Cloud-edge collaborative real-time video analysis task execution delay modeling and deployment method

Abstract

The application discloses a cloud-edge collaborative real-time video analysis task execution delay modeling and deployment method, comprising: a video analysis task execution delay model construction method under a cloud-edge environment; the delay model is an additional delay introduced in a video analysis task process, comprising a delay caused by edge end processing, a delay caused by intermediate data transmission between the edge and the cloud, and a delay caused by cloud processing; according to the video analysis task execution delay model under the cloud-edge environment, an approximate algorithm based on the Lyapunov algorithm and the Markov approximation algorithm is used to select a more energy-saving task deployment scheme for the video analysis task, wherein the task deployment scheme comprises edge end machine specification selection, cloud machine instance specification and a task division scheme. The application can quickly adjust the task deployment scheme when the task load of the video analysis task and the cloud-edge network environment change, thereby effectively reducing the long-term operation energy consumption of the system.

Description

Technical Field

[0001] This invention relates to the field of edge computing, and to a cloud-edge collaborative system energy-saving deployment method, specifically a cloud-edge collaborative real-time video analysis task execution latency modeling and deployment method. Background Technology

[0002] With the development of mobile internet, a large number of highly interactive applications have emerged, such as the recently popular AR applications and various facial and license plate recognition applications. For such highly interactive applications, users often have high latency requirements. The transmission latency from the terminal to the data center is often high due to the need for multiple layers of network forwarding, including base stations and backbone networks. The transmission of large amounts of high-quality video also puts significant bandwidth pressure on the wide area network (WAN). To reduce access latency, these applications have introduced edge computing. Therefore, edge computing is gradually becoming a new network infrastructure in the mobile internet era. Edge computing reduces task response latency by deploying services closer to users, such as in urban backbone networks, base stations, or edge devices, while also reducing the enormous pressure on WAN bandwidth caused by raw data transmission. Furthermore, offloading services from the data center to the user edge avoids sending raw data to the cloud, thereby improving user data privacy.

[0003] Real-time video streaming and real-time video surveillance are new video applications that have emerged with the development of technologies such as CDN and edge computing. They are widely used in intelligent systems such as autonomous driving and disaster monitoring. These applications require real-time video streaming with low latency.

[0004] When real-time video stream analysis tasks have long video processing pipelines, large neural network models used for video analysis, or poor computing power at the edge, it is difficult to meet the computing power requirements by deploying real-time video stream analysis tasks solely at the edge. On the other hand, the traffic overhead and transmission latency caused by placing them entirely in the cloud are unacceptable. Therefore, cloud-edge collaborative deployment of real-time video stream analysis tasks becomes a more reasonable deployment solution.

[0005] Choosing a better deployment solution is crucial in a real production environment. A poor solution will not only result in higher power consumption or expenses, but also longer task processing time.

[0006] Current mainstream cloud service providers, such as Amazon Web Services, Azure, and Alibaba Cloud, offer a large number of cloud-side configuration options, sometimes reaching thousands. Similarly, there are multiple edge configurations available. Furthermore, for large neural networks or long processing pipelines, the number of strategies that can be used to divide the cloud and edge can reach dozens or even hundreds, resulting in a large total number of cloud-edge deployment methods. To determine the additional latency introduced by a specific deployment method when executing a particular task, it is necessary to actually run the corresponding deployment method. Therefore, measuring the service quality of each deployment method is extremely costly. Measuring all deployment methods cannot meet the requirements for rapid task deployment and execution. Thus, selecting the optimal deployment method for the current task from a large number of available deployment methods, whose actual performance can vary, is quite difficult. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for modeling and deploying the execution latency of real-time video analytics tasks in a cloud-edge collaborative manner.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] A method for latency modeling and energy-saving deployment of real-time video analytics tasks in a cloud-edge collaborative manner includes the following steps:

[0010] A. A method for constructing a video analysis task execution latency model in a cloud-edge environment; the latency model refers to the additional latency introduced during the video analysis task process, including latency caused by edge processing, latency caused by the transmission of intermediate data between the edge and the cloud, and latency caused by cloud processing.

