CN115457369A - Real-time video analysis task execution delay modeling and deployment method based on cloud-edge collaboration - Google Patents

Abstract

The invention discloses a cloud-edge collaborative real-time video analysis task execution delay modeling and deployment method, which comprises the following steps: a video analysis task execution delay model construction method under a cloud edge environment; the time delay model is extra time delay introduced in the process of a video analysis task, and comprises time delay brought by edge end processing, time delay brought by data transmitted between an edge and a cloud end and time delay brought by cloud end processing; according to a video analysis task execution time delay model in a cloud edge environment, selecting a more energy-saving task deployment scheme for the video analysis task through an approximation algorithm based on a Lyapunov algorithm and a Markov approximation algorithm, wherein the task deployment scheme comprises edge end machine specification selection, cloud end machine instance specification and a task division scheme. According to the invention, when the task load of the video analysis task and the cloud side network environment change, the task deployment scheme can be rapidly adjusted, so that the energy consumption of the system in long-term operation is effectively reduced.

Description

Real-time video analysis task execution delay modeling and deployment method based on cloud-edge collaboration

technical field

The invention relates to the field of edge computing, and relates to a cloud-edge collaborative system energy-saving deployment method, in particular to a cloud-edge collaborative real-time video analysis task execution delay modeling and deployment method.

Background technique

With the development of the mobile Internet, a large number of highly interactive applications have begun to appear, such as the recently popular AR applications and various face and license plate recognition applications. For such highly interactive applications, users often have more High response latency requirements. The transmission delay from the terminal to the data center needs to be forwarded through multi-layer networks such as base stations and backbone networks. The network delay is often high, and a large number of high-quality videos transmitted will also bring great bandwidth pressure to the WAN. In order to reduce access latency, these applications introduce edge computing. Therefore, edge computing is gradually becoming a new network infrastructure in the mobile Internet era. Edge computing can reduce task response delay by deploying services closer to users, such as urban backbone networks, base stations, or edge devices, and reduce the huge pressure on WAN bandwidth caused by raw data transmission. In addition, offloading services from the data center to the user's edge can avoid sending raw data to the cloud, thereby improving the privacy of user data.

Real-time video broadcasting and real-time video monitoring are new types of video applications emerging with the development of technologies such as CDN and edge computing, and are widely used in intelligent systems such as autonomous driving and disaster monitoring. Such applications require real-time video streaming for extended periods of time.

When the real-time video stream analysis task has a long video processing pipeline, the neural network model used for video analysis itself is large, or the computing power of the edge is too poor, it is difficult to deploy the real-time video stream analysis task on the edge alone to meet the computing power requirements. However, the traffic overhead and transmission delay caused by completely placing in the cloud are unacceptable, and the cloud-side collaborative deployment of real-time video stream analysis tasks has become a more reasonable deployment solution.

Choosing an optimal deployment solution is very important in the actual production environment. A poor solution will not only bring higher power consumption or cost, but also make the task processing time longer.

When current mainstream cloud service providers, such as Amazon Web Service, Azure, Alibaba Cloud and other companies provide cloud services, the number of cloud optional configurations is large or even reaches thousands of available configurations. For large-scale neural networks or long processing pipelines, there are also dozens or even hundreds of strategies that can be divided between cloud and edge, so the total number of cloud-edge deployment methods is large. In order to obtain the additional delay caused by a specific deployment method when performing a specific task, it is necessary to actually run the corresponding deployment method, so it will cost a lot of money to measure the service quality of each deployment method. The method of measuring all deployment methods cannot meet the requirements of rapid deployment execution of tasks. Therefore, it is difficult to choose the optimal deployment method for the current task situation among the large number of optional deployment methods whose actual performance will change.

Contents of the invention

In order to solve the deficiencies of the prior art, the object of the present invention is to provide a cloud-side collaborative real-time video analysis task execution delay modeling and deployment method.

