CN108021451B

CN108021451B - Self-adaptive container migration method in fog computing environment

Info

Publication number: CN108021451B
Application number: CN201711288967.4A
Authority: CN
Inventors: 贾维嘉; 唐志清; 周小杰
Original assignee: Shanghai Jiaotong University
Current assignee: Shanghai Jiaotong University
Priority date: 2017-12-07
Filing date: 2017-12-07
Publication date: 2021-08-13
Anticipated expiration: 2037-12-07
Also published as: CN108021451A

Abstract

The invention provides a self-adaptive container migration method in a fog computing environment, which comprises the following steps: establishing a fog computing frame based on a container, wherein the container is positioned on a fog node, a mobile application is positioned on a user body, and a task of the user is executed in the container; modeling a target of container migration in a fog calculation scene, wherein the migration target comprises time delay, power consumption and migration overhead; setting a state space and an action space, defining a return function and setting a Q iteration function; reducing the dimension of the state space through a deep neural network; and the dimension reduction of the action space is realized by optimizing the action selection. Finally, the prototype of the container self-adaptive migration system is realized and the whole process is verified. The self-adaptive container migration method in the fog computing environment provided by the invention can better plan resources in the fog computing, reduce time delay between a user and a fog node and reduce energy consumption overhead of the fog node.

Description

Self-adaptive container migration method in fog computing environment

Technical Field

The invention belongs to the field of fog calculation in a computer network, relates to methods such as fog calculation, moving edge calculation, reinforcement learning and deep reinforcement learning, and particularly relates to a self-adaptive container migration method in a fog calculation environment.

Background

Fog computing has become a promising computing paradigm in recent years, providing a flexible architecture to support distributed area-specific, domain-specific applications with cloud-like quality of service. Fog computing deploys a large amount of lightweight computing and storage infrastructure (called fog nodes) near mobile users. Therefore, the mobile application can be lowered to a proper fog node, and the access delay of the user to the application is shortened. In addition, the fog nodes are flexible and have expansibility, and the mobility of mobile users can be supported.

The existing technology of container migration in a fog computing environment by using a deep reinforcement learning method is rare. The related research field is mainly a scene of virtual machine management in a data center, and the main method is to dynamically migrate a virtual machine to partial nodes through migration of the virtual machine, so that idle fog nodes are closed, and the effect of reducing power consumption is achieved. The method for obtaining the migration scheme mainly predicts the resource requirements to obtain a pre-distribution scheme, or obtains some heuristic algorithms of the resource requirements through regression analysis based on historical information.

For the task scheduling problem in the fog computing scene, in the prior art, a two-dimensional simple Markov decision process model is mainly considered, when a single user moves, the distance between the user and a fog node is considered and modeling is performed to obtain a simple state space, and whether a task is migrated or not is judged by calculating a migration value function in the state space.

Directly migrating the method for managing the virtual machine in the data center to a fog computing environment can bring a series of problems, including a dimension disaster problem caused by a high-dimension state space and an action space, and a mobility problem of a mobile user is not considered in a modeling process, so that a time delay problem in a mobile scene cannot be well solved.

The existing task scheduling method for fog calculation only considers the condition of a single user when establishing a state space, and does not consider the actual condition of multiple users. And the transition probabilities between states are assumed to be fixed, whereas in practice the transition probabilities between states are unknown.

To overcome the above problems, the present invention proposes a container-based fog computing framework, placing applications in containers, and containers on fog nodes. In order to achieve optimal container scheduling, the container migration problem is regarded as a random optimization problem, and an algorithm suitable for a large Markov decision process state space and an action space is designed based on Q learning and deep learning strategies, so that the problem of dimension disaster is solved. On the basis of the system, the invention realizes a prototype system for container migration.

Disclosure of Invention

The invention provides a self-adaptive container migration method in a fog computing environment, which can better plan resources in fog computing, reduce time delay between a user and a fog node and reduce energy consumption overhead of the fog node.

In order to achieve the above-mentioned purpose and solve the above-mentioned problems, we first propose a set of container-based fog computing framework, then model the delay, power consumption and migration overhead in the fog computing scene under the framework, and design the adaptive container migration algorithm based on deep reinforcement learning, and finally realize the prototype of a container adaptive migration system and verify the whole process.

