumption, M is the total number of clusters in the directed acyclic graph, and m is the cluster of the mth directed acyclic graph; Among them, the terminal device D m The energy consumption is calculated as follows: in, in, For terminal device D m The static energy consumption, Terminal device D in local mode m Execute real-time offloading taskτ m,i The dynamic energy consumption when When performing real-time offloading tasks in D2D mode, the terminal device D m The communication energy When processing real-time offloading tasks in D2D mode, another terminal device D π energy consumption, Delivering real-time offloading tasks in D2D mode m,i The energy dissipation Send real-time offload tasks for remote execution modeτ m,i Energy consumption, τ n,j G is a cluster of a directed acyclic graph n The jth real-time offloading task in (1≤n≤M,1≤j≤ξ n ), M is the total number of clusters in the directed acyclic graph, ξ n is a cluster G of a directed acyclic graph n The total number of real-time offload tasks in τ m,p is a cluster G of a directed acyclic graph m The pth real-time offloading task in m G is a cluster of a directed acyclic graph m The total number of real-time offload tasks in , i is the cluster G of the i-th directed acyclic graph m In the real-time offloading task, h is the h-th container, H is the total number of containers on the edge server, and α m,i is the first and first variable, β n,j,m is the 201st variable, γ m,i,h is a zero-one variable, δ i,j is a zero-one variable; The constraints of the problem of minimizing the energy consumption of the terminal device are as follows: (1) Each real-time offloading task meets the time constraint, and the calculation formula is as follows: Among them, To offload the task τ in real time m,i The start time, W m,i To offload the task τ in real time m,i The number of CPU instruction cycles at terminal m, f m For terminal device D m The operating frequency, β m,i,π To offload the task τ in real time m,i In terminal device D m The relationship factor with the cluster π of the directed acyclic graph, π is the cluster type of the directed acyclic graph, T m For terminal device D m The deadline for executing the task; (2) Each real-time offloading task can only execute one execution mode, and the calculation formula is as follows: (3) Each real-time offloading task satisfies the reliability constraint, which is calculated as follows: in, For the real-time offloading task τ in the reliability model m,i The reliability, R m is the reliability target of the m-th DAG cluster; (4) The energy consumption of each terminal device has an upper limit, which can be calculated as follows: in, For terminal device D m Energy consumption threshold; (5) The real-time offloading tasks should satisfy the priority constraint, which is calculated as follows: in, To offload the task τ in real time m,i The completion time, To offload the task τ in real time m,i The starting time, e m,i,p Real-time offloading taskτ m,i In terminal device D m communication data; (6) All real-time offloading tasks are executed within their respective durations without overlap. The calculation formula is as follows: Among them, ν n is the voltage at point n, ν m is the power supply voltage, Z is a constant, τ n,j G is a cluster of a directed acyclic graph n The jth real-time offloading task in m,i G is a cluster of a directed acyclic graph m The i-th real-time offloading task in To offload the task τ in real time n,j The start time of τ n,j ∈ν n , and τ m,i ≠τ n,j , we get the following inequality: same, τ n,j ∈ν n ,τ m,i ≠τ n,j ,π∈[1,M],h∈[1,H], we get the following inequality: in, To offload the task τ in real time n,j The completion time, To offload the task τ in real time n,j The starting time, α n,j It is the 501st variable.

7. The clustering-based directed acyclic graph cluster task offloading method according to claim 1, characterized in that: The method for obtaining a computing offloading destination of each real-time offloading task in the cluster of the directed acyclic graph of computation-intensive type adopts an incremental control scheme according to the optimal solution of energy consumption of the energy consumption minimization problem, including: For the cluster of computationally intensive directed acyclic graphs, calculate the optimal solution to the energy consumption minimization problem of the cluster of directed acyclic graphs; Expressing the optimal solution of energy consumption as a function of the dependency relationship between processor characteristics and tasks; Under the premise of maintaining the topological order of the directed acyclic graph, all real-time offload tasks in the cluster of the directed acyclic graph are sorted in descending order according to the processor characteristics and the dependencies between the tasks; Among the sorted real-time offloading tasks, PID control technology is used to iteratively search for the optimal offloading solution, and the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph is obtained.

