CN110413392B

CN110413392B - Method for formulating single task migration strategy in mobile edge computing scene

Info

Publication number: CN110413392B
Application number: CN201910675567.1A
Authority: CN
Inventors: 方娟; 徐玮豪; 陈勇
Original assignee: Beijing University of Technology
Current assignee: Beijing University of Technology
Priority date: 2019-07-25
Filing date: 2019-07-25
Publication date: 2022-11-29
Anticipated expiration: 2039-07-25
Also published as: WO2021012584A1; JP6942397B2; JP2021525402A; CN110413392A

Abstract

The invention provides a method for formulating a single task migration strategy in a mobile edge calculation scene, which solves the problems of interaction loss and increased base station bandwidth pressure caused by adopting an integral migration scheme in the mobile edge scene. The method comprises the following concrete steps: firstly, a task needing migration calculation is divided into different subtasks with mutual dependency relationship, each subtask can be guaranteed to be capable of independently performing calculation processing, and meanwhile, the position of a subtask node which cannot be migrated in a graph is determined. Secondly, a weighted directed acyclic graph is generated according to the dependency relationship among the subtasks, each node in the graph represents the calculated amount of the data, and each edge represents the communication amount of the data among different components. And then, iteratively calculating the specific execution position of each migratable subtask by using an ant colony algorithm, namely confirming whether the migration to the edge computing server or the local completion of the operation is carried out. And finally obtaining a suboptimal solution of the single task migration strategy based on the ant colony algorithm with the aim of reducing the energy consumption of the mobile equipment.

Description

Method for formulating single task migration strategy in mobile edge computing scene

Technical Field

The invention belongs to the field of mobile edge calculation, and aims to reduce the task migration energy consumption of mobile equipment.

Background

The popularization of mobile terminals such as smart phones or tablet computers has a profound impact on mobile and wireless networks, and thus has led to the revolution of global mobile networks. Such mobile devices face a network environment of low storage capacity, high energy consumption, low bandwidth and high latency. Mobile Cloud Computing (MCC), as an integration of cloud computing with mobile computing, brings considerable capabilities to mobile devices and provides storage, computing, and energy through a centralized cloud. However, with the advent of large numbers of mobile devices, MCC is facing more severe challenges such as high latency, security holes, and low network coverage. These problems may become more pronounced in the context of next generation mobile networks (e.g., 5G). According to recent reports of the Cisco visual network index, by 2020, 116 billion mobile connected devices will be used globally. To address the increasing network demand, the concept of Mobile Edge Computing (MEC) was born. The main purpose of MEC is to solve the challenges from MCC systems. MEC enhances the capability of MCC by deploying cloud resources (e.g., storage and computing capabilities) to the edge of the radio access network. This provides the end-user with fast and powerful computing power, energy efficiency, storage capacity, mobility and context-aware support. Previously, internet edge technology called cloudlet has been introduced to deploy mobile cloud services. However, clouldets still do not address the existing challenges due to their limited WiFi coverage. There are still a number of problems to be solved in the research field of MEC.

In academia, rudenko a et al originally proposed task migration to effectively reduce energy consumption of mobile devices and extend operating time. The study verified the above guess by an experiment performed by a laptop uploading a computationally complex program to a remote desktop. Wen Y and Zhang W et al propose a joint optimization task migration scheme that minimizes computational energy consumption by optimally scheduling the clock frequency of a mobile device when the mobile application executes locally on the mobile device, and minimizes transmission energy consumption by configuring the transmission power of a wireless channel when the mobile application executes in a cloned cloud. Since running the task migration algorithm itself consumes computational resources and energy of the mobile device, huerta-Canepa G et al propose a scheme for task migration based on the execution history of the mobile application and the current application state. The scheme sets the self-adaptive adjustment of task migration decision: when the self computing resources and the battery electric quantity of the mobile equipment are sufficient, a dynamic decision scheme is used, and the task execution performance is improved; when the resources of the mobile equipment are insufficient, a migration decision is made before the task is executed according to the historical state, and the extra expense caused by dynamic decision is reduced. When a task is migrated, a wireless network also affects the energy consumption for task completion, huang D and Wang P and the like propose a method for dynamically changing a task migration strategy according to the wireless network environment, construct a mobile application into a directed acyclic graph model of a plurality of subtasks, and minimize the execution energy consumption of the mobile application by allocating the execution positions of the subtasks.

