An efficient distributed algorithm task scheduling method

Abstract

The present invention discloses an efficient distributed algorithm task scheduling method, which realizes real-time frame-level processing of tens of thousands of video streams through key technologies such as parallel decoding of video streams, dynamic load balancing, and multi-model reasoning service orchestration; adopts a multimodal fusion detection algorithm to improve the recognition accuracy of complex scenes, and combines with a rule engine to realize alarm event classification, providing highly reliable and low-latency intelligent decision support for scenarios such as smart cities and industrial safety. Compared with traditional architectures, it has significant advantages and can effectively solve the key pain points in the current field of video intelligent analysis. This system adopts a distributed architecture design, supports horizontal expansion and dynamic resource scheduling, and can easily realize parallel processing of tens of thousands of video streams. Through an efficient load balancing mechanism and containerized deployment, the system can flexibly allocate computing resources according to actual business needs, significantly improving the overall throughput and processing efficiency, while reducing the risk of single point failures.

Description

Technical Field

[0001] The present invention relates to the field of video processing technology, and in particular to an efficient distributed algorithm task scheduling method. Background Art

[0002] With the rapid adoption of video surveillance technology and the advancement of smart city development, massive amounts of video data have become a crucial information carrier for public safety, industrial monitoring, traffic management, and other fields. Traditional manual monitoring models are limited by low processing efficiency and poor real-time performance, making them incapable of handling the real-time analysis of multiple concurrent video streams. Furthermore, emergencies such as fire warnings and area intrusions are often sudden and destructive, necessitating the rapid identification and response of intelligent methods.

[0003] The core challenges facing current technology are mainly reflected in two aspects: First, traditional stand-alone architectures face significant bottlenecks when dealing with large-scale video stream processing. Stand-alone systems have limited computing resources, making it difficult to simultaneously support efficient decoding and frame-by-frame extraction of multiple video streams. This is especially true in high-concurrency scenarios, where resource contention and processing delays are prone to problems. Furthermore, stand-alone architectures lack elastic scalability and cannot dynamically adjust computing resources based on business needs, resulting in the system's overall throughput and real-time performance failing to meet actual application requirements. Second, unified management of multi-source alarm events faces technical difficulties. Traditional systems typically adopt independent deployment and decentralized management, resulting in isolated alarm event information and a lack of correlation, making it difficult to achieve cross-system collaborative analysis and unified response. Furthermore, alarm events from different sources may suffer from time asynchrony and inconsistent data formats, further increasing the complexity of event fusion and management. These challenges severely restrict the overall effectiveness and practicality of video intelligent analysis systems. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an efficient distributed algorithm task scheduling method, which realizes real-time frame-level processing of tens of thousands of video streams through key technologies such as parallel decoding of video streams, dynamic load balancing, and multi-model (compatible with common video image recognition models, such as the YOLO object detection algorithm model, which can be specifically applied to traffic management, urban governance, industrial manufacturing, agriculture and other fields, such as pedestrian red light running identification, vehicle illegal parking identification, green space trampling identification, unsafe construction behavior identification, etc.); adopts multimodal fusion detection algorithm to improve the recognition accuracy of complex scenes, and combines with rule engine to realize alarm event classification, providing highly reliable and low-latency intelligent decision-making support for scenarios such as smart cities and industrial safety.

[0005] The present invention adopts the following technical solutions to achieve the invention objectives:

[0006] An efficient distributed algorithm task scheduling method, characterized by comprising the following steps:

[0007] S1: User creates a task;

[0008] S2: The task scheduling system passes the frame extraction task to the video frame extraction system;

[0009] S3: The video frame extraction system sends a request to the streaming platform to obtain the video stream;

[0010] S4: The streaming platform returns the video stream address;

[0011] S5: The video frame extraction system extracts frames and adopts an adaptive frame extraction strategy;

[0012] S6: The video frame extraction system uploads the image to the image resource pool and generates thumbnails using a space-saving strategy.

[0013] As a further limitation of this technical solution, the following steps are also included:

[0014] S7: The image resource pool publishes the image;

[0015] S8: The image resource pool returns the image address to the video frame extraction system;

[0016] S9: The video frame extraction system stores the video address information;

[0017] S10: The task scheduling system initiates a request to the video frame extraction system to obtain camera images;

[0018] S11: The video frame extraction system returns the camera image address;

[0019] S12: The real-time resource monitoring system dynamically monitors the load status of each node in the algorithm service cluster, and the task scheduling system preferentially assigns recognition tasks to nodes with lower current computing resource utilization.