In order to achieve the above object, the present invention provides an adaptive container migration method in a fog computing environment, including the following steps:

establishing a fog computing frame based on a container, wherein the container is positioned on a fog node, a mobile application is positioned on a user body, and a task of the user is executed in the container;

modeling a target of container migration in a fog calculation scene, wherein the migration target comprises time delay, power consumption and migration overhead;

setting a state space and an action space, defining a return function and setting a Q iteration function;

reducing the dimension of the state space through a deep neural network;

and the dimension reduction of the action space is realized by optimizing the action selection.

Furthermore, each fog node has position data and a total amount of computing resources, wherein the computing resources include CPU resources, memory resources, storage resources, and bandwidth resources.

Further, each container has a resource request amount and an actual resource allocation amount, and each mobile application has location data and request data for the container.

Further, the time delay in the migration target is calculated by the following formula:

d_total＝d_net-k×d_comp，

wherein d is_netThe overhead generated by data transmission in the network is related to the distance between the user and the container, and is defined by path loss; d_compThe calculation time delay on the fog node is determined by the violation degree of the service level agreement with the fog node.

Further, the power consumption of the fog node is defined as follows:

wherein p is_idleAnd p_maxRefers to the power consumption, u, at CPU utilization of 0 and 100%_i(t) is the resource utilization of the fog node.

Further, the container migration overhead is defined as follows:

wherein m is_migIs a container C_iIncluding the transmission delay, 1{ "is eferson bracket.

Furthermore, the dimensionality reduction of the motion space comprises motion utilization, and after the state is obtained each time, the corresponding optimal Q value and the corresponding motion are selected from the Q value list.

Furthermore, the dimensionality reduction of the action space comprises action exploration, the agent randomly selects a state each time, limits the selection action, defines return income, and encourages migration when the income is positive.

Furthermore, the dimensionality reduction of the state space is to store all state information into a deep neural network, so that the dimensionality of the state space is reduced.

The self-adaptive container migration method under the fog computing environment has the following beneficial effects:

(1) according to the method, the user mobility is considered into the model, and the time delay between the user and the fog node is modeled, so that the time delay of the user task in the fog computing environment is well reduced, and the fog computing environment is better adapted.

(2) The method does not make any assumption on the transition probability, adaptively learns the actions to be taken in different states through the model-free autonomous learning algorithm, and can be well suitable for different fog computing environments.

(3) According to the method, the Q matrix of the storage state space is converted into the three-layer neural network similar to the Q matrix of the storage state space by using the deep neural network, so that the dimensionality of the state space in the fog computing environment is well reduced, and the problem of dimensionality disaster is solved.

(4) According to the method, the return revenue function which is beneficial to action selection is set through analysis of the specific conditions of fog calculation, and the condition that negative actions are selected during action selection can be effectively reduced, so that the convergence speed of the whole algorithm is effectively increased, and unnecessary energy consumption loss is better reduced.

(5) The method adopts container packaging application instead of the traditional virtual machine, can effectively reduce the cost generated during migration in the fog computing environment, and is more suitable for the fog computing environment with more limited resources.

Drawings

FIG. 1 shows a fog calculation framework and a user movement diagram.

Fig. 2 is a flow chart of an adaptive container migration method in a fog computing environment according to a preferred embodiment of the present invention.

FIG. 3 shows different omega₁Average delay versus case.

FIG. 4 shows different omega₁Average energy consumption in case is plotted.

FIG. 5 shows different omega₁Overhead versus situation.

FIG. 6 shows different omega₂Average delay versus case.

FIG. 7 shows different omega₂Average energy consumption in case is plotted.

FIG. 8 shows different omega₂Overhead versus situation.

FIG. 9 is a graph comparing the CPU overhead of a container and a virtual machine under different load conditions.

FIG. 10 is a graph comparing migration costs of containers and virtual machines under different load conditions.

Detailed Description

The following description will be given with reference to the accompanying drawings, but the present invention is not limited to the following embodiments. Advantages and features of the present invention will become apparent from the following description and from the claims. It is noted that the drawings are in greatly simplified form and that non-precision ratios are used for convenience and clarity only to aid in the description of the embodiments of the invention.

