MEC task offloading strategy method based on task division and D2D

Abstract

The present invention provides a MEC task offloading strategy method based on task division and D2D, which includes: solving the average delay and average battery power of offloaded subtasks and local intrinsic subtasks calculated in the local server of the user device, solving the average delay and average battery power of the offloaded subtask in the output port calculation and transmission process of the user device, all D2D devices, edge cloud and remote cloud, and the average evaluation in the edge cloud and remote cloud, and then obtaining the total average delay, total average battery power and total average evaluation of the task in the server group, and finding a set of optimal Pareto frontier points of delay, battery power and evaluation by improving the fast elite non-dominated sorting genetic algorithm, and obtaining an intelligent optimization scheme for task allocation probability. Aiming at the problems of limited local storage and computing power of user devices, the present invention considers computing offloading technology, realizes a cloud-edge-end collaborative offloading strategy method, reduces task response time, and improves user experience.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-access edge computing and task offloading, and in particular to a MEC task offloading strategy method based on task partitioning and D2D. Background Art

[0002] With the rapid development of the Internet of Things, wireless communications, and big data technologies, user devices and data traffic have exploded, and a large number of computationally intensive and delay-sensitive applications such as healthcare, virtual reality, and autonomous driving have emerged. The computing resources and storage capacity of user devices are limited and cannot meet the growing demand for data computing. The emergence of cloud computing has alleviated this problem. However, traditional cloud computing transmits massive data tasks to cloud servers for processing, which may bring a heavy link burden to the core network. At the same time, since the deployment location of cloud servers is far away from terminal devices, problems such as response delays, energy consumption, and data security may occur during the transmission process. Although the computing power of cloud servers is very powerful and can provide sufficient computing and storage resources for Internet of Things applications, with the continuous development of user devices and network applications, traditional cloud computing has gradually failed to meet users' requirements for low latency and high service quality. In addition, with the development of 5G, data transmission between terminal devices and remote clouds may cause backhaul network paralysis.

[0003] In recent years, in order to solve the problems of high latency and high battery power in traditional cloud computing, a paradigm shift has occurred in mobile computing, from centralized cloud computing to edge computing. Multi-access edge computing (MEC) is a new network architecture model. Its main feature is to sink computing, storage, processing and other functions from the centralized cloud platform to the edge of the network, providing real-time services to users anytime and anywhere, reducing network latency, and thus improving the user's service experience. Unlike traditional cloud computing, MEC offloads the data computing tasks of terminal devices to edge servers that are closer to the terminal devices, providing users with computing and storage capabilities nearby, reducing congestion in the core network, and thus improving service quality and efficiency. Servers distributed on the edge of the network (also called edge nodes) can reduce interactions with centralized cloud data centers and significantly reduce queuing response time in data exchange. Since edge servers have a certain amount of storage space and computing power and are closer to terminal devices, it is possible to consider offloading computing-intensive or latency-sensitive computing tasks to edge servers for computing. Therefore, as a research hotspot in edge computing, computing offloading has always attracted much attention. A large number of studies have shown that computing offloading technology can well solve the contradiction between resource-constrained terminal devices and computing-intensive applications. In the research on computational offloading strategy methods, most of them are aimed at optimizing latency and battery power, studying reasonable offloading strategy methods, and obtaining the optimal offloading strategy method. Therefore, based on the above research background, it is urgent and necessary to seek a MEC task offloading strategy method based on task division and D2D to achieve multi-objective optimization with task average latency, system average battery power and average evaluation as the goals. Summary of the invention

[0004] In view of the defects in the above-mentioned prior art, the present invention proposes a MEC task offloading strategy method based on task division and D2D. Wherein D2D represents device to device. The method includes solving the average delay and average battery power of the offloaded subtask and the local inherent subtask calculated in the local server of the user device, solving the average delay and average battery power of the offloaded subtask in the output port calculation and transmission process of the user device, all D2D devices, edge cloud and remote cloud, and the average evaluation in the edge cloud and remote cloud, and then obtaining the total average delay, total average battery power and total average evaluation of the task in the server group, and finding a set of optimal Pareto frontier points of delay, battery power and evaluation by improving the fast elite non-dominated sorting genetic algorithm, and obtaining an intelligent optimization scheme for task allocation probability. In view of the problems of limited local storage and computing power of user devices, the present invention considers computing offloading technology, realizes a cloud-edge-end collaborative offloading strategy method, reduces task response time, and improves user experience.