[0011] B. Based on the video analysis task execution latency model in the cloud-edge environment, construct the optimal task deployment scheme problem in the cloud-edge environment; and select a more energy-efficient task deployment scheme for video analysis tasks through an approximation algorithm 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 partitioning scheme.

[0012] Furthermore, the video analysis task execution latency model construction method in step A includes estimating the computational load required for the analysis task and estimating the machine's computational capabilities. The task processing latency is estimated by considering the computational capabilities and computational load. The task deployment scheme in step B is obtained by solving the optimal task deployment scheme problem in the cloud-edge environment, which is a long-term optimization problem. This long-term optimization problem aims to minimize the long-term energy consumption of the system, with the constraint that the system's processing latency at each moment is lower than the task requirements.

[0013] In step A, when estimating the computational load of the analysis task, the computational load required for the analysis task is modeled according to the progress of the analysis task (such as the total number of deep neural networks placed on the machine, or the total number of neural network operators placed on the machine); when estimating the machine specification's computational capability, the machine specification's computational capability is modeled as a function of CPU model and quantity, GPU model and quantity, memory size, and hard disk size; when modeling the execution latency of the real-time video analysis task, a benchmark machine specification is first selected, and the time taken by the benchmark machine specification to execute a complete real-time video analysis task is measured as the benchmark time. Then, the benchmark machine specification is used to execute each possible segmentation point of the task, and the execution time before each segmentation point is measured as the segmentation point benchmark time. Subsequently, the execution time of different task segmentation methods for different machine specifications can be determined by the segmentation point benchmark time and the computational capability coefficient between different machine specifications and the benchmark specification.

[0014] In step B, when constructing the optimal task deployment scheme for the cloud-edge environment, let the available machine specification sets at the edge and in the cloud be C1 and C2, respectively, and use c 1t c 2t U represents the machine specifications selected at the edge and in the cloud at time t, respectively. t u t These represent the total computational load of the current real-time video analytics task and the computational load deployed at the edge, respectively. e1, e2, and e3 represent the resources consumed in processing and transmitting data at the edge and in the cloud, respectively; these values ​​are affected by machine specifications and task processing time. w1, w2, and w3 represent the time consumed in processing and transmitting data at the edge and in the cloud, respectively; these values ​​are affected by machine specifications and the computational load required by the task. B t The current available bandwidth is represented by L, the maximum acceptable latency for task execution is represented by f, and the compression ratio of the intermediate data compression algorithm used by the system is represented by f (for example, if transmitted in a video format, the ratio of the bitrate of the video stream coding algorithm to the frame size can be used as the compression ratio). Table 1 lists the symbols mentioned in this paper and their meanings.

[0015] Table 1. Symbols mentioned in this article and their meanings.

[0016]

[0017]

[0018] In step B, when constructing the optimal task deployment scheme for the cloud-edge environment, the objective of the constructed optimal task deployment scheme for the cloud-edge environment is:

[0019]

[0020] The optimization objective is to minimize long-term energy consumption, where e1(c 1t w1(c 1t u t The ) represents the energy consumed by the edge processing process in a single time slot (in actual use, each time slot can be measured in minutes or hours), and the parameter represents the resources e1 consumed in processing data and the machine specifications c. 1t And related to the machine processing time w1 (this is because the machine specifications affect the machine's power, and the processing time and machine power determine the resources consumed in processing), e2(c 2t w2(c 2t U t -u t ) represents the resources consumed by the cloud processing process in a single time slot, and the parameter represents e2 and machine specification c. 2t And related to machine processing time w2; further due to U t u represents the total computational cost. t U represents the computational cost at the edge, therefore t -u t This represents the amount of cloud computing power; the parameters for w1 and w2 are similar. Machine specifications affect computing power, and the amount of computing power and computing power together determine the processing time. This represents the latency caused by the transmission of intermediate data, due to d t With B t These represent data volume and bandwidth, respectively, and f represents the compression ratio. It can be used to represent the time required to send data;

[0021] The constraints are:

[0022]

[0023]

[0024] in This represents the time required to transmit intermediate data.