In addition, some studies optimize both task execution time and energy consumption. WuHuaming et al propose a task migration scheme that trades off between shortening execution time and saving energy consumption, achieving flexible, on-demand allocation of cloud computing resources. Li Tianze et al propose an optimization scheme that integrates energy consumption, time delay, and server execution cost for task migration in MEC environment. The complexity of a task migration algorithm with multiple targets coexisting is often too high, and in order to reduce the time complexity of the migration algorithm, wang J et al proposes a low-complexity task migration algorithm based on Lyapunov (Lyapunov) optimization theory, and can reduce the execution time and the energy consumption of a mobile device at the same time.

Compared with the traditional mobile cloud computing environment, in the new mobile edge computing environment, the distance between the user and the MEC server is closer, so that the communication overhead of task migration on data transmission is greatly reduced. In the traditional task migration strategy, a task is taken as a whole during preparation, if the task is migrated, the task is completely handed to an MEC server for processing, and if the task is not migrated, the task is executed locally. Such migration strategies are clearly not optimal for mobile devices that are in frequent data communication with the server. The task migration mode of splitting the single task and making the migration decision can improve the task execution performance, reduce the task execution overhead and divide the calculation task with finer granularity. At this time, it is very important to design a set of task migration algorithm by using the specific characteristics of the tasks (task topology, task computation amount, size of data transmission amount between tasks, etc.).

Disclosure of Invention

In the prior art, when a task migration decision is made, the migration decision is made by taking one task as a unit. There may be many tasks in a real scene that require frequent interaction with the mobile device and these operations must be performed locally. Therefore, the current technical solution is a part that is not applicable to a real scene. In a moving edge scenario, a user's mobile device may frequently interact with a base station of an operator, and if an overall migration scheme is adopted, the bandwidth pressure of the base station may be increased while the interaction capability is lost, which obviously is not suitable for practical application.

The invention designs a task migration strategy based on an ant colony algorithm in a single-user MEC system, calculates the probability of the current subtask migration decision by converting application into a directed graph containing a plurality of subtasks and introducing the concept of pheromone in the ant colony algorithm, generates a group of scheduling strategies by using the minimization of energy consumption of mobile equipment as an optimization target, iterates the algorithm for multiple times, and continuously optimizes the generated scheduling strategies to approach the optimal task scheduling strategy. The invention fully considers the condition of each subtask to formulate an overall migration strategy, ensures that each non-migratable subtask can be executed locally to meet the requirement of user interaction, is more suitable for real scenes, and improves the optimization efficiency.

In order to achieve the purpose, the invention designs the following scheme, which comprises the following steps:

step 1, when a random task arrives, the random task is temporarily stored in a buffer queue, when a system schedules the task to execute, the task is divided into a plurality of subtasks which can be independently executed, and the subtasks can be expressed as a set V = { V = (V) = ₁ ,v ₂ ,…v _i ,…v _j ,…,v _N And N is expressed as the total number of the subtasks. Simultaneous partitioning of V into two distinct sets V _loc 、

Wherein V _loc Representing a non-migratable component, V, that must be executed locally _off Represents a collection of components for decision making that can be migrated to the MEC server. And each task set V has a unique entry transaction and an exit transaction, wherein the entry transaction has no predecessor transaction, and the exit transaction has no successor transaction. The invention will also define a binary variable E _ij E {0,1} to represent the dependencies between the various tasks:

for

All have an e _ij And is used for representing the data transmission quantity between the task i and the task j. And finally, forming a directed acyclic graph G = (V, E) through the subtask set V and the transaction dependency set E.

And step 2, determining and initializing various parameters of the mobile edge calculation model, and establishing an energy consumption model.