[0005] The present invention provides a MEC task offloading strategy method based on task division and D2D, which includes the following steps:

[0006] S1. In the server cluster, considering the burst arrival mode of tasks, the Poisson process is extended to the Markov modulated Poisson process MMPP. The task division module is used to divide the large number of tasks continuously generated when the user device is working into offload subtasks and local inherent subtasks according to a certain ratio, and the average delay and average battery power of the offload subtasks and local inherent subtasks calculated on the local server of the user device are solved. The server group includes user equipment, base stations, D2D devices, edge clouds and remote clouds;

[0007] S2. Calculate the average delay and average battery power T of the offload subtask during the calculation and transmission process at the output port of the user device. out ,E out ,T tra ,E tra ;

[0008] S3. Calculate the average delay and average battery power T of the offload subtask on all D2D devices d2d ,E d2d ;

[0009] S4. Solve the average delay, average battery power and average evaluation T of the offloaded subtask in the edge cloud edg ,E edg ,C edg ;

[0010] S5. Solve the average delay and average evaluation T of the offloaded subtask in the remote cloud clo ,C clo ;

[0011] S6, solve the total average delay, total average battery power and total average evaluation T,E,C of the task in the server group;

[0012] S61, the average delays of a local inherent subtask and an offloaded subtask in the server cluster are T1 and T2 respectively:

[0013]

[0014]

[0015] Where p represents the offloading probability; q represents the total probability of offloading to n D2D devices; ξ represents the probability of offloading to the edge cloud; T out represents the average delay of the offload subtask calculated at the output port of the user equipment; T tra represents the transmission time of the offload subtask being offloaded from the output port of the user device; represents the average latency of the local intrinsic subtask calculated on the local processor of the user device; represents the average latency of offloading subtasks calculated on the local processor of the user device;

[0016] S62: The task is calculated simultaneously on the user device and multiple offloading terminals including the D2D device, the edge cloud and the remote cloud, and a task is divided into a user device part and all offloading parts. Then the total average delay T of a task in the server group is the maximum of the delays of the user device part and all offloading parts:

[0017] T=max{2(1-1)T1,2lT2} (17)

[0018] Among them, l represents the proportion of uninstalled subtasks;

[0019] S63, the average battery power of a local inherent subtask and an offloaded subtask in the server cluster is E1, E2 respectively:

[0020]

[0021]

[0022] in, represent the average battery power of the local inherent subtask and the offloaded subtask on the local server respectively; E out ,E tra They represent the average battery power of the offload subtask calculated at the output port and offloaded transmitted from the output port, respectively;

[0023] S64, assuming that the average battery power generated by each task being processed in the server group is the total average battery power E, then:

[0024] E = 2(1-1)E1+21E2 (20);

[0025] S65. The total average evaluation of a task in the server group is:

[0026] C=2l(C edg +C clo ) (twenty one);

[0027] S7. Improve the fast elite non-dominated sorting genetic algorithm and use MATLAB software to conduct optimization experiments to find the optimal Pareto frontier point (T * ,E * ,C * ), the total average delay T * , total average battery capacity E * With the overall average rating C* is relatively low and in equilibrium, and obtains the task assignment probability (p * ,q * ,ξ * ) intelligent optimization solution, where p * represents the optimized uninstall probability; q * represents the total probability of optimized offloading to n D2D devices; ξ * Represents the probability of optimized offloading to edge cloud.

[0028] Further, the step S1 specifically includes the following steps:

[0029] S11. Based on queuing theory, the average delay of the local inherent subtask and the offload subtask in the local server of the user device is obtained.

[0030]

[0031]

[0032] Where λ represents the task arrival rate; L represents the capacity of the local buffer queue for offloading subtasks; π i,j,k,g represents the probability distribution in steady state, that is, the probability that in steady state, when the number of local intrinsic subtasks is i, the number of offloaded subtasks is j, the underlying Markov chain of the local intrinsic subtask is in state k, and the underlying Markov chain of the offloaded subtask is in state g;

[0033] S12, assuming that the average power required for local server calculation is P loc , then the average battery power of the local inherent subtask and the offloaded subtask on the local server is for:

[0034]

[0035] Preferably, the step S2 specifically includes the following steps:

[0036] S21. Based on queuing theory, the average delay T calculated by the offloading subtask at the output port of the user device is obtained respectively. out and its transmission time T to be unloaded from the output port of the user equipment tra :

[0037]

[0038]

[0039] Where r represents the uplink data transmission rate; W represents the channel bandwidth; θ represents the average amount of data contained in an offload subtask; represents the signal-to-noise ratio, S represents the signal power, and N represents the noise power; π i,j represents the probability distribution in steady state, that is, the probability that the Markov chain underlying the unloading subtask is in state j when the number of unloading subtasks is i in steady state;

[0040] S22, assuming that the average power required for the output port processing calculation of the user equipment is P out , then the offloading subtask calculates and offloads the average battery power E transmitted from the output port of the user device at the output port of the user device out ,E tra for:

[0041]

[0042]

[0043] Preferably, the step S3 specifically includes the following steps:

[0044] S31, average delay of offloading subtasks in all D2D devices T d2d for:

[0045]

[0046] Where n represents the number of D2D devices; B represents the capacity of the buffer queue of the offloaded subtask in the D2D device; π i,j,k represents the probability distribution in steady state, that is, the probability that the number of local tasks of the D2D device is i, the number of offloaded subtasks is j, and the underlying Markov chain of the offloaded subtask is in state k in steady state;

[0047] S32. Assume that the average power calculated by a D2D device is P d2d , then the average battery power E of all D2D devices for the offloaded subtask d2d for:

[0048]

[0049] Preferably, the step S4 specifically includes the following steps:

[0050] S41, average delay of offloading subtasks in edge cloud T edg for:

[0051]

[0052] Among them, π i,j,k represents the probability distribution in steady state, that is, the probability of whether the edge cloud low-speed service group is in working state k when the number of offloaded subtasks is i and the underlying Markov chain of the offloaded subtask is in state j in steady state;

[0053] S42, assuming that the average power of edge cloud computing is P edg , then the average battery power of the offloaded subtask in the edge cloud is E edg for:

[0054]

[0055] S43, assuming that the unit evaluation of edge cloud resources is C1, then the average evaluation of the offloaded subtask in the edge cloud is C edg for:

[0056] C edg =2lp(1-q)ξλC1 (12).