The object of the present invention is achieved through the following technical solutions:

A cloud-edge collaborative real-time video analysis task execution delay modeling and energy-saving deployment method, which includes the following steps:

A. The construction method of the video analysis task execution delay model in the cloud-edge environment; the delay model is the additional delay introduced in the process of the video analysis task, including the delay caused by the processing at the edge, and the transmission between the edge and the cloud The delay caused by data and the delay caused by cloud processing;

B. According to the video analysis task execution delay model in the cloud-edge environment, construct the optimal task deployment plan for the cloud-edge environment; and use the approximate algorithm based on Lyapunov algorithm and Markov approximation algorithm to select more energy-saving video analysis tasks The task deployment scheme includes the selection of edge machine specifications, cloud machine instance specifications, and task division scheme.

Further, the video analysis task execution delay model construction method in the step A includes the estimation of the calculation amount required for the analysis task and the calculation capacity estimation of the machine specification, and estimates the task processing delay through the calculation capacity and the calculation amount; the step B The task deployment scheme in is obtained by solving the optimal task deployment scheme problem in the cloud-edge environment, that is, a long-term optimization problem. The long-term optimization problem aims to minimize the long-term energy consumption of the system, and the constraint condition is the processing time Delay falls short of task requirements.

In the step A, when estimating the calculation amount of the analysis task, the calculation amount required for the analysis task is calculated according to the progress of the analysis task (such as the total amount of the deep neural network placed on the machine, or the neural network calculation amount placed on the machine). The total number of sub-units) for modeling; when estimating the computing power of machine specifications, the computing power of machine specifications is modeled as a function of CPU model and quantity, GPU model and quantity, memory size, and hard disk size; when performing real-time video analysis tasks When performing delay modeling, first select the benchmark machine specification, measure the benchmark machine specification and execute the complete real-time video analysis task as the benchmark time, then use the benchmark machine specification to execute each possible split point of the task, and measure each split The execution time before the point is used as the reference time of the split point, and the execution time of different task split methods of different machine specifications can be determined by the reference time of the split point and the computing power coefficients of different machine specifications and benchmark specifications.

In the step B, when constructing the optimal task deployment scheme for the cloud-edge environment, let the available machine specification sets of the edge end and the cloud be C ₁ and C ₂ respectively, and use c _1t and c _2t to represent the edge end and the cloud at time t respectively. The machine specifications selected by the cloud, U _t , u _t respectively represent the total calculation amount of the current real-time video analysis task and the calculation amount deployed at the edge. e ₁ , e ₂ , and e ₃ respectively represent the resources consumed by processing data and sending data at the edge, in the cloud, and the values will be affected by machine specifications and task processing time. w ₁ , w ₂ , and w ₃ respectively represent the time consumed to process data and send data at the edge, on the cloud, and the values will be affected by machine specifications and the amount of computation required for the task. B _t represents the available bandwidth at the current moment, L represents the maximum time delay that the task can accept, and f represents the compression rate of the intermediate data compression algorithm used by the system (for example, if it is transmitted in a video format, the code rate of the video stream coding algorithm can be used with The ratio of the size of the frame as the compression rate). Table 1 lists the symbols and their meanings mentioned in this article.

Table 1 Symbols and their meanings mentioned in this article

In the step B, when constructing the optimal task deployment scheme for the cloud-edge environment, the goal of constructing the optimal task deployment scheme for the cloud-edge environment is:

代表传输中间数据带来的时延，由于d_t与B_t分别代表数据量以及带宽，f代表压缩率，因此

即可以用来表示发送数据需要的时间；The optimization goal is to minimize long-term energy consumption, where e ₁ (c _1t , w ₁ (c _1t , u _t )) represents the energy consumed by the edge processing in a single time slot (in actual use, each time slot can be measured in minutes or hours), the parameter represents the resource e ₁ consumed by processing data is related to the machine specification c _1t and the machine processing time w ₁ (this is because the machine specification affects the power of the machine, and the processing time and machine power determine the resources consumed by processing), e ₂ (c _2t , w ₂ (c _2t , U _t -u _t )) represents the resources consumed by the cloud processing process in a single time slot, and the parameter represents that e ₂ is related to the machine specification c _2t and the machine processing time w ₂ ; further Since U _t represents the total calculation amount, and u _t represents the edge calculation amount, U _t -u _t represents the cloud computing amount; the parameters of w ₁ and w ₂ are the same, and the machine specifications affect the computing power, and the computing power and computing power are common determine processing time;

Represents the delay caused by the transmission of intermediate data, since d _t and B _t represent the amount of data and bandwidth respectively, and f represents the compression rate, so

That is, it can be used to indicate the time required to send data;

The constraints are:

代表传输中间数据需要花费的时间。in

Represents the time it takes to transfer intermediate data.