[0025] In step B, when solving the optimal task deployment scheme problem in the cloud-edge environment, the Lyapunov algorithm is first used to decouple the optimal task deployment scheme problem in the cloud-edge environment into subproblems in each time slot, and then the Markov approximation algorithm is used to find the optimal solution in the solution space.

[0026] First, a virtual dynamic queue q is used. t The description analyzes the changes in task processing latency. When the actual latency at the current moment exceeds a threshold, the queue length is shortened; conversely, the queue length is increased, and the queue length is never less than 0. The definition of a virtual dynamic queue is as follows:

[0027]

[0028] The subproblem of optimal task deployment in the decoupled cloud-edge environment is: where V is an adjustment parameter used to adjust the algorithm's tolerance for constraint violations.

[0029]

[0030]

[0031]

[0032]

[0033] c 1t c 2t u t ∈Z +

[0034] The subproblem of optimal task deployment scheme is solved using the Markov approximation method;

[0035] 1) First, c 1t c 2t u t All values ​​are randomly assigned, and the objective function resulting from the analysis task at this point is estimated. This point is then denoted as...

[0036] 2) Then randomly change c 1t c 2t u t The value of is used to recalculate the objective function brought about by the analysis task, and this delay is denoted as 0. The probability is then calculated. Where e is the natural logarithm, with which the newly selected option is accepted;

[0037] 3) Repeat step 2) until the maximum number of times set by the program is reached, or until 10 consecutive random schemes are executed without updating the scheme;

[0038] 4) Deploy and execute the final solution, and during execution, obtain the actual edge execution time, cloud execution time, and data transmission time, which are denoted as follows:

[0039] 5) Calculate the length of the virtual queue in the next subproblem.

[0040] The beneficial effects of this invention are as follows:

[0041] Based on Lyapunov optimization, an online algorithm is designed to select the optimal combination of cloud and edge specifications and a deep neural network segmentation scheme for real-time video stream analysis in a cloud-edge collaborative environment. The system first uses a hill-climbing algorithm to construct a quality of service (QoS) model with linear time overhead. Then, using Markov approximation, it searches for the optimal task configuration in a configuration space of hundreds of thousands, ensuring QoS while minimizing resource consumption. Next, Lyapunov optimization decouples the long-term optimization problem into sub-problems at each time step. At each time step, a near-optimal solution is found using Markov approximation. When network latency and task load change, the system quickly and accurately adjusts the task deployment method, minimizing system energy consumption.

[0042] The analysis task latency model construction method in this invention can construct an analysis task latency model with a number of measurements not exceeding the linear sum of the number of cloud configurations, edge configurations, and task segmentation methods, and has good sensitivity.

[0043] Time complexity analysis and simulation comparison show that the present invention can effectively reduce the energy consumption of the system when network latency and task load change. Attached Figure Description

[0044] Figure 1 This is a schematic block diagram of the cloud-edge collaborative real-time video analysis task execution latency modeling and deployment method provided in this embodiment;

[0045] Figure 2 This is a non-loss-of-generalization example illustrating the real-time video analytics task execution latency modeling process in the cloud-edge collaborative real-time video analytics task execution latency modeling and deployment method provided in this embodiment. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to the embodiments. These embodiments do not constitute a limitation on the present invention.