8. The cluster-based directed acyclic graph cluster task offloading method according to claim 1, characterized in that: The cluster of the communication-intensive directed acyclic graph adopts a group intelligent task offloading scheme according to the energy consumption optimal solution of the energy consumption minimization problem to obtain the calculation offloading destination of each real-time offloading task in the cluster of the directed acyclic graph, including: The PSOSR method is used to transform the cluster of communication-intensive directed acyclic graph into the cluster of computation-intensive directed acyclic graph; For the cluster of computationally intensive directed acyclic graphs, calculate the optimal solution to the energy consumption minimization problem of the cluster of directed acyclic graphs; The optimal solution for energy consumption is expressed as a product of a processor-dependent parameter and a task-dependent parameter; Under the premise of maintaining the topological order of the directed acyclic graph, all real-time offloading tasks in the cluster of the directed acyclic graph are sorted in descending order according to the parameters on which the tasks depend; Among the sorted real-time offloading tasks, PID control technology is used to iteratively search for the optimal offloading solution, and the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph is obtained.

9. The clustering-based directed acyclic graph cluster task offloading method according to claim 8, characterized in that: The method of using the PSOSR method to convert a cluster of a communication-intensive directed acyclic graph into a cluster of a computation-intensive directed acyclic graph includes: Step S6.1.1: taking all real-time offloading tasks in the cluster of the communication-intensive directed acyclic graph as particles in the particle swarm algorithm; Step S6.1.2: Divide the position vector of each particle into a segment encoding continuous variables and a segment encoding discrete variables; Step S6.1.3: using the first replication scheme to create a first offspring position vector connected to the encoding continuous variable segment and the discrete variable segment, using the second replication scheme to create a second offspring position vector connected to the encoding continuous variable segment and the discrete variable segment, connecting the first offspring position vector with the second offspring position vector to obtain a complete offspring particle; Step S6.1.3: According to step S6.1.2, multiple complete offspring particles are obtained to form initial feasible particles, and the fitness of each initial feasible particle is calculated, where the fitness is the difference between the CCR value of the particle and a predefined first CCR threshold; Step S6.1.4: using particle swarm algorithm to iterate particles; Step S6.1.5: Determine whether the termination condition is met after iteration; Step S6.1.6: If the termination condition is not met, all particles are sorted in descending order according to their fitness. For each particle after sorting, a binary variable is randomly selected to relax the value range of the particle's fitness. Steps S6.1.2 to S6.1.3 are re-executed to regenerate complete offspring particles. Based on multiple complete offspring particles, feasible particles are re-formed. Step S6.1.7: Apply feasible particles as real-time offloading tasks to the cluster of the directed acyclic graph; Step S6.1.8: Determine whether the CCR value of the cluster of the directed acyclic graph is lower than a predefined second CCR threshold; Step S6.1.9: If the CCR value of the cluster of the directed acyclic graph is lower than a predefined second CCR threshold, mark the current cluster of the directed acyclic graph as true; Step S6.1.10: Select the particle with the largest fitness in the cluster of the directed acyclic graph marked as true as the optimal solution, and take the optimal solution as the real-time offloading task in the cluster of the converted dense directed acyclic graph; Step S6.1.11: If the CCR value of a cluster in the directed acyclic graph is higher than or equal to a predefined second CCR threshold, part of the load is diverted to other clusters in the directed acyclic graph.

10. A directed acyclic graph cluster task offloading system based on clustering, characterized in that: include: A model building module, a problem optimization module, a value quantization module, a clustering grouping module, an increment control module and an intelligent unloading module, wherein the model building module is connected to the problem optimization module, the value quantization module is connected to the clustering grouping module, the problem optimization module and the clustering grouping module are respectively connected to the increment control module, and the increment control module is connected to the intelligent unloading module; A model building module, used to build a basic model, wherein the basic model includes: a hardware architecture, an application model, a reliability model, and an energy model; A problem optimization module, used for constructing a problem of minimizing the energy consumption of a terminal device according to the basic model; A value quantification module, used for quantifying energy requirements of different directed acyclic graphs using CCR values, wherein the directed acyclic graphs represent the degree of dependency between real-time offloading tasks of different applications; A clustering grouping module is used to cluster and group the clusters of the directed acyclic graph based on the CCR value by using an iterative self-organizing data analysis technology algorithm to obtain clustering results, wherein the clustering results include: clusters of computationally intensive directed acyclic graphs and clusters of communication-intensive directed acyclic graphs, wherein the clusters of the directed acyclic graphs are multiple different directed acyclic graphs, and the directed acyclic graphs include multiple different applications; The incremental control module is used for obtaining the computing offloading destination of each real-time offloading task in the cluster of the directed acyclic graph according to the optimal solution of energy consumption of the energy consumption minimization problem, and adopting the incremental control scheme for the cluster of the directed acyclic graph, wherein the destination includes: a local terminal device, a non-local terminal device, and a virtualized container in the edge server S; The intelligent offloading module is used for clusters of communication-intensive directed acyclic graphs. According to the optimal solution to the energy consumption minimization problem, a group intelligent task offloading scheme is adopted to obtain the computational offloading destination of each real-time offloading task in the cluster of the directed acyclic graph.

Images