The task migration strategy provided by the invention aims to optimize the execution energy consumption of the mobile equipment end, so that the position of task execution (local execution or MEC end execution) needs to be defined, and a set A = { A = (A execution on local execution or MEC end) is used ₁ ,A ₂ ,……,A _N Denotes the execution position of each task, and

the invention uses omega (CPU cycles) to represent the task calculation amount, f represents the CPU execution rate of the device, and T represents the execution time of the task. If the task is executed locally, the local execution time may be: t is _l ＝ω _i f _l ^-1 . If the task is at the calculation speed f _c When the MEC server end executes, the time required for task completion is as follows: t is _c ＝ω _i f _c ^-1 。

Assuming that P is the unit of power when the CPU performs the task is (W), the power consumption of the CPU of the mobile device performing the task locally can be expressed as: e _l ＝P _l T _l . If the task is executed at the MEC server side, the mobile terminal equipment does not need to perform task operation at the moment, but still needs to consume basic energy to maintain the equipment to operate, and the energy consumption is E _b ＝P _b T _c And (4) showing. Wherein P is _b (W) represents power at which the mobile device CPU is idle, T _c (s) represents mobile idle time. Due to P _b Much less than P _l Therefore, taskThe migration policy may save energy consumption for the mobile device.

In terms of data transmission consumption, with R _s And R _r Respectively representing the channel rate (from the mobile end to the MEC end) of data uploading and the channel rate (from the MEC end to the mobile end) of data downloading in the unit of (bit/s), P _s And P _r The unit is (W) indicating communication power at the time of data transmission and data reception, respectively.

When the task j is executed at the MEC server side and the preposed task i is executed at the mobile equipment side, the transmission time of the task is as follows:

the energy consumed was:

when the task j is executed at the mobile terminal and the preceding task is executed at the MEC server terminal, the transmission time of the task is as follows:

the energy consumed was:

according to the energy consumption model constructed above, the total energy consumption of the entire mobile device to execute to complete a single application can be expressed as:

n denotes the total number of subtasks, the second part of the equation on the right represents the sum of the energy consumptions from the first to the second last subtask, where [ E _l (1-A _i )+E _b A _i ]Representing the power consumption of the mobile device CPU. Of formula (3)

The part represents the total transmission energy consumption of the task, | A _i -A _j I is used for judging task i and the subsequent taskIf the affair j is operated at the same position, and if the affair j is operated at the mobile equipment or the MEC server, transmission energy consumption cannot be generated. Since the last subtask is determined to be executing locally and there are no subsequent tasks, the energy consumption of its task is added to the front-most end of the energy consumption calculation model.

So far we have established tasks as a model to minimize the total energy consumption E (a) of the mobile device, since there are two options for each migratable task, migration or not, then the total migration decision for N tasks will exist 2 ^N And (4) solving. If the enumeration method is used for calculating the optimal energy consumption solution of the task, the time complexity is too high, and the method is not suitable for actual production. The present invention uses ant colony algorithms to solve this complex task model.

And 3, initializing pheromone concentration, task cycle times t and ant number m in each path.

In order to calculate the specific energy consumption of the task, the actual calculation position of each subtask has to be determined, i.e. the set a = { a = { (a) } ₁ ,A ₂ ,……,A _N The value of. The task calculation amount and the data communication amount required by each subtask are different, in order to reduce task energy consumption as much as possible, the calculation-dependent subtask is more prone to be transmitted to the edge calculation node for operation, and the subtask with high data communication amount but low calculation amount is handed to the mobile device for local processing. The ant colony algorithm calculates the task migration probability according to the concentration of the pheromones on different paths, and therefore the task migration strategy is obtained. The invention uses tau _c (0)＝{τ _1c (0)、……、τ _Nc (0)}、τ _l (0)＝{τ _1l (0)、……、τ _Nl (0) Denotes the pheromone concentration of each subtask on the migration path and the non-migration path when the ant colony algorithm starts to execute respectively, and the concentration of each subtask on the migration path and the non-migration path is calculated

Having a value of _ic (0)＝τ _il (0) = δ, (δ ∈ (0, 1)). For the

Having a value of _il (0) = + ∞. And initializing the number of task cycles and the number m of ants in each cycle.