FIG. 1 shows a fog calculation framework and a user movement diagram. A total of five levels are included in fig. 1: user layer, access network layer, fog layer, core network layer and cloud layer. The user layer includes the mobile user and the mobile application being run on the mobile user. The mobile application accesses the fog layer through the access network layer and generates certain time delay. The fog nodes are located on the fog layer, the containers are located on the fog nodes, resources of the fog nodes are requested, and therefore the fog nodes generate expenses such as energy consumption. The fog nodes are connected with the cloud layer through the core network layer. The distance between the mobile user and the requesting container is increased due to the fact that the mobile user moves among the fog nodes, so that time delay is increased, and whether the container needs to be migrated along with the mobile user or not becomes a problem needing decision making. If the container is migrated, only the runtime library necessary for the application contained in the container and the application itself need to be migrated; and migrating the virtual machine needs to include the whole virtual machine system.

Referring to fig. 2, fig. 2 is a flow chart illustrating a method for adaptive container migration in a fog computing environment according to a preferred embodiment of the invention. The invention provides a self-adaptive container migration method in a fog computing environment, which comprises the following steps:

step S100: establishing a fog computing frame based on a container, wherein the container is positioned on a fog node, a mobile application is positioned on a user body, and a task of the user is executed in the container;

step S200: modeling a target of container migration in a fog calculation scene, wherein the migration target comprises time delay, power consumption and migration overhead;

step S300: setting a state space and an action space, defining a return function and setting a Q iteration function;

step S400: reducing the dimension of the state space through a deep neural network;

step S500: and the dimension reduction of the action space is realized by optimizing the action selection.

The invention first establishes a container-based fog calculation framework, and enables F to be equal to F₁,F₂,…,F_m}，C＝{C₁,C₂,…,C_n}，M＝{M₁,M₂,…,M_lRepresents a set of fog nodes, a set of containers, and a set of mobile applications, respectively. The container is located on the fog node, the mobile application is located on the user, and the user's tasks are performed within the container. Each fog node has a position F_iL, and the total amount of computing resources F_iAnd c, the computing resources mainly comprise CPU resources, memory resources, storage resources and bandwidth resources, and because the computing capacity is mainly considered in scheduling, the CPU resources are mainly considered here, and the memory resources, the storage resources and the bandwidth resources are considered to be sufficient. For containers, each container has a position C_iL (t) and each container has a resource request quantity C_iR (t) and an actual resource allocation C_iA (t). Furthermore, for mobile applications, there is one location M per mobile application_iL (t) and request for container M_i.r(t)。

We then model the target of container migration in the fog calculation scenario. The migration target mainly comprises the following aspects:

1. and (4) time delay. Time delay d_totalComprising two aspects, d_netAnd d_comp。d_netIs the overhead generated by data transmission in the network, mainly related to the distance between the user and the container, which can be defined by the path loss:

where f is the signal frequency, d_i(t) is the distance between the mobile application on the mobile user and the fog node where the corresponding container is located, h_bIs the height of the fog node, c_m3dB, ah in urban scene_mIs defined by:

ah_m＝3.20(log₁₀(11.75h_r))²-4.97，f＞400MHz，

wherein h is_rIs the user height.

In addition, d_compIs the calculated delay at the fog node, which is proven to be determined primarily by the degree of Service Level Agreement (SLA) violation (SLAV) of the fog node. And SLAV is defined as follows:

from the above, d_compCan be defined as:

and d_totalCan be defined as:

d_total＝d_net+k×d_comp。

2. power consumption. Power consumption p_totalRefers to the power consumption of all the fog nodes. If the fog node is in sleep mode, then the power consumption is approximately 0, and in addition to this, the power consumption of the fog node is defined as follows:

wherein p is_idleAnd p_maxRefers to the power consumption when the CPU utilization is 0 and 100%. u. of_i(t) is the resource utilization of the fog node, defined as follows:

3. container migration overhead. The container migration overhead is defined as follows:

wherein m is_migIs C_iThe migration overhead of (2) includes transmission delay. 1{ "is eferson bracket.