In the step B, when solving the problem of the optimal task deployment plan for the cloud-side environment, first use the Lyapunov algorithm to decouple the problem of the optimal task deployment plan for the cloud-side environment into subproblems in each time slot, and then use the Mark The Kove approximation algorithm finds the optimal solution in the solution space.

First, a virtual dynamic queue q _t is used to describe the change of the delay of analysis task processing. When the actual delay at the current moment exceeds the threshold, the queue length is shortened, otherwise the queue length is increased, and the queue length is always not less than 0. A virtual dynamic queue is defined as follows:

The decoupled sub-problem of the optimal task deployment scheme in the cloud-edge environment is, where V is an adjustment parameter, which is used to adjust the algorithm's tolerance for constraint violations:

c _1t , c _2t , u _t ∈ Z ⁺

Use the Markov approximation method to solve the optimal task deployment scheme sub-problem;

1) First, randomly assign c _1t , c _2t , and u _t to estimate the objective function brought by the analysis task at this time, and record this delay as

其中e为自然对数，以该概率接受此新选取的方案；2) Afterwards, randomly change the values of c _1t , c _2t , and u _t , calculate the objective function brought by the analysis task again, record this delay as o, and calculate the probability

Where e is the natural logarithm, accept this newly selected solution with this probability;

3) Repeat step 2) until the maximum number of times set by the program, or the program is not updated for 10 consecutive random programs during the execution process;

4) Deploy and execute the final solution, and obtain the actual edge execution time, cloud execution time, and data sending time during the execution process, which are recorded as

5) Calculate the virtual queue length in the next subproblem

The beneficial effects of the present invention are as follows:

An online algorithm based on Lyapunov optimization design selects the optimal cloud-side specification combination and deep neural network cutting scheme for real-time video stream analysis in a cloud-side collaborative environment. The system first uses the hill-climbing algorithm to build a service quality model in the case of linear time overhead. Use Markov approximation to find the optimal configuration of tasks in the 100,000-level configuration space to ensure the quality of service and minimize resource overhead. Then, the long-term optimization problem is decoupled into sub-problems at each time through Lyapunov optimization. At each moment, a better solution is obtained through the Markov approximation method. When the network delay and task load change, quickly and accurately adjust the task deployment method to minimize the energy consumption of the system.

The analysis task delay model construction method in the present invention can construct the analysis task delay model under the condition that the number of measurements does not exceed the cloud configuration, edge configuration, and task segmentation methods and the number of types is linearly accumulated, and has good sensitivity.

Through time complexity analysis and simulation comparison analysis, it can be seen that the present invention can effectively reduce the energy consumption of the system when the network time delay and the task load change.

Description of drawings

FIG. 1 is a schematic block diagram of a cloud-edge collaborative real-time video analysis task execution delay modeling and deployment method provided in this embodiment;

FIG. 2 is an illustration without loss of generality of the real-time video analysis task execution delay modeling process in the cloud-edge collaboration real-time video analysis task execution delay modeling and deployment method provided in this embodiment.

detailed description

The present invention will be further described in detail below in conjunction with the examples, which are not intended to limit the present invention.