[0020] S13: The algorithm service cluster generates an alarm and uploads the alarm information to the alarm receiving program;

[0021] S14: The alarm receiving program sends a notification to the alarm video caching system;

[0022] S15: The alarm video cache system initiates a video stream acquisition request to the streaming media platform;

[0023] S16: The streaming platform returns the video stream address;

[0024] S17: The alarm video caching system intercepts, caches, and publishes the alarm video.

[0025] As a further limitation of this technical solution, the frame extraction interval formula of the video frame extraction system is:

[0026] (1);

[0027] in: is the gradient change between frames;

[0028] 、 and It is an adjustable parameter.

[0029] As a further limitation of the present technical solution, the thumbnail of step S6 is:

[0030] (2);

[0031] in: is the original image;

[0032] The thumbnail to be generated;

[0033] is the convolution operation;

[0034] G ( s =1.5) is a Gaussian kernel with a standard deviation of 1.5;

[0035] This is a downsampling operation with a sampling interval of 4.

[0036] As a further limitation of the present technical solution, the node load evaluation model of the load condition of each node in step S12 is:

[0037] (3);

[0038] Among them: CPU_util represents CPU utilization, that is, the percentage of CPU capacity used at a certain moment;

[0039] MEM_util represents memory utilization, which is the percentage of total memory currently in use;

[0040] GPU_util represents GPU utilization, that is, the percentage of GPU capacity used at a certain moment;

[0041] The task allocation strategy adopts the improved softmax strategy:

[0042] (4);

[0043] in: Is the selection action probability;

[0044] It's action the value of

[0045] is the temperature parameter;

[0046] is the total number of actions.

[0047] As a further limitation of the present technical solution, the video stream request obtained in step S15 adopts the least recently used-K algorithm:

[0048] (5);

[0049] in: It is the cache elimination strategy;

[0050] Represents a data item timestamp;

[0051] Represents a data item Frequency of visits;

[0052] is a tuning parameter;

[0053] To adjust the access frequency using a logarithmic function;

[0054] The expression combines the effects of timestamp and access frequency;

[0055] Indicates that in all data items , select Make Expression The smallest data items are eliminated.

[0056] As a further limitation of this technical solution, the calculation of the cached video segment duration in step S17 is:

[0057] (6);

[0058] in: Indicates the alarm time;

[0059] Indicates the start time;

[0060] Use the maximum value function to compare 30 and the dynamic time length, and take the larger value as the cache expiration time;

[0061] System key performance indicators:

[0062] End-to-end delay constraints: ; Indicates that the system should ensure that the response time for 95% of requests does not exceed 2 seconds;

[0063] Frame processing throughput :

[0064] (7);

[0065] in: Indicates the total number of frames processed;

[0066] They represent the rates of the three processing stages: network IO, decoding, and encoding rates;

[0067] (8);

[0068] Where: MTBF represents the mean time between two consecutive failures of the system;

[0069] MTTR refers to the average time from when a system fails to when it returns to normal;

[0070] A represents the probability of system availability.

[0071] As a further limitation of this technical solution, the efficient distributed algorithm task scheduling system consists of six subsystems: task scheduling system, video frame extraction system, image resource pool, alarm receiving program, alarm video caching system and real-time resource monitoring system.

[0072] As a further limitation of this technical solution, the efficient distributed algorithm task scheduling system also includes a supporting system, which includes a streaming media platform and an algorithm service cluster.

[0073] Compared with the prior art, the advantages and positive effects of the present invention are:

[0074] 1. This system design offers significant advantages over traditional architectures and can effectively address key pain points in the current field of intelligent video analysis. First, at the system architecture level, traditional stand-alone architectures are limited by hardware resources and processing power, making it difficult to handle the concurrent processing requirements of large-scale video streams. This system, however, utilizes a distributed architecture design that supports horizontal expansion and dynamic resource scheduling, making it easy to achieve parallel processing of tens of thousands of video streams. Through efficient load balancing mechanisms and containerized deployment, the system can flexibly allocate computing resources based on actual business needs, significantly improving overall throughput and processing efficiency while reducing the risk of single points of failure.