[0047] To minimize system resource consumption when the load of analysis tasks and the network environment change, this invention provides a cloud-edge collaborative real-time video analysis task execution latency modeling and energy-saving deployment method. First, a video analysis task execution latency model is constructed in a cloud-edge environment. This latency model represents the additional latency introduced during the video analysis task, including latency from edge processing, latency from data transmission between the edge and the cloud, and latency from cloud processing. Then, based on the video analysis task execution latency model in the cloud-edge environment, an optimal task deployment scheme problem is constructed. Finally, an approximation algorithm based on the Lyapunov algorithm and Markov approximation algorithm is used to select a more energy-efficient task deployment scheme for the video analysis task. This task deployment scheme includes the selection of edge machine specifications, cloud machine instance specifications, and task partitioning scheme. This effectively reduces the resources required for long-term system execution while meeting the latency requirements of the analysis task. In this embodiment, as shown... Figure 1 The process includes the following steps:

[0048] 1) Construct a latency model for video analytics tasks in a cloud-edge environment;

[0049] The latency model refers to the additional latency introduced during the video analysis task, including latency caused by edge processing, latency caused by the transmission of intermediate data between the edge and the cloud, and latency caused by cloud processing.

[0050] 2) The problem of constructing the optimal task deployment scheme in the cloud-edge environment;

[0051] First, describe the problem, then define the objective function and various constraints.

[0052] 3) Solve the problem of optimal task deployment scheme in cloud-edge environment;

[0053] An approximation algorithm based on Lyapunov and Markov approximation is used to select a more energy-efficient task deployment scheme for video analysis tasks. The task deployment scheme includes the selection of edge machine specifications, cloud machine instance specifications, and task partitioning scheme.

[0054] In this embodiment, constructing a video analytics task execution latency model in a cloud-edge environment includes the following steps:

[0055] 4) Determine the computational load required before processing segmentation points for different video analysis tasks;

[0056] The video analysis frame rate and the input image size were fixed at the highest frame rate and the maximum input image size, and the total processing time before different task processing split points was measured using benchmark machine specifications.

[0057] 5) Determine the computational load required for different video analysis frame rates;

[0058] The task processing split point and the input image size for video analysis were fixed at the first task processing split point and the maximum input image size. The total processing time for different video analysis frame rates was measured using benchmark machine specifications.

[0059] 6) Determine the computational load required for different input image sizes in video analysis;

[0060] The task processing split point and video analysis frame rate were fixed at the first task processing split point and the maximum video analysis frame rate, and the total processing time for different video analysis input image sizes was measured using benchmark machine specifications.

[0061] 7) Based on the measurement results of steps 4)-6), construct a table of computational coefficients for different task partitioning and input schemes of the benchmark machine;

[0062] For any task segmentation point, the computational complexity coefficients for the video analysis frame rate and the input image size are:

[0063]

[0064] The processing latency of the benchmark machine for the analysis task at any task segmentation point, video analysis frame rate, and input image size can be estimated as follows:

[0065]

[0066] in The processing time for different segmentation points determined in step 4); The processing time for different video analysis frame rates determined in step 5); where The processing time for different video analysis input image sizes was measured in step 6); These represent the maximum frame rate, input image size, total processing time, and the time for the first segmentation point, respectively.

[0067] 8) Model the computing capabilities of different machine specifications in the cloud and at the edge;

[0068] 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 modeling the execution latency of real-time video analysis tasks, a baseline machine specification is first selected, and the execution time of the baseline machine specification for a complete real-time video analysis task is measured as the baseline time. Then, the baseline machine specification is used to execute the task at each possible segmentation point, and the execution time before each segmentation point is measured as the segmentation point baseline time. Subsequently, the execution time of different task segmentation methods for different machine specifications can be determined by the segmentation point baseline time and the computing power coefficient between different machine specifications and the baseline specification.