In order to reduce the resource consumption of the system as much as possible when the load of the analysis task and the network environment change, the present invention provides a cloud-edge collaborative real-time video analysis task execution delay modeling and energy-saving deployment method. Firstly, a video analysis task execution delay model in the cloud-edge environment is constructed; the delay model refers to the additional delay introduced during the video analysis task process, including the delay caused by edge processing, and the intermediate data transmission between the edge and the cloud. The delay caused by the cloud processing and the delay caused by the cloud processing; then according to the video analysis task execution delay model in the cloud-edge environment, the optimal task deployment plan for the cloud-edge environment is constructed; and based on the Lyapunov algorithm and the Markov The approximate algorithm of the approximate algorithm selects a more energy-saving task deployment scheme for the video analysis task, and the task deployment scheme includes the selection of the edge machine specification, the cloud machine instance specification, and the task division scheme. In this way, the resources required for long-term execution of the system can be effectively reduced while satisfying the analysis task execution delay. As shown in Figure 1, the present embodiment specifically includes the following steps:

1) Build a video analysis task execution delay model in the cloud-edge environment;

The delay model is the additional delay introduced in the video analysis task process, including the delay caused by edge processing, the delay caused by the transmission of intermediate data between the edge and the cloud, and the delay caused by cloud processing;

2) Constructing the optimal task deployment scheme for the cloud-edge environment;

First describe the problem, and then define the objective function and various constraints.

3) Solve the problem of optimal task deployment scheme in cloud-edge environment;

A more energy-efficient task deployment scheme is selected for the video analysis task through an approximation algorithm based on the Lyapunov algorithm and the Markov approximation algorithm. The task deployment scheme includes edge machine specification selection, cloud machine instance specification, and task division scheme.

In this embodiment, constructing a video analysis task execution delay model in a cloud-edge environment includes the following steps:

4) Determining the amount of calculation required before processing the segmentation point for different video analysis tasks;

Fix the video analysis frame rate and video analysis input picture size to the highest frame rate and maximum input picture size, and use the benchmark machine specifications to measure the total processing time before the split point for different tasks.

5) Determining the amount of calculation required for different video analysis frame rates;

The task processing split point and the video analysis input image size are fixed to the first task processing split point and the maximum input image size, and the total processing time of different video analysis frame rates is measured using the benchmark machine specification.

6) Determining the calculation amount required for different video analysis input image sizes;

Fix the task processing split point and video analysis frame rate as the first task processing split point and maximum video analysis frame rate, and use the benchmark machine specifications to measure the total processing time of different video analysis input image sizes.

7) According to the measurement results of steps 4)-6), the different tasks of the benchmark machine are divided and the calculation amount coefficient table of the input scheme is constructed;

For any task split point, video analysis frame rate and video analysis input image size, the calculation coefficient is:

The processing delay of the analysis task of the benchmark machine at any task split point, video analysis frame rate, and video analysis input image size can be estimated as:

为步骤4)测定的不同分割点的处理时间；

为步骤5)测定的不同视频分析帧率的处理时间；其中

为步骤6)测定的不同视频分析输入图片大小的处理时间；

分别为最高帧率，输入图片大小，完整处理流程的时间以及最高帧率，最高输入图片大小，第一个分割点的时间。in

For step 4) the processing time of the different segmentation points measured;

Be the processing time of the different video analysis frame rates of step 5) measurement; Wherein

Analyze the processing time of the input picture size for the different videos measured in step 6);

They are the highest frame rate, the input image size, the time of the complete processing process and the highest frame rate, the highest input image size, and the time of the first split point.

8) Model the computing power of different machine specifications in the cloud and at the edge;

When estimating the computing power of the machine specification, the computing power of the machine specification is modeled as a function of the CPU model and quantity, GPU model and quantity, memory size, and hard disk size; when performing real-time video analysis task execution delay modeling, first select Determine the benchmark machine specification, measure the benchmark machine specification to execute the complete real-time video analysis task as the benchmark time, then use the benchmark machine specification to execute each possible split point of the task, and measure the execution time before each split point as the split point benchmark After that, the execution time of different task division methods of different machine specifications can be determined by the division point reference time and the computing power coefficients of different machine specifications and reference specifications.