[0057] Preferably, the step S5 specifically includes the following steps:

[0058] S51, average delay of offloading subtasks in remote cloud T clo for:

[0059]

[0060] Among them, T o represents the fixed time delay of task request transmitted from the load balancer to the remote cloud; μ clo Indicates the cloud server service rate;

[0061] S52, assuming that the unit evaluation of the remote cloud resources is C2, then the average evaluation of the offloaded subtasks in the remote cloud is C clo for:

[0062] C clo =2lp(1-q)(1-ξ)λC2 (14).

[0063] Preferably, in the step S1, the capacity of the buffer queue of the offloading subtask in the local server is L, the capacity of the buffer queue of the local inherent subtask in the local server is unlimited, and the queuing model of the local server is an MMPP+MMPP / M / 1 queue; in the step S2, the queuing cache of the offloading subtask for calculation and processing at the output port of the user equipment and queuing for offloading transmission is unlimited, and the queuing model of the output port of the user equipment is an MMPP / M / 1 queue; in the step S3, the capacity of the buffer queue of the offloading subtask in the D2D device is B, the capacity of the buffer queue of the local task in the D2D device is unlimited, the local task of the D2D device has a strong preemption priority, and the queuing model in the D2D device is an M+MMPP / M / 1 queue; in the step S4, the edge cloud provides a buffer queue of unlimited capacity for the offloaded subtask. In the edge cloud, a limited number of servers are clustered according to the service rate, and are divided into two clusters, a high-speed server cluster and a low-speed server cluster. When the number of subtasks in the edge cloud is less than a certain number, only the high-speed server cluster works. When the number of subtasks is greater than a certain number, both the high-speed server cluster and the low-speed server cluster are in working state. The queuing model in the edge cloud is an MMPP / M / c queue; in the step S5, the remote cloud has enough cloud servers and the computing power is strong enough. The tasks offloaded to the remote cloud do not need to wait and can receive services without delay when they arrive. The queuing model in the remote cloud is an MMPP / M / ∞ queue.

[0064] Preferably, the local inherent subtask is only processed and calculated on the local server of the user device, and the offloading subtask is determined with a certain probability whether it is processed on the local server or queued at the output port of the user device and waits to be offloaded; if the offloading subtask is queued at the output port of the user device, then after leaving the output port, a part of it is evenly offloaded to multiple adjacent D2D devices, and the other part is transmitted to the load balancer in the edge cloud via the base station, and the load balancer decides whether to offload to the edge cloud or the remote cloud based on the load balancing situation.

[0065] Preferably, the improved fast elite non-dominated sorting genetic algorithm in step S9 uses Tent mapping to initialize the population, and dynamically changes the crossover and mutation operators in the genetics to achieve adaptive evolution of the population.

[0066] Compared with the prior art, the technical effects of the present invention are:

[0067] 1. The present invention designs a MEC task offloading strategy method based on task division and D2D. Considering the computing offloading technology in edge computing, the local tasks are offloaded to the adjacent D2D devices, edge cloud or remote cloud, to realize the cloud-edge-end collaborative offloading strategy method, which effectively solves the problems of limited local storage and computing power of user devices. The partial offloading technology considering task division is designed to reduce the task response time and improve the user experience.

[0068] 2. The present invention designs a MEC task offloading strategy method based on task division and D2D. It uses queuing theory knowledge to establish different queuing theory models locally and at multiple offloading ends respectively, and uses matrix geometry solution to perform steady-state analysis on the model to obtain performance indicators such as average task delay, average system battery power and average evaluation. It considers the multi-objective optimization problem of the system, improves the fast elite non-dominated sorting genetic algorithm, and obtains the optimal Pareto frontier point and optimization scheme for task allocation probability. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Other features, objects and advantages of the present application will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings.

[0070] Figure 1 It is a flow chart of the MEC task offloading strategy method based on task division and D2D of the present invention;

[0071] Figure 2 It is a schematic diagram of the composition of a server group of the MEC task offloading strategy method based on task division and D2D of the present invention;

[0072] Figure 3 is a graph showing a changing trend of the average task delay in a specific embodiment of the present invention;

[0073] Figure 4 is a graph showing a change trend of the average battery power of the system in a specific embodiment of the present invention;

[0074] Figure 5 is a graph showing a changing trend of the average evaluation of the system in a specific embodiment of the present invention;

[0075] Figure 6 is a Pareto front point diagram in a specific embodiment of the present invention;

[0076] Figure 7 It is a fitness function convergence curve in a specific embodiment of the present invention.