[0069] In this embodiment, the problem of constructing the optimal task deployment scheme in a cloud-edge environment includes the following steps:

[0070] 9) Describe the problem;

[0071] In step B, when constructing the optimal task deployment scheme for the cloud-edge environment, let the available machine specification sets at the edge and in the cloud be C1 and C2, respectively, and use c 1t c 2t U represents the machine specifications selected at the edge and in the cloud at time t, respectively. t u t These represent the total computational load of the current real-time video analytics task and the computational load deployed at the edge, respectively. e1, e2, and e3 represent the resources consumed in processing and transmitting data at the edge and in the cloud, respectively; these values ​​are affected by machine specifications and task processing time. w1, w2, and w3 represent the time consumed in processing and transmitting data at the edge and in the cloud, respectively; these values ​​are affected by machine specifications and the computational load required by the task. B t L represents the available bandwidth at the current moment, L represents the maximum acceptable latency for task execution, and f represents the compression ratio of the intermediate data compression algorithm used by the system (for example, if transmitted in a video format, the ratio of the bitrate of the video stream coding algorithm to the frame size can be used as the compression ratio).

[0072] 10) Define the objective function;

[0073] The objective of constructing the optimal task deployment scheme in the cloud-edge environment is:

[0074]

[0075] Where e1(c 1t w1(c 1t u t The ) represents the energy consumed by the edge processing process in a single time slot (in actual use, each time slot can be measured in minutes or hours), and the parameter represents the resources e1 consumed in processing data and the machine specifications c.1t And related to the machine processing time w1 (this is because the machine specifications affect the machine's power, and the processing time and machine power determine the resources consumed in processing), e2(c 2t w2(c 2t U t -u t ) represents the resources consumed by the cloud processing process in a single time slot, and the parameter represents e2 and machine specification c. 2t And it is related to the machine processing time w2; the parameters w1 and w2 are similar, the machine specifications affect the computing power, and the amount of computation and computing power together determine the processing time; furthermore, due to U t u represents the total computational cost. t U represents the computational cost at the edge, therefore t -u t Represents cloud computing volume; This represents the latency caused by the transmission of intermediate data, due to d t With B t These represent data volume and bandwidth, respectively, and f represents the compression ratio. It can be used to represent the time required to send data;

[0076] 11) Define the various constraints;

[0077] The constraints are:

[0078]

[0079]

[0080] in This represents the time required to transmit intermediate data.

[0081] In this embodiment, solving the problem of optimal task deployment scheme in the cloud-edge environment includes the following steps:

[0082] 12) The Lyapunov algorithm is used to decouple the optimal task deployment scheme problem in the cloud-edge environment into sub-problems in each time slot;

[0083] Use a virtual dynamic queue q t The description analyzes the changes in task processing latency. When the actual latency at the current moment exceeds a threshold L, the queue length is shortened; conversely, the queue length is increased, and the queue length is never less than 0. The definition of a virtual dynamic queue is as follows:

[0084]

[0085] The subproblem of optimal task deployment in the decoupled cloud-edge environment is: where V is an adjustment parameter used to adjust the algorithm's tolerance for constraint violations.

[0086]

[0087]

[0088]

[0089]

[0090] c 1t c 2t u t ∈Z +

[0091] The Markov approximation algorithm is used to find the optimal solution in the solution space;

[0092] 1) First, c 1t c 2t u t All values ​​are randomly assigned, and the objective function resulting from the analysis task at this point is estimated. This point is then denoted as...

[0093] 2) Then randomly change c 1t c 2t u t The value of is used to recalculate the objective function brought about by the analysis task, and this delay is denoted as 0. The probability is then calculated. Where e is the natural logarithm, with which the newly selected option is accepted;

[0094] 3) Repeat step 2) until the maximum number of times set by the program is reached, or until 10 consecutive random schemes are executed without updating the scheme;

[0095] 4) Deploy and execute the final solution, and during execution, obtain the actual edge execution time, cloud execution time, and data transmission time, which are denoted as follows:

[0096] 5) Calculate the length of the virtual queue in the next subproblem.