In this embodiment, the problem of constructing the most task deployment solution in the cloud-edge environment includes the following steps:

9) Describe the problem;

In the above step B, when constructing the optimal task deployment solution for the cloud-edge environment, let the available machine specification sets of the edge end and the cloud be C ₁ and C ₂ respectively, and use c _1t and c _2t to represent the edge end at time t respectively. and the machine specification selected by the cloud, U _t , u _t respectively represent the total calculation amount of the current real-time video analysis task and the calculation amount deployed at the edge. e ₁ , e ₂ , and e ₃ respectively represent the resources consumed by processing data and sending data at the edge, in the cloud, and the values will be affected by machine specifications and task processing time. w ₁ , w ₂ , and w ₃ respectively represent the time consumed to process data and send data at the edge, on the cloud, and the values will be affected by machine specifications and the amount of computation required for the task. B _t represents the available bandwidth at the current moment, L represents the maximum time delay that the task can accept, and f represents the compression rate of the intermediate data compression algorithm used by the system (for example, if it is transmitted in a video format, the code rate of the video stream coding algorithm can be used with The ratio of the size of the frame as the compression rate).

10) Define the objective function;

The goal of the optimal task deployment scheme for the constructed cloud-edge environment is:

代表传输中间数据带来的时延，由于d_t与B_t分别代表数据量以及带宽，f代表压缩率，因此

即可以用来表示发送数据需要的时间；where e ₁ (c _1t , w ₁ (c _1t , u _t )) represents the energy consumed by the edge-end processing in a single time slot (in actual use, each time slot can be in minutes or hours), and the parameter represents the processing The resource e1 consumed by the data is related to the machine specification c _1t and the machine processing time w ₁ (this is because the machine specification affects the power of the machine, and the processing time and machine power determine the resources consumed by processing), e ₂ (c _2t , w ₂ (c _2t , U _t -u _t )) represents the resources consumed by the cloud processing process in a single time slot, and the parameter e ₂ is related to the machine specification c _2t and the machine processing time w ₂ ; the parameters of w ₁ and w ₂ are the same, and the machine The specification affects the computing power, and the computing power and computing power together determine the processing time; further, since U _t represents the total computing power, and u _t represents the edge computing power, U _t -u _t represents the cloud computing power;

Represents the delay caused by the transmission of intermediate data, since d _t and B _t represent the amount of data and bandwidth respectively, and f represents the compression rate, so

That is, it can be used to indicate the time required to send data;

11) Define each constraint condition;

The constraints are:

代表传输中间数据需要花费的时间。in

Represents the time it takes to transfer intermediate data.

In this embodiment, when solving the problem of the optimal task deployment scheme in the cloud-edge environment, the following steps are included:

12) Use Lyapunov algorithm to decouple the problem of optimal task deployment scheme in cloud-edge environment into sub-problems in each time slot;

A virtual dynamic queue q _t is used to describe the change of the delay of analysis task processing. When the actual delay at the current moment exceeds the threshold L, the queue length is shortened, otherwise the queue length is increased, and the queue length is always not less than 0. A virtual dynamic queue is defined as follows:

The decoupled sub-problem of the optimal task deployment scheme in the cloud-edge environment is, where V is an adjustment parameter, which is used to adjust the algorithm's tolerance for constraint violations:

c _1t , c _2t , u _t ∈ Z ⁺

Use the Markov approximation algorithm to find the optimal solution in the solution space;

1) First, randomly assign c _1t , c _2t , and u _t to estimate the objective function brought by the analysis task at this time, and record this delay as

其中e为自然对数，以该概率接受此新选取的方案；2) Afterwards, randomly change the values of c _1t , c _2t , and u _t , calculate the objective function brought by the analysis task again, record this delay as o, and calculate the probability

Where e is the natural logarithm, accept this newly selected solution with this probability;

3) Repeat step 2) until the maximum number of times set by the program, or the program is not updated for 10 consecutive random programs during the execution process;

4) Deploy and execute the final solution, and obtain the actual edge execution time, cloud execution time, and data sending time during the execution process, which are recorded as

5) Calculate the virtual queue length in the next subproblem

Figure 1 describes a typical system optimization process using the cloud-edge collaborative video analysis task deployment method.

The following is an example without loss of generality of the construction method of the cloud-side collaborative video analysis task delay model involved in this embodiment in combination with FIG. 2 :

According to the task deployment method provided in this embodiment, it is assumed that at a certain moment there is a video analysis task that needs to be deployed cloud-side collaboratively. The frame rate of the mode and the input image size are different. The variable set of frame rate is {1,2,3,5,10} (referring to processing 1 frame, 2 frames, etc. per second), and the variable set of image input size during execution is {416*416,512*512,608*608}( 416, 512, 606 are the three standard input image sizes of the target recognition neural network yolov3).