[0077] In the figure:

[0078] 1-User equipment; 2-Base station; 3-D2D device; 4-Edge cloud; 5-Remote cloud. DETAILED DESCRIPTION

[0079] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0080] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0081] Figure 1 The present invention shows a MEC task offloading strategy method based on task division and D2D, which includes the following steps:

[0082] S1. In the server cluster, considering the burst arrival mode of tasks, the Poisson process is extended to the Markov modulated Poisson process MMPP. Using the task division module, a large number of tasks continuously generated when the user device 1 is working are divided into offload subtasks and local inherent subtasks according to a certain ratio. The average delay and average battery power of the offload subtasks and local inherent subtasks calculated on the local server of the user device 1 are solved.

[0083] like Figure 2 As shown, the server group includes user equipment 1, base station 2, D2D device 3, edge cloud 4 and remote cloud 5. The local inherent subtask is only processed and calculated in the local server of user equipment 1. In the offload subtask, it is determined with a certain probability whether it is processed in the local server or queued at the output port of user equipment 1 and waits to be offloaded; if the offload subtask is queued at the output port of user equipment 1, then after leaving the output port, part of it is evenly offloaded to multiple adjacent D2D devices 3, and the other part is transmitted to the load balancer in the edge cloud 4 through the base station 2, and the load balancer decides to offload to the edge cloud 4 or the remote cloud 5 according to the load balancing situation.

[0084] The capacity of the buffer queue of the offload subtask on the local server is L, and the capacity of the buffer queue of the local inherent subtask on the local server is unlimited. Because the local inherent subtask can only be executed on the local server, it is given a strong preemptive priority. It can be seen that the longer the waiting time of the offload subtask on the local server, the more impatient it will become and leave the system early. The local server queue model is a MMPP+MMPP / M / 1 queue.

[0085] S11, based on queuing theory, the average delay of the local inherent subtask and the offload subtask in the local server of user device 1 is obtained

[0086]

[0087]

[0088] Where λ represents the task arrival rate; L represents the capacity of the local buffer queue for offloading subtasks; π i,j,k,g It represents the probability distribution in steady state, that is, the probability that when the number of local intrinsic subtasks is i, the number of offloaded subtasks is j, the underlying Markov chain of the local intrinsic subtasks is in state k, and the underlying Markov chain of the offloaded subtasks is in state g.

[0089] S12, assuming that the average power required for local server calculation is P loc , then the average battery power of the local inherent subtask and the offloaded subtask on the local server is for:

[0090]

[0091] S2. Calculate the average delay and average battery power of the offloading subtask during calculation and transmission at the output port of user device 1. The queue cache of the offloading subtask for calculation, processing and queuing for offloading transmission at the output port of user device 1 is infinite, and the queue model of the output port of user device 1 is an MMPP / M / 1 queue.

[0092] S21. Based on queuing theory, the average delay (delay for data processing and queuing at the output port) T of the offload subtask calculated at the output port of the user device 1 is obtained respectively. out and its transmission time T unloaded from the output port of user equipment 1 tra :

[0093]

[0094]

[0095] Where r represents the uplink data transmission rate; W represents the channel bandwidth; θ represents the average amount of data contained in an offload subtask; represents the signal-to-noise ratio, S represents the signal power, and N represents the noise power; π i,j It represents the probability distribution in steady state, that is, the probability that the Markov chain under the unloading subtask is in state j when the number of unloading subtasks is i in steady state.

[0096] S22, assuming that the average power required for the output port processing calculation of user equipment 1 is P out , then the offloading subtask calculates and offloads the average battery power E transmitted from the output port of user device 1 at the output port of user device 1 out ,Etra for:

[0097]

[0098]

[0099] S3, solve the average delay and average battery power T of the offloading subtask in all D2D devices 3 d2d ,E d2d . Since D2D device 3 has fewer local tasks and most of them are in idle state, there are a lot of idle computing resources for the unloading subtask service of the user equipment. The capacity of the buffer queue of the unloading subtask in D2D device 3 is B, and the capacity of the buffer queue of the local task in D2D device 3 is unlimited. Among them, the local task of D2D device 3 has a strong preemptive priority. The queuing model in D2D device 3 is an M+MMPP / M / 1 queue.

[0100] S31, average delay T of offloading subtasks in all D2D devices 3 d2d for:

[0101]

[0102] Where n represents the number of D2D devices; B represents the capacity of the buffer queue of the offloaded subtask in the D2D device; π i,j,k It represents the probability distribution in steady state, that is, the probability that when the number of local tasks of the D2D device is i, the number of offloaded subtasks is j, and the underlying Markov chain of the offloaded subtask is in state k in steady state.