4. A problem model. So far, a model of the whole problem can be obtained

After modeling the above problem, the state space and the operation space are set. Since the experiment is mainly related to M_iL (t) and C_iR (t) about, define:

wherein:

furthermore, combining C.l (t) and C.a (t), the state space of the system can be derived:

according to the practical situation, the obtained corresponding action space is as follows:

since a minimum of overhead is required, the reward function is defined as:

R_τ＝-(d_total(τ)+ω_1ptotal(τ)+ω₂m_total(τ))

then, setting a Q iteration function:

obviously, a huge state space brings dimensionality disasters, so the dimensionality reduction is carried out on the state space through a deep neural network, and the dimensionality reduction of an action space is realized through the optimization of action selection. The method mainly comprises the following three steps:

1. and (4) utilizing the action. Maintaining a Q value list of optimal Q values

Each of which is composed of

And (4) forming. In the step of action utilization, after the state is obtained each time, the corresponding optimal Q value and the corresponding action are selected from the Q value list.

2. And (5) action exploration. In the action exploration phase, the agent randomly selects a state at a time. In order not to make the selected state too random resulting in a negative optimization, certain restrictions are placed on the selection action. And (3) defining return income:

when in use

The revenue is positive and migration is encouraged. Migration probability:

the action is finally obtained:

wherein:

a random action selection algorithm can thus be derived:

3. the deep neural network reduces the state space. And storing all state information into the deep neural network so as to reduce the state space dimension. The training target of the neural network is defined as:

L(θ_τ)＝E[(y(τ)-Q(S_τ，A_τ；θ_τ))²]

wherein:

y(τ)＝E[(1-α)Q(S_τ-1，A_τ-1；θ_τ-1)+α[R_τ-1+γmaxQ(S_τ，A_τ；θ_τ-1)]|S_τ-1，A_τ-1].

furthermore, through empirical playback, the association between each training is reduced. The training algorithm is as follows:

and a final container adaptive migration algorithm:

finally, on the basis, a set of prototype system for container migration is realized.

The invention uses Python for programming, simulating a fog node, a container and a user. The fog node class comprises a position initialization module, a fog node distance calculation module, a power consumption calculation module, a CPU resource module, a container list module, a user list module and a bandwidth module. The container class comprises a number maintenance module, a CPU resource use module, a position updating module, a position module, a migration overhead module and a size module. The user class comprises a position initialization module, a position updating module, a request initialization module, a request updating module, a distance calculation module with the fog node and a time delay module. The fog nodes, containers and users make up the environment of the entire system. In addition, the core part is of a Q learning class and comprises an Agent class of an intelligent Agent, a Brain class of a deep neural network part and a Memory playback Memory class. The Agent class of the Agent comprises an optimal action obtaining module, a current (state, action, return value and next state) tuple obtaining module, a Memory playback module, a preprocessing module and a neural network training module, wherein the deep neural network Brain class comprises a network structure module, and the Memory class comprises a Memory storage module, a Memory extraction module and a Memory table.

Experimental data were from real taxi cab data from san francisco, with data sets ranging from 32.87 to 50.31 latitudes and-127.08 to-122.0 longitudes. The area is divided, 7 fog nodes are deployed, and the movement situation of more than 200 users is considered. All mobile users are active and use 0's and 1's to indicate whether they are getting on or off the vehicle, thereby indicating a switch requested by the application.

For the setting of the parameters, the power consumption of the CPU is given by the following table:

CPU Utilization(％)	0％	10％	20％	30％	40％	50％
							HP ProLiant G4	86	89.4	92.6	96	99.5	102
CPU Utilization(％)	60％	70％	80％	90％	100％
							HP ProLiant G4	106	108	112	114	117

TABLE 1 CPU utilization and energy consumption relationship table

For other parameters, let f be 2.5MHz, h_b＝35m，h_r＝1m，c_m3dB, and

di(t)＝|M_i.l(t)-F_j.l(t)|

further, X_scale＝3，alpha＝0.1，gamma＝0.9，epsilon＝0.9。

To compare the experimental results, two baseline algorithms were chosen. The algorithm is abbreviated as ODQL, in addition, the DBQL algorithm is obtained by discretizing the traditional Q learning algorithm, and the basic algorithm of the invention is also an approximately greedy algorithm Myovic.