First, build a video analysis task execution delay model in the cloud-edge environment. Specifically, it is divided into the following steps. 1. Measure the amount of calculation required before processing the segmentation point for different video analysis tasks;

The video analysis frame rate and video analysis input picture size are fixed at a frame rate of 10 and an input picture size of 608, and the total processing time before different task processing split points (ie, from 1 to 5) is measured using the benchmark machine specification.

2. Determine the amount of calculation required for different video analysis frame rates;

Fix the task processing split point and video analysis input image size to the first task processing split point and 608, and use the benchmark machine specifications to analyze the total Processing time is measured.

3. Determine the amount of calculation required for different video analysis input image sizes;

Fix the task processing split point and video analysis frame rate as the first task processing split point and frame rate 10, and use the benchmark machine specifications to analyze the total processing time of different video input image sizes (ie 416*416, 512*512, 608*608) Take measurements.

According to the measurement results of steps 1-3, construct the different task division of the benchmark machine and the calculation amount coefficient table of the input scheme;

For any task split point, video analysis frame rate and video analysis input image size, the calculation coefficient is:

The processing delay of the analysis task of the benchmark machine at any task split point, video analysis frame rate, and video analysis input image size can be estimated as:

The meaning of each symbol is as described above.

There are many specific application ways of the present invention, and the above-mentioned method, especially the adjustment mode of the redundancy rate, is only a preferred embodiment of the present invention. It should be pointed out that the above-mentioned examples do not limit the present invention, and relevant workers do not deviate from the scope of the technical idea of the present invention Various changes and modifications made within the scope of the present invention all fall within the protection scope of the present invention.

Claims (8)

1. A cloud-side collaborative real-time video analysis task execution delay modeling and deployment method, is characterized in that, comprises the following steps: A. The construction method of the video analysis task execution delay model in the cloud-edge environment; the delay model is the additional delay introduced in the process of the video analysis task, including the delay caused by the processing at the edge, and the transmission between the edge and the cloud The delay caused by data and the delay caused by cloud processing; B. According to the video analysis task execution delay model in the cloud-edge environment, construct the optimal task deployment plan for the cloud-edge environment; and select a more energy-efficient task deployment for video analysis tasks based on the Lyapunov algorithm and the Markov approximation algorithm The task deployment scheme includes the selection of edge machine specifications, cloud machine instance specifications, and task division schemes. 2. The real-time video analysis task execution delay modeling and deployment method according to claim 1, characterized in that: in the step A, the video analysis task execution delay model construction method includes the calculation amount estimation required for the analysis task and Estimate the computing power of the machine specification, and estimate the task processing delay through the computing power and the amount of calculation; in the step B, the task deployment plan is obtained by solving the problem of the optimal task deployment plan in the cloud-edge environment, that is, a long-term optimization problem, the long-term The goal of the optimization problem is to minimize the long-term energy consumption of the system, and the constraint condition is that the processing delay of the system at each moment is lower than the task requirement. 3. The real-time video analysis task execution delay modeling and deployment method according to claim 1, characterized in that: in the step A, when the analysis task calculation amount is estimated, the calculation amount required for the analysis task is calculated according to the analysis Model the task progress; when estimating the computing power of the machine specification, model the computing power of the machine specification as a function of the CPU model and quantity, GPU model and quantity, memory size, and hard disk size; when performing real-time video analysis task execution delay When modeling, first select the benchmark machine specification, measure the benchmark machine specification and execute the complete real-time video analysis task as the benchmark time, then use the benchmark machine specification to execute each possible split point of the task, and measure the time before each split point The execution time is used as the reference time of the segmentation point, and the execution time of different task segmentation methods of different machine specifications can be determined by the reference time of the segmentation point and the computing power coefficients of different machine specifications and reference specifications. 4. The real-time video analysis task execution delay modeling and deployment method according to claim 1, characterized in that: in the step B, when constructing the optimal task deployment solution for the cloud-side environment, the edge terminal and the cloud The set of available machine specifications are C ₁ , C ₂ , respectively, c _1t , c _2t represent the machine specifications selected at the edge and cloud at time t, respectively, U _t , u _t represent the total calculation amount and The amount of computing deployed at the edge; e ₁ , e ₂ , and e ₃ respectively represent the resources consumed by processing data and sending data on the edge at the cloud, and this value will be affected by the comprehensive impact of machine specifications and task processing time; w ₁ , w ₂ and w ₃ respectively represent the time spent on processing data and sending data at the edge, in the cloud, and this value will be affected by machine specifications and the amount of calculation required by the task; B _t represents the available bandwidth at the current moment, and L represents the task execution performance The maximum time delay accepted, f represents the compression rate of the intermediate data compression algorithm used by the system; d _t represents the amount of data that needs to be transmitted at the current moment. 5. The real-time video analysis task execution delay modeling and deployment method according to claim 4, characterized in that: in the step B, when constructing the cloud edge environment optimal task deployment plan problem, the constructed cloud edge environment The objective of the optimal task deployment problem is:

The optimization goal is to minimize long-term energy consumption, where e ₁ (c _1t , w ₁ (c _1t , u _t )) represents the energy consumed by the edge processing in a single time slot, and the parameter represents the resources e ₁ and The machine specification c _1t is related to the machine processing time w ₁ , e ₂ (c _2t , w ₂ (c _2t , U _t -u _t )) represents the resources consumed by the cloud processing process in a single time slot, and the parameter represents e ₂ and the machine The specification c _2t is related to the machine processing time w ₂ ; further, because U _t represents the total calculation amount, and u _t represents the edge calculation amount, so U _t -u _t represents the cloud computing amount; the parameters of w ₁ and w ₂ are the same, Machine specifications affect the computing power, and the amount of computing and computing power together determine the processing time;

Represents the delay caused by the transmission of intermediate data, since d _t and B _t represent the amount of data and bandwidth respectively, and f represents the compression rate, so

That is, it can be used to indicate the time required to send data; The constraints are:

in

Represents the time it takes to transfer intermediate data. 6. The real-time video analysis task execution delay modeling and deployment method according to claim 5, characterized in that: in the step B, when solving the problem of the optimal task deployment plan for the cloud-side environment, first use Lyapunov The algorithm decouples the problem of the optimal task deployment scheme in the cloud-edge environment into sub-problems in each time slot, and then uses the Markov approximation algorithm to find the optimal solution in the solution space. 7. The real-time video analysis task execution delay modeling and deployment method according to claim 6, characterized in that: the solution algorithm for the problem of optimal task deployment scheme in cloud-side environment is as follows: First, a virtual dynamic queue q _t is used to describe the change of the delay of analysis task processing. When the actual delay at the current moment exceeds the threshold, the queue length is shortened, otherwise the queue length increases, and the queue length is always not less than 0; the virtual dynamic queue is defined as follows:

The decoupled sub-problem of the optimal task deployment scheme in the cloud-edge environment is, where V is an adjustment parameter, which is used to adjust the algorithm's tolerance for constraint violations:

c _1t , c _2t , u _t ∈ Z ⁺ . 8. The real-time video analysis task execution delay modeling and deployment method according to claim 6, characterized in that: using the Markov approximation method to solve the optimal task deployment scheme sub-problem, specifically as follows: 1) First, randomly assign c _1t , c _2t , and u _t to estimate the objective function brought by the analysis task at this time, and record this delay as

2) Afterwards, randomly change the values of c _1t , c _2t , and u _t , calculate the objective function brought by the analysis task again, record this delay as o, and calculate the probability

Where e is the natural logarithm, accept this newly selected solution with this probability; 3) Repeat step 2) until the maximum number of times set by the program, or the program is not updated for 10 consecutive random programs during the execution process; 4) Deploy and execute the final solution, and obtain the actual edge execution time, cloud execution time, and data sending time during the execution process, which are recorded as

5) Calculate the virtual queue length in the next subproblem

Images