[0103] S32: Since D2D device 3 has few local tasks, its battery power is not considered. Assume that the average power calculated by D2D device 3 is P d2d , then the average battery power E of all D2D devices 3 in the offload subtask d2d for:

[0104]

[0105] S4. Solve the average delay, average battery power and average evaluation T of the offloaded subtask in the edge cloud edg ,E edg ,C edg .

[0106] Edge Cloud 4 provides a buffer queue with unlimited capacity for offloading subtasks. In Edge Cloud 4, a limited number of servers are clustered according to the service rate into two clusters: a high-speed server cluster and a low-speed server cluster. All tasks in Edge Cloud 4 are waiting for service in a waiting queue. As the number of tasks in Edge Cloud 4 changes, different tasks may receive different services in the two clusters, which is beneficial to saving the battery power of the edge server. When the number of subtasks in Edge Cloud 4 is less than a certain number, only the high-speed server cluster works. When the number of subtasks is greater than a certain number, both the high-speed server cluster and the low-speed server cluster are in working state. The queuing model in Edge Cloud 4 is an MMPP / M / c queue.

[0107] S41, average delay T of offloading subtasks in edge cloud 4 edg for:

[0108]

[0109] Among them, π i,j,k It represents the probability distribution in steady state, that is, the probability of whether the edge cloud low-speed service group is in working state k when the number of offloaded subtasks is i and the underlying Markov chain of the offloaded subtask is in state j in steady state.

[0110] S42, assuming that the average power calculated by edge cloud 4 is P edg , then the average battery power E of the offloaded subtask in edge cloud 4 edg for:

[0111]

[0112] S43: User device 1 must pay for using the resources of edge cloud 4. Assuming that the unit evaluation of the resources used by edge cloud 4 is C1, the average evaluation of the offloaded subtask in edge cloud 4 is C edg for:

[0113] C edg =2lp(1-q)ξλC1 (12).

[0114] S5. Calculate the average delay and average evaluation T of the offloaded subtask in remote cloud 4 clo ,C clo Compared with edge cloud 4, remote cloud 5 has enough cloud servers and strong computing power in terms of geographical location. Tasks offloaded to remote cloud 5 do not need to wait and can receive services without delay when they arrive. The queuing model in remote cloud 5 is a MMPP / M / ∞ queue.

[0115] S51, average delay T of offloading subtasks in remote cloud 5 clo for:

[0116]

[0117] Among them, T o represents the fixed time delay of the task request being transmitted from the load balancer to the remote cloud 5; μ clo Indicates the cloud server service rate.

[0118] Due to its powerful computing power, the battery power of the subtask on the remote cloud 5 is not considered.

[0119] S52: User device 1 must pay for using the resources of remote cloud 5. Assuming that the unit evaluation of the resources used by remote cloud 5 is C2, the average evaluation of the offloaded subtask in remote cloud 5 is C clo for:

[0120] C clo =2lp(1-q)(1-ξ)λC2 (14).

[0121] S6. Calculate the total average latency, total average battery power, and total average evaluation of the task in the server cluster.

[0122] S61, the average delays of a local intrinsic subtask and an offloaded subtask in the server cluster are T1 and T2 respectively:

[0123]

[0124]

[0125] Where p represents the offloading probability; q represents the total probability of offloading to n D2D devices; ξ represents the probability of offloading to the edge cloud; T out represents the average delay of the offloading subtask calculated at the output port of user equipment 1; T tra represents the transmission time of the offloading subtask being offloaded from the output port of the user device 1; represents the average latency of the local intrinsic subtask calculated on the local processor of the user device; represents the average latency of offloading subtasks calculated on the local processor of the user device;

[0126] S62, the task is calculated simultaneously on the user device 1 and multiple offloading terminals including the D2D device 3, the edge cloud 4 and the remote cloud 5, and a task is divided into a user device part and all offloading parts, then the total average delay T of a task in the server group is the maximum of the delays of the user device part and all offloading parts:

[0127] T=max{2(1-1)T1,21T2} (17)

[0128] Among them, l represents the proportion of uninstalled subtasks.

[0129] S63, the average battery power of a local inherent subtask and an offloaded subtask in the server cluster is E1, E2 respectively:

[0130]

[0131]

[0132] in, represent the average battery power of the local inherent subtask and the offloaded subtask on the local server respectively; E out ,E tra They represent the average battery power of the offloaded subtasks calculated at the output port and offloaded transmitted from the output port, respectively.

[0133] S64, assuming that the average battery power generated by each task being processed in the server group is the total average battery power E, then:

[0134] E=2(1-1)E1+21E2 (20).

[0135] S65, the local inherent subtask in the server group is only processed and calculated in the local server of the user device 1, and the offloaded subtask is offloaded to the edge cloud 4 and the remote cloud 5 for processing. Then the total average evaluation of a task in the server group is:

[0136] C=2l(C edg +C clo ) (twenty one).