[0097] Figure 1 This paper describes a typical process of system optimization using a cloud-edge collaborative video analytics task deployment method.

[0098] The following is combined with Figure 2 The latency model construction method for cloud-edge collaborative video analytics tasks involved in this embodiment is illustrated with an example without loss of generality:

[0099] According to the task deployment method provided in this embodiment, assuming that a video analysis task needs to be deployed collaboratively between cloud and edge at a certain moment, the task has 5 divisible segmentation points. The task can have multiple running modes during execution, and these modes have different frame rates and input image sizes. The variable set of frame rates is {1, 2, 3, 5, 10} (referring to processing 1 frame per second, 2 frames per second, etc.), and the variable set of input image sizes during execution is {416*416, 512*512, 608*608} (416, 512, 606 are the three standard input image sizes for the YOLOv3 object recognition neural network).

[0100] First, we construct a video analytics task execution latency model in a cloud-edge environment. Specifically, this involves the following steps: 1. Determine the computational load required before different video analytics task processing segmentation points.

[0101] With the video analytics frame rate and input image size fixed at 10 frames per second and 608 pixels per image, the total processing time before different task processing split points (i.e., from 1 to 5) was measured using benchmark machine specifications.

[0102] 2. Determine the computational load required for different video analysis frame rates;

[0103] The task processing split point and the video analysis input image size were fixed at the first task processing split point and 608, and the total processing time for different video analysis frame rates (i.e., {1,2,3,5,10}) was measured using the baseline machine specifications.

[0104] 3. Determine the computational load required for different input image sizes in video analysis;

[0105] The task processing split point and video analysis frame rate were fixed at the first task processing split point and a frame rate of 10. The total processing time for different video analysis input image sizes (i.e., 416*416, 512*512, 608*608) was measured using the benchmark machine specifications.

[0106] Based on the measurement results of steps 1-3, construct a table of computational coefficients for different task partitioning and input schemes of the benchmark machine;

[0107] For any task segmentation point, the computational complexity coefficients for the video analysis frame rate and the input image size are:

[0108]

[0109] The processing latency of the benchmark machine for the analysis task at any task segmentation point, video analysis frame rate, and input image size can be estimated as follows:

[0110]

[0111] The meanings of each symbol are as described above.

[0112] This invention has many specific applications. The methods described above, especially the redundancy rate adjustment method, are only preferred embodiments of this invention. It should be noted that the above embodiments do not constitute a limitation on this invention. Various changes and modifications made by relevant personnel without departing from the technical concept of this invention all fall within the protection scope of this invention.