Step 4, obtaining a path selected by each ant by utilizing an ant colony algorithm, and selecting minimum energy consumption E from all paths selected by m ants according to the total task energy consumption model designed in the step 2 _min (A) The corresponding path is taken as an optimal task migration strategy under the task cycle, and the step 5 is continuously executed after all m ants finish the task under the task cycle

Wherein, the path selected by the ant is determined by the migration calculation probability, and under the t-th task cycle, each ant migrates the subtask i with the calculated probability P _ic The calculation formula of (t) is as follows:

the meaning of each symbol in the above formula is as follows:

t represents the number of task cycles, also the time of day;

·τ _ic (t) represents the concentration of pheromones on the path for migrating the task i to the MEC server at the time t, t _il (t) task i calculates the concentration of pheromones on the path locally at t moment;

α represents a pheromone heuristic factor (α ∈ [0,5 ]), which reflects the role of pheromones on ant path selection;

·

is a heuristic function which represents the expected degree of the task i needing to be migrated, and takes the value as

Thus can see e _ij The smaller the

The larger, i.e. the higher the expectation for task i migration;

beta represents a heuristic function factor (beta ∈ [0,5 ]), reflecting the relative importance of the heuristic function in directing ant colony searches;

after the execution positions of the subtasks are obtained by the formula 4, the number k of ants is reset, each ant executes the task according to the task migration strategy of the round, the energy consumption of the path selected by each ant is calculated according to the energy consumption model designed in the step 2, and the lowest energy consumption E is updated _min (A) In that respect And (5) continuing to execute the step after all m ants in the round complete the task.

Step 5, if the preset task cycle times are not reached, updating the pheromone concentration, and returning to the step 4 to continuously search for a more optimal task migration strategy; and if the preset task cycle number is reached, continuing to execute the step 6.

The pheromone concentration updating formula is as follows:

τ _ic (t+1)＝(1-ρ)*τ _ic (t)+Δτ _ic (t,t+1) (5)

wherein rho is a pheromone volatilization factor (rho epsilon [0.1,0.99)]) 1-p denotes the residual pheromone factor, Δ τ _ic (t, t + 1) is expressed as the increment of the pheromone after one task iteration, and is calculated by formula 6:

m is the total number of ants in one circulation,

the pheromone left on the path of the kth ant in task migration at task i is shown, and the pheromone left on the migration path of each ant at task i is shown by formula 7, wherein Q is a normal number (Q is epsilon [1,10 ])]) For controlling the amount of pheromones left by each ant.

And 6, the optimal task migration strategy obtained by the last task cycle is the optimal task migration strategy, task allocation is carried out according to the optimal task migration strategy, and edge calculation is executed.

Compared with the prior art, the invention has the following characteristics:

the design of the invention considers that in the actual MEC scene, a plurality of applications which need to interact with mobile users frequently exist, and the integral migration operation of the application program undoubtedly increases communication overhead greatly, resulting in higher energy consumption of mobile equipment. According to the method, the application to be processed is firstly converted into the directed graph comprising a plurality of subtasks, then the ant colony algorithm is utilized to traverse the task graph to be processed for multiple times, and finally the task migration strategy suboptimal solution taking energy consumption as an optimization target is obtained. Compared with other algorithms, on the basis of ensuring the task execution efficiency, the method reduces the time complexity of task execution, and meanwhile, the method of splitting the task in a fine granularity reduces the energy consumption of the mobile equipment to the maximum extent and improves the service quality of the whole MEC system.

Drawings

In order to make the purpose of the present invention more comprehensible, the present invention will be further described with reference to the accompanying drawings.

FIG. 1 is a fine-grained task partitioning diagram of the present invention;

FIG. 2 is a task execution flow diagram;

Detailed Description

Step 1, as shown in fig. 1, is a fine-grained task partitioning diagram of an application, and the present invention partitions the application into multiple sub-tasks that are executed independently and is represented by a directed graph G = (V, E). The node V ∈ V in fig. 1 represents the divided subtask, and the edge e in fig. 1 _ij E represents the transmission data between tasks, for example: e.g. of a cylinder _ij Indicating that task i is completed, e is transmitted _ij The data of the task j is sent to the task j, and the task j can only start to execute after the data transmitted after the task i is executed is received. The subtasks in the figure can be divided into two categories 2: one type is tasks that have to be performed locally (e.g. interaction of the user's audio-video capture with the mobile terminal, etc.), represented as

solid tasks

1, 4, 6 in fig. 1, represented as

Another type is migratable tasks, such as the

hollow tasks

2, 3, 5 in FIG. 1, denoted as

The invention defines a binary variable E _ij E {0,1} to represent the dependencies between the various tasks. Formula 1 shows that when E _ij When =1, the task j can start to execute after the task i completes execution, otherwise E _ij =0. For example, in FIG. 1E ₁₂ ＝1，E ₄₅ ＝1，E ₆₃ =0, and if there are two or more pre-tasks for a task, it must be executed until all the pre-tasks are completed.