[0137] S7. Improve the fast elite non-dominated sorting genetic algorithm and use MATLAB software to conduct optimization experiments to find the optimal Pareto frontier point (T) of delay, battery power and evaluation. * ,E * ,C * ), the total average delay T * , total average battery capacity E * With the overall average rating C * is relatively low and in equilibrium, and obtains the task assignment probability (p * ,q * ,ξ * ) intelligent optimization solution, where p * represents the optimized uninstall probability; q * represents the total probability of optimized offloading to n D2D devices; ξ * Represents the probability of optimized offloading to edge cloud. The improved fast elite non-dominated sorting genetic algorithm uses Tent mapping to initialize the population and dynamically changes the crossover and mutation operators in genetics to achieve adaptive evolution of the population.

[0138] In a specific embodiment, for a MEC task offloading strategy method based on task partitioning and D2D, numerical experiments are conducted to further analyze the impact of task offloading probability on different performance indicators under different arrival rates.

[0139] Numerical experiments were carried out using the MATLAB R2018a environment to reveal the changing trends of various performance indicators and verify the rationality of the strategy.

[0140] Considering the system stability conditions, the experimental parameters are set. When conducting system experiments, in order to achieve the purpose of controlling a single variable, the task allocation probability q = ξ = 0.5 is taken as an example to study the impact of the unloading probability p on different performance indicators. The average delay, average battery power and average evaluation of the system tasks under different arrival rates are obtained, respectively, with the change trend of the unloading probability, as shown in the figure below: Figure 3 to Figure 5 shown.

[0141] from Figure 3 It can be observed that when the average arrival rate λ of the task is fixed, as the offloading probability p increases, the average delay T of the system task shows a trend of first decreasing and then increasing. When the probability of task offloading is small, most of the offloaded subtasks are processed on the local server. At this time, the average delay of the local server is the main factor. Due to the large number of tasks on the local server, the local load pressure is large, resulting in a large average delay. As the probability of task offloading gradually increases, the load pressure on the local server is relieved, and the average delay gradually decreases. However, when the probability of task offloading exceeds a certain size, the task delay in the system increases with the increase of the probability of task offloading. This is because when the probability of task offloading is large, the number of tasks to be offloaded is large, and the load pressure processed at the local output port increases, resulting in an increase in the average delay of tasks in the system. At the same time, this trend of decreasing first and then increasing shows that as long as a reasonable task offloading probability is set, the minimum average task delay can be obtained.

[0142] In addition, when the offloading probability p is fixed, the average task delay T increases with the increase of the average task arrival rate. The larger the average task arrival rate, the more tasks arrive, and the load pressure on the processor and server in the system increases accordingly, which increases the average task delay.

[0143] from Figure 4It can be observed that the average battery power E of the system shows a trend of first decreasing and then increasing with the increase of the offloading probability p. The system battery power consists of local server processing, local output port processing, local transmission, D2D device server processing, edge cloud transmission and edge cloud processing. When the probability of task offloading is small, most of the offloaded subtasks are processed on the local server. At this time, the average battery power generated by the local server is the main factor. As the probability of task offloading gradually increases, the load pressure of the local server is relieved, and the average battery power gradually decreases. However, when the probability of task offloading exceeds a certain size, the average battery power in the system increases with the increase of the probability of task offloading. This is because when the probability of task offloading is large, the number of tasks to be offloaded is large, and because the average power of the output port and each offloading server is large, the average battery power of the tasks in the system increases. At the same time, this trend of decreasing first and then increasing shows that as long as a reasonable task offloading probability is set, the minimum system average battery power can be obtained.

[0144] At the same time, the greater the average arrival rate of tasks, the greater the number of tasks arriving, and the correspondingly greater load pressure in the system, thereby increasing the average battery power of the system.

[0145] from Figure 5 It can be observed that the system average evaluation C is linearly related to the offloading probability p, and increases with the increase of the offloading probability. When the offloading probability increases, the number of offloading tasks increases, resulting in an increase in the evaluation values ​​generated in the edge cloud and the remote cloud.

[0146] In order to further obtain the Pareto frontier points and the Pareto optimal solution set, the system optimization experiment was carried out in the MATLAB R2018a environment using the relevant parameters of the numerical experiment to verify the rationality of the optimization method.

[0147] like Figure 6 As shown in Figure 1, the Pareto frontier scatter distribution obtained after iterative optimization is shown. The Pareto optimal solution set that meets the conditions, that is, the optimal task allocation probability, is selected from the figure, as shown in Table 1.

[0148]

[0149] Table 1

[0150] In the Pareto optimal solution set, there is no unique global optimal solution. Due to the contradictions between the various objectives, in practical applications, decision makers should select the optimal solution configuration from the Pareto optimal solution set based on the optimization focus. For example, in real life, based on the preferences of end users and MEC service providers, appropriate optimal solutions can be selected to form the required optimization solution.

[0151] Taking the average task delay as an example, we use it as the fitness function to compare the convergence of the traditional and improved fast elite non-dominated sorting genetic algorithm NSGA II. Figure 7 shown.

[0152] from Figure 7 From the comparison curve, we can see that the convergence of the improved optimization algorithm is better, and the fitness function can get the optimal value when the number of iterations is close to 15. This shows that the proposed improvement method has played a role, accelerating the convergence speed of the algorithm and improving the solution rate.