By targeting different omega₁For comparison, the following results were obtained. Referring to FIGS. 3-5, FIG. 3 shows different omega₁Average delay versus time for the case, FIG. 4 shows the average energy consumption versus time for different omega1, FIG. 5 shows the average energy consumption versus time for different omega₁Overhead versus situation.

In addition, the present invention is also applicable to different omega₂The comparison was performed, and the experimental results are shown below. Referring to FIGS. 6-8, FIG. 6 shows different omega₂Average delay versus time for the case, FIG. 7 shows different omega₂Average energy consumption in the case of a graph, FIG. 8 shows different omega₂Overhead versus situation. The experimental results show that the algorithm provided by the invention has better effect than the other two algorithms.

In addition, the invention also builds a prototype of the container migration system. A CPU of E5-1650v2@3.5GHz, a memory of 16.0 GB and a desktop of an operating system of Ubuntu 16.04 LTS are used as fog nodes, a CPU of i7-4600U @2.1GHz, a memory of 8.0 GB and a notebook computer of an operating system of Windows 10 are used for simulating a user group. The desktop is provided with a Docker container engine, and an Nginx website server, a WordPress site and MySQL database, a Ghost site and SQLite3 database and a Docker container of a static page are arranged through the Docker. Different Ubuntu containers are managed on the notebook through a Docker container engine, a Webbenchmark simulation user request is installed in each Ubuntu container, and time delay is changed through a tc tool. Meanwhile, the virtual machine migration environment under the same hardware environment is set and compared. The comparative results are as follows. As shown in fig. 9 and 10, fig. 9 is a diagram showing a comparison of CPU overhead of the container and the virtual machine under different loads, and fig. 10 is a diagram showing a comparison of migration overhead of the container and the virtual machine under different loads. According to experimental results, under the same hardware condition, the migration cost of the container is far less than that of the virtual machine, so that the container self-adaptive migration system in the fog computing environment provided by the invention is very effective.

Although the present invention has been described with reference to the preferred embodiments, it is not intended to be limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. An adaptive container migration method in a fog computing environment, comprising the steps of:

modeling a migration target of container migration in a fog calculation scene, wherein the migration target comprises time delay, power consumption and migration overhead;

reducing the dimension of the state space through a deep neural network;

2. The method of adaptive container migration in a fog computing environment of claim 1, wherein each of the fog nodes has location data and a total amount of computing resources, wherein computing resources include CPU resources, memory resources, storage resources, and bandwidth resources.

3. The method of adaptive container migration in a fog computing environment of claim 1, wherein each of the containers has a requested amount of resources and an actual allocated amount of resources, and each of the mobile applications has a location data and requested data for a container.

4. The adaptive container migration method in a fog computing environment according to claim 1, wherein the time delay in the migration target is calculated by the following formula:

d_total＝d_net+k×d_comp，

wherein d is_netThe overhead generated by data transmission in the network is related to the distance between the user and the container, and is defined by path loss; d_compThe calculation time delay on the fog node is determined by the violation degree of the service level agreement of the fog node.

5. The adaptive container migration method in a fog computing environment as claimed in claim 1, wherein the power consumption of the fog node is defined as follows:

6. The adaptive container migration method in a fog computing environment of claim 1, wherein the container migration overhead is defined as follows:

wherein m is_migiIs a container C_iIncluding transmission delay, 1 {. is eferson brackets.

7. The method of claim 1, wherein the dimensionality reduction of the motion space comprises motion utilization, and each time a state is obtained, a corresponding optimal Q value and a corresponding motion are selected from a Q value list.

8. The method of claim 1, wherein the dimensionality reduction of the action space comprises action exploration, wherein an agent randomly selects a state each time, limits selection actions, defines a return benefit, and encourages migration when the benefit is positive.

9. The adaptive container migration method in the fog computing environment according to claim 1, wherein the state space is reduced in dimension by storing all state information in a deep neural network.