And 2, establishing an energy consumption model for calculating the moving edge, and initializing task parameters. The energy consumption model is divided into two parts, namely local energy consumption calculation and migration energy consumption calculation. Let the calculation amount of each sub-task be ω _i (CPU cycles) with CPU execution rate f _l Power P when performing the calculation _l Then the task local execution energy consumption can be expressed as: e _l ＝P _l ω _i f _l ^-1 . If the task needs migration calculation, standby energy consumption of the mobile equipment during task migration can be used _b ＝P _b ω _i f _c ^-1 The invention is used respectively because the task migration must generate communication energy consumption

And

representing the energy consumption of data upload and download. The task migration strategy provided by the invention aims to optimize the execution energy consumption of the mobile equipment end, and therefore, the task execution position (local execution or MEC end execution) needs to be defined, and A is used _i E {0,1} represents the execution position of a certain task. Obviously, all must be inLocally executed tasks can only be operated on by the mobile device, so for

The total energy consumption of the task can be represented by equation 3.

Step 3, on the basis of step 2, only the set A = { A } needs to be determined to obtain the specific energy consumption of task execution ₁ ,A ₂ ,……,A _N The value of. First of all according to the initialized pheromone concentration tau _c (0)、τ _l (0) And formula 4, calculating the first round task migration probability P _ic (0). Then the first round of k ants passes through P _ic (0) Can respectively obtain the first time task migration policy a = { a = { (a) ₁ ,A ₂ ,……,A _N Value of (solving method is as follows: suppose P _1c (0) = λ, in which case a number μ e (0, 1) is generated by a completely random method, if 0<λ is less than or equal to μ A ₁ =1, if 1>λ>Mu then A ₁ =0. ) Finally, the total energy consumption of the task can be obtained by using a formula 3, and the minimum value in the k energy consumptions is selected and recorded as E _min (A)。

Step 4, according to the minimum energy consumption E _min (A) And equations 5-7, update pheromone concentrations on different paths. Calculating the task allocation probability P of the first round by using the new pheromone concentration and formula 4 _ic (t) by P _ic (t) a new task migration strategy A can be obtained, and then the optimal energy consumption E is updated _min (A)。

Step 5, if the preset task cycle times t are not reached, returning to the step 4 to continuously search for a more optimal task migration strategy; and if the preset task cycle number t is reached, continuing to execute the step 6.

And 6, the optimal task migration strategy obtained by the last task cycle is the optimal task migration strategy, task allocation is carried out according to the optimal task migration strategy, and edge calculation is executed. The specific execution flow chart is shown in fig. 2.

Claims

1. A method for making a single task migration strategy under a mobile edge computing scene is characterized by comprising the following steps:

step 1, when a random task arrives, the random task is temporarily stored in a buffer queue, and when a system schedules the task to execute, the task is divided into a plurality of subtasks which can be independently executed, and the subtasks can be represented as a set V = { V = (V) } in a set manner ₁ ,v ₂ ,…v _i ,…v _j ,…,v _N N is expressed as the total number of the subtasks, and V is divided into two different sets V _loc 、

Wherein V _loc Representing a non-migratable subtask that must be executed locally, V _off Representing a set of subtasks for decision making, which can be migrated to the MEC server, and each task set V has a unique entry transaction and exit transaction, wherein the entry transaction has no predecessor transaction and the exit transaction has no successor transaction, and a binary variable E is defined _ij E {0,1} to represent the dependencies between the various tasks:

for the

All have an e _ij The system is used for representing the data transmission quantity between the task i and the task j, and finally a directed acyclic graph G = (V, E) is formed through a task set V and a transaction dependency set E;

step 2, combining task calculation and a task transmission energy consumption model to establish a task total energy consumption model, and initializing each parameter of the total energy consumption model;