[0153] The present invention designs a MEC task offloading strategy method based on task division and D2D, which considers the computing offloading technology in edge computing, offloads local tasks to adjacent D2D devices, edge clouds or remote clouds, and realizes a cloud-edge-end collaborative offloading strategy method, which effectively solves the problems of limited local storage and computing power of user devices; designs a partial offloading technology considering task division, reduces task response time, and improves user experience; utilizes queuing theory knowledge to establish different queuing theory models locally and at multiple offloading ends, uses matrix geometry solutions to perform steady-state analysis on the models, and obtains performance indicators such as average task delay, average system battery power, and average evaluation; considers the multi-objective optimization problem of the system, improves the fast elite non-dominated sorting genetic algorithm, and obtains the optimal Pareto frontier point and optimization scheme for task allocation probability.

[0154] Finally, it should be noted that the above embodiments are only intended to illustrate rather than limit the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the present invention can still be modified or replaced by equivalents. Any modification or partial replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A MEC task offloading strategy method based on task division and D2D, characterized in that: It includes the following steps: S1. In the server cluster, considering the burst arrival mode of tasks, the Poisson process is extended to the Markov modulated Poisson process MMPP. The task division module is used to divide the large number of tasks continuously generated when the user device is working into offload subtasks and local inherent subtasks according to a certain ratio, and the average delay and average battery power of the offload subtasks and local inherent subtasks calculated on the local server of the user device are solved. The server group includes user equipment, base stations, D2D devices, edge clouds and remote clouds; S2. Solve the offloading subtask to calculate and offload the average battery power E transmitted at the output port of the user device out ,E tra , the average delay T calculated by the offloading subtask at the output port of the user equipment out and the transmission time T of the offload subtask from the output port of the user device tra ; S3. Calculate the average delay and average battery power T of the offload subtask on all D2D devices d2d ,E d2d ; S4. Solve the average delay, average battery power and average evaluation T of the offloaded subtask in the edge cloud edg ,E edg ,C edg ; S5. Solve the average delay and average evaluation T of the offloaded subtask in the remote cloud clo ,C clo ; S6, solve the total average delay, total average battery power and total average evaluation T,E,C of the task in the server group; S61, the average delays of a local inherent subtask and an offloaded subtask in the server cluster are T1 and T2 respectively: Where p represents the offloading probability; q represents the total probability of offloading to n D2D devices; ξ represents the probability of offloading to the edge cloud; T out represents the average delay of the offload subtask calculated at the output port of the user equipment; T tra represents the transmission time of the offload subtask being offloaded from the output port of the user device; represents the average latency of the local intrinsic subtask calculated on the local server of the user device; represents the average latency of offloading subtasks calculated on the local server of the user device; S62: The task is calculated simultaneously on the user device and multiple offloading terminals including the D2D device, the edge cloud and the remote cloud, and a task is divided into a user device part and all offloading parts. Then the total average delay T of a task in the server group is the maximum of the delays of the user device part and all offloading parts: T=max{2(1-1)T1,21T2} (17) Among them, l represents the proportion of uninstalled subtasks; S63, the average battery power of a local inherent subtask and an offloaded subtask in the server cluster is E1, E2 respectively: in, represent the average battery power of the local inherent subtask and the offloaded subtask on the local server respectively; E out ,E tra They represent the average battery power of offloading subtasks for calculation and offloading transmission at the output port of the user device, respectively; S64, assuming that the average battery power generated by each task being processed in the server group is the total average battery power E, then: E = 2(1-1)E1+21E2 (20); S65. The total average evaluation of a task in the server group is: C=2l(C edg +C clo ) (21); S7. Improve the fast elite non-dominated sorting genetic algorithm and use MATLAB software to conduct optimization experiments to find the optimal Pareto frontier point (T * ,E * ,C * ), the total average delay T * , total average battery capacity E * With the overall average rating C * is relatively low and in equilibrium, and obtains the task assignment probability (p * ,q * ,ξ * ) intelligent optimization solution, where p * represents the optimized uninstall probability; q * represents the total probability of optimized offloading to n D2D devices; ξ * represents the probability of optimized offloading to edge cloud; In the step S1, the capacity of the buffer queue of the offloading subtask in the local server is L, the capacity of the buffer queue of the local inherent subtask in the local server is unlimited, and the queuing model of the local server is an MMPP+MMPP / M / 1 queue; in the step S2, the queuing cache of the offloading subtask for calculation and processing at the output port of the user equipment and queuing for offloading transmission is unlimited, and the queuing model of the output port of the user equipment is an MMPP / M / 1 queue; in the step S3, the capacity of the buffer queue of the offloading subtask in the D2D device is B, the capacity of the buffer queue of the local task in the D2D device is unlimited, the local task of the D2D device has a strong preemption priority, and the queuing model in the D2D device is an M+MMPP / M / 1 Queue; in the step S4, the edge cloud provides a buffer queue of unlimited capacity for the offloaded subtask. In the edge cloud, a limited number of servers are clustered according to the service rate, and are divided into two clusters, a high-speed server cluster and a low-speed server cluster. When the number of subtasks in the edge cloud is less than a certain number, only the high-speed server cluster works. When the number of subtasks is greater than a certain number, both the high-speed server cluster and the low-speed server cluster are in working state. The queuing model in the edge cloud is an MMPP / M / c queue; in the step S5, the remote cloud has enough cloud servers and the computing power is strong enough. The tasks offloaded to the remote cloud do not need to wait, and can receive services without delay when they arrive. The queuing model in the remote cloud is an MMPP / M / ∞ queue.