Claims

1. A method for modeling and deploying the execution latency of real-time video analytics tasks in a cloud-edge collaborative manner, characterized in that, Comprising the following steps: A. A video analysis task execution delay model construction method under a cloud-edge environment; the delay model is an additional delay introduced in the process of a video analysis task, 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; B. According to the video analysis task execution delay model under the cloud-edge environment, a cloud-edge environment optimal task deployment scheme problem is constructed; and a more energy-saving task deployment scheme is selected for the video analysis task based on the Lyapunov algorithm and the Markov approximation algorithm, the task deployment scheme including edge machine specification selection, cloud machine instance specification, and task partitioning scheme; The step B, in constructing the cloud edge environment optimal task deployment scheme problem, let the available machine specification set of edge and cloud be c1, c2, use c 1t , c 2t Respectively represent the selected machine specification of edge and cloud at t time, U t , u t Respectively represent the total computing amount of current real-time video analysis task and the computing amount deployed in edge; e1, e2, e3 respectively represent the resource consumed by processing data and sending data in edge, in cloud, the value will be affected by machine specification and task processing time; w1, w2, w3 respectively represent the time consumed by processing data and sending data in edge, in cloud, the value will be affected by machine specification and task required computing amount; B t Represent the available bandwidth at current time, L represents the maximum time delay that task execution can accept, f represents the compression rate of intermediate data compression algorithm used by the system; d t Represent the data amount that needs to be transmitted at current time; In step B, when constructing the cloud-edge environment optimal task deployment scheme problem, the goal of the constructed cloud-edge environment optimal task deployment scheme problem is: The optimization goal is to minimize the long-term energy consumption, where e1(c 1t , w1(c 1t , u t )) represents the energy consumed by the edge processing in a single time slot, and the parameter e1 is related to the machine specification c 1t and the machine processing time w1, e2(c 2t , w2(c 2t , U t -u t )) represents the energy consumed by the cloud processing in a single time slot, and the parameter e2 is related to the machine specification c 2t and the machine processing time w2; further, since U t represents the total computation amount, u t represents the edge computation amount, and thus U t -u t represents the cloud computation amount; the parameters of w1 and w2 are the same, the machine specification affects the computation ability, and the computation amount and the computation ability jointly determine the processing time; represents the time delay caused by the transmission of intermediate data, since d t and B t represent the data amount and the bandwidth respectively, and f represents the compression rate, thus , which can be used to represent the time required for sending data; The constraint condition is: wherein represents the time spent in transmitting the intermediate data; In step B, when solving the cloud-edge environment optimal task deployment scheme problem, first decouple the cloud-edge environment optimal task deployment scheme problem into a sub-problem in each time slot using the Lyapunov algorithm, and then find the optimal solution in the solution space using the Markov approximation algorithm; The solution algorithm of the cloud-edge environment optimal task deployment scheme problem is as follows: First, a virtual dynamic queue q is used t The change of the time delay of the analysis task processing is described. When the actual time delay at the current time exceeds a threshold, the queue length is shortened, and vice versa. The queue length is always not less than 0. The virtual dynamic queue is defined as follows: The decoupled cloud-edge environment optimal task deployment scheme sub-problem is where V is a tuning parameter used to adjust the tolerance of the algorithm for constraint violations: c 1t , c 2t , u t ∈ Z + .

2. The real-time video analytics task execution latency modeling and deployment method of claim 1, wherein: In step A, the video analysis task execution delay model construction method includes analysis task computation amount estimation and machine specification computing capacity estimation, which estimates the task processing delay by computing capacity and computation amount; in step B, the task deployment scheme is obtained by solving the cloud-edge environment optimal task deployment scheme problem, which is a long-term optimization problem with the goal of minimizing the long-term energy consumption of the system, and the constraint condition is that the processing delay of the system at each time is lower than the task requirement. 3.The real-time video analytics task execution latency modeling and deployment method of claim 1, wherein: In step A, when performing analysis task computation amount estimation, the analysis task computation amount is modeled according to the analysis task progress; when performing machine specification computing capacity estimation, the machine specification computing capacity 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 task execution delay modeling, first select a reference machine specification, measure the execution time of the complete real-time video analysis task as the reference time, then use the reference machine specification to perform each possible partition point of the task, measure the execution time before each partition point as the partition point reference time, and then the execution time of different task partitioning methods of different machine specifications can be determined by the partition point reference time and the computing capacity coefficient of different machine specifications and the reference specification.

4. The real-time video analytics task execution latency modeling and deployment method of claim 1, wherein: The optimal task deployment scheme sub-problem is solved using the Markov approximation method, as follows: 1) First, c 1t , c 2t , u t are randomly assigned, and the objective function brought by the analysis task at this time is estimated, and the delay at this time is denoted as 2) then randomly change c 1t , c 2t , u t , and again compute the objective function for the analysis task, call this value o, and compute the probability where e is the natural logarithm, and accept this new choice with this probability. 3) Repeat step 2) until the maximum number of times set by the program, or the random scheme does not update the scheme continuously for 10 times during execution; 4) Deploy the final solution and get the actual edge execution time, cloud execution time and data transmission time during the execution, denoted as calculating the virtual queue length in the next sub-problem

Images