step 3, initializing pheromone concentration, task cycle times t and the number m of ants in each cycle in each path;

step 4, obtaining the path selected by each ant by utilizing an ant colony algorithm, and selecting the minimum energy consumption E from all paths selected by m ants according to the total task energy consumption model designed in the step 2 _min (A) The corresponding path is used as an optimal task migration strategy under the task cycle, when all m ants finish the task under the task cycle, the step 5 is continuously executed,

wherein, the path selected by the ant is determined by the migration calculation probability, and under the t-th task cycle, the probability P calculated by migration of each ant to the subtask i _ic The calculation formula of (t) is as follows:

the symbols in the above formula have the following meanings:

t represents the number of task cycles;

·τ _ic (t) indicates the concentration of pheromones on the path for migrating task i to the MEC server at the time t,

τ _il (t) the task i calculates the concentration of the pheromone on the path locally at the time t;

α represents an pheromone heuristic factor, which reflects the effect of pheromones on ant routing;

·

is a heuristic function which represents the expected degree of migration required by the task i and takes the value of

It can be seen that e _ij The smaller the

The larger, i.e. the higher the expected value of task i migration;

beta represents a heuristic function factor reflecting the relative importance of the heuristic function in directing ant colony search;

step 5, if the preset task cycle times are not reached, updating the pheromone concentration, and returning to the step 4 to continuously search for a more optimal task migration strategy; if the preset number of task cycles is reached, the execution continues to step 6,

the pheromone concentration updating formula is as follows:

τ _ic (t+1)＝(1-ρ)*τ _ic (t)+Δτ _ic (t,t+1) (5)

where ρ is pheromone volatility factor, 1- ρ represents residual pheromone factor, Δ τ _ic (t, t + 1) is expressed as the increment of the pheromone after one task iteration, and is calculated by formula 6:

m is the total number of ants in one circulation,

the pheromone left on the path of task migration of the kth ant at task i is shown, the pheromone left on the path of migration of each ant at task i is shown by formula 7, wherein Q is a normal number used for controlling the number of pheromones left by each ant,

herein E _min (A) Representing the minimum energy consumption corresponding to the t-th task cycle;

step 6, the optimal task migration strategy obtained by the last task cycle is the optimal task migration strategy, task allocation is carried out according to the optimal task migration strategy, and edge calculation is executed;

wherein:

the task calculation energy consumption model in step 2 is as follows: if the task is executed locally, the energy consumption of the mobile device for executing the task locally is as follows: e _l ＝P _l T _l Wherein, P _l Power when executing tasks for local CPU, task execution time

ω _i Represents the amount of computation of task i, f _l Representing the CPU execution rate of the local device;

if the task is executed at the MEC server side, the basic energy consumption E of the mobile equipment _b ＝P _b T _c Wherein P is _b Indicating power and task execution time of mobile device CPU idle

f _c Representing the CPU execution rate of the MEC server;

the task transmission energy consumption model in step 2 is as follows:

when the task j is executed at the MEC server side and the pre-task i is executed at the mobile device side, the consumed energy is:

wherein the transmission time of the task

When the task j is executed at the mobile terminal and the pre-task is executed at the MEC server terminal, the consumed energy is:

wherein, the transmission time of the task is as follows:

wherein R is _s And R _r Respectively representing the channel rate for data upload and the channel rate for data download, P _s And P _r Representing the power of the mobile device at the time of data transmission and data reception, respectively;

the total task energy consumption model described in step 2 is as follows:

wherein the set a = { a = ₁ ,A ₂ ,……,A _N Denotes the execution position of each task, and

2. the method according to claim 1, wherein the method for formulating the single task migration policy in the mobile edge computing scenario comprises:

the pheromone concentration in the step 3 comprises the pheromone concentration tau of each subtask on the migration path when the ant colony algorithm starts to execute _c (0)＝{τ _1c (0)、……、τ _Nc (0) And pheromone concentration tau of each subtask on non-migratory paths _l (0)＝{τ _1l (0)、……、τ _Nl (0) And to

With tau _ic (0)＝τ _il (0) δ, δ ∈ (0, 1), for

With tau _il (0)＝+∞。