2. According to claim 1, the MEC task offloading strategy method based on task division and D2D is characterized in that: The step S1 specifically includes the following steps: S11. Based on queuing theory, the average delay of the local inherent subtask and the offload subtask in the local server of the user device is obtained. Where λ represents the task arrival rate; L represents the capacity of the local buffer queue for offloading subtasks; π i,j,k,g represents the probability distribution in steady state, that is, the probability that in steady state, when the number of local intrinsic subtasks is i, the number of offloaded subtasks is j, the underlying Markov chain of the local intrinsic subtask is in state k, and the underlying Markov chain of the offloaded subtask is in state g; S12, assuming that the average power required for local server calculation is P loc , then the average battery power of the local inherent subtask and the offloaded subtask on the local server is for:

3. According to claim 1, the MEC task offloading strategy method based on task division and D2D is characterized in that: The step S2 specifically includes the following steps: S21. Based on queuing theory, the average delay T calculated by the offloading subtask at the output port of the user device is obtained respectively. out and its transmission time T to be unloaded from the output port of the user equipment tra : Where λ represents the task arrival rate, r represents the uplink data transmission rate; W represents the channel bandwidth; θ represents the average amount of data contained in an offloaded subtask; represents the signal-to-noise ratio, S represents the signal power, and N represents the noise power; π i,j represents the probability distribution in steady state, that is, the probability that the Markov chain underlying the unloading subtask is in state j when the number of unloading subtasks is i in steady state; S22, assuming that the average power required for the output port processing calculation of the user equipment is P out , then the offloading subtask calculates and offloads the average battery power E transmitted from the output port of the user device at the output port of the user device out ,E tra for:

4. According to claim 1, the MEC task offloading strategy method based on task division and D2D is characterized in that: The step S3 specifically comprises the following steps: S31, average delay of offloading subtasks in all D2D devices T d2d for: Where λ represents the task arrival rate, n represents the number of D2D devices; B represents the capacity of the buffer queue of the offloaded subtask in the D2D device; π i,j,k represents the probability distribution in steady state, that is, the probability that the number of local tasks of the D2D device is i, the number of offloaded subtasks is j, and the underlying Markov chain of the offloaded subtask is in state k in steady state; S32. Assume that the average power calculated by a D2D device is P d2d , then the average battery power E of all D2D devices for the offloaded subtask d2d for:

5. The MEC task offloading strategy method based on task division and D2D according to claim 1 is characterized in that: The step S4 specifically comprises the following steps: S41, average delay of offloading subtasks in edge cloud T edg for: Among them, λ represents the task arrival rate, π i,j,k represents the probability distribution in steady state, that is, the probability of whether the edge cloud low-speed service group is in working state k when the number of offloaded subtasks is i and the underlying Markov chain of the offloaded subtask is in state j in steady state; S42, assuming that the average power of edge cloud computing is P edg , then the average battery power of the offloaded subtask in the edge cloud is E edg for: S43, assuming that the unit evaluation of edge cloud resources is C1, then the average evaluation of offloading subtasks in the edge cloud is C edg for: C edg =2lp(1-q)ξλC1 (12)。 6. The MEC task offloading strategy method based on task division and D2D according to claim 1 is characterized in that: The step S5 specifically comprises the following steps: S51, average delay of offloading subtasks in remote cloud T clo for: Among them, T o represents the fixed time delay of task request transmitted from the load balancer to the remote cloud; μ clo Indicates the cloud server service rate; S52, assuming that the unit evaluation of the remote cloud resources is C2, then the average evaluation of the offloaded subtasks in the remote cloud is C clo for: C clo =2lp(1-q)(1-ξ)λC2 (14); Among them, λ represents the task arrival rate.

7. The MEC task offloading strategy method based on task division and D2D according to claim 1 is characterized in that: The local inherent subtask is only processed and calculated by the local server of the user device, and the offload subtask is determined with a certain probability to be processed in the local server or queued at the output port of the user device and wait to be offloaded; If the offloading subtask is queued at the output port of the user equipment, then after leaving the output port, a part of it is evenly offloaded to multiple adjacent D2D devices, and the other part is transmitted to the load balancer in the edge cloud through the base station, and the load balancer decides whether to offload to the edge cloud or the remote cloud based on the load balancing situation.

8. The MEC task offloading strategy method based on task division and D2D according to claim 1 is characterized in that: The improved fast elite non-dominated sorting genetic algorithm described in step S9 uses Tent mapping to initialize the population, and dynamically changes the crossover and mutation operators in the genetics to achieve adaptive evolution of the population.

Images