[0075] 2. Secondly, in terms of video processing capabilities, traditional systems usually adopt a serial processing mode, and the efficiency of frame-by-frame extraction and analysis is low, making it difficult to meet real-time requirements. This system has greatly improved the frame-level processing speed by introducing high-performance video decoding technology and a multi-channel parallel frame extraction mechanism, combined with GPU accelerated reasoning, to ensure low-latency analysis in complex scenarios. In addition, for multi-source alarm event management, traditional architectures often have information island problems, resulting in scattered alarm events and delayed responses. Through a unified event bus and standardized data format, this system realizes centralized management and intelligent correlation analysis of multi-source alarm events, supports cross-system collaborative response and regularized processing, and significantly reduces false alarm and missed alarm rates.

[0076] 3. Finally, regarding system maintainability and scalability, traditional architectures often rely on customized development and hard-coded rules, making it difficult to quickly adapt to business changes. This system utilizes a modular design and microservices architecture, supporting dynamic loading and updating of target recognition models while providing flexible rule engine configuration for rapid adaptation to diverse scenario requirements. This design not only reduces system maintenance costs but also provides a solid foundation for future functional expansion and technological upgrades. Overall, this system outperforms traditional architectures in performance, real-time performance, scalability, and intelligence, providing a more efficient and reliable video intelligent analysis solution for smart cities, industrial security, and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0078] A specific embodiment of the present invention is described in detail below with reference to the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiment.

[0079] This invention primarily addresses three core technical issues. First, addressing the challenge of concurrently processing large-scale video streams, traditional stand-alone architectures, limited by computing resources and processing power, struggle to efficiently complete real-time decoding and frame-by-frame extraction of multiple video streams. This design, through a distributed architecture and dynamic resource scheduling mechanism (embodied in S12), enables parallel processing of highly concurrent video streams, supports elastic scalability, and significantly improves system throughput and processing efficiency. Second, traditional video analysis systems struggle to balance target recognition accuracy and speed, resulting in low recognition accuracy and high processing latency in complex scenarios. This design utilizes a multimodal fusion detection algorithm (compatible with common video image recognition models, such as the YOLO object detection algorithm, and specifically applicable to traffic management, urban governance, industrial manufacturing, agriculture, and other fields, such as identifying pedestrians running red lights, illegally parked vehicles, green space trampling, and unsafe construction activities) and GPU-accelerated inference technology (using OpenCV technology, integrated with IPP technology at the underlying layer, leveraging the GPU's parallel computing power to accelerate video stream parsing). This approach ensures high recognition accuracy while keeping end-to-end processing latency within acceptable service limits, meeting real-time requirements. Finally, to address the decentralized management and delayed response of multi-source alarm events, traditional systems lack a unified event management mechanism, leading to information silos and collaboration difficulties. This design, through a standardized event bus and intelligent correlation analysis of real-time alarm and task events, achieves centralized management, unified processing, and cross-system linkage of multi-source alarm events, significantly improving the accuracy and timeliness of event responses. Addressing these technical issues provides reliable technical support for large-scale intelligent video analysis.

[0080] Fusion detection algorithm:

[0081] (1) Target detection modeling: B, C = YOLO-Head (F);

[0082] It means inputting the feature map F into the YOLO detection head (YOLO-Head) and outputting the bounding box (BoundingBoxes) and category confidence (Class Confidence Scores);

[0083] F represents the input feature map, which is usually the feature representation extracted by the convolutional neural network (CNN);

[0084] YOLO-Head is the detection head part in the YOLO model, which is responsible for predicting the location and category of the target from the feature map F;

[0085] YOLO-Head usually includes multiple convolutional layers and some specific processing steps, such as anchor box prediction, bounding box regression, and category confidence calculation;

[0086] B represents the predicted bounding boxes, each of which contains the location information of the target (such as center coordinates, width and height);

[0087] C represents the class confidence scores, each of which represents the probability that a bounding box belongs to a specific class.

[0088] (2) Detection loss:

[0089] (9);

[0090] in: Indicates the overall loss;

[0091] Represents classification cross entropy loss;

[0092] represents the CIoU bounding box loss;

[0093] represents the confidence loss;

[0094] GPU accelerated inference technology:

[0095] (1) Use OpenCV's hardware acceleration module (such as cv::cuda::GpuMat) to upload frame data to GPU memory:

[0096] CPU→GPU:Mhost→Mdevice

[0097] The process of data transfer from the CPU (Central Processing Unit) to the GPU (Graphics Processing Unit).

[0098] Host memory (Mhost) refers to memory that the CPU can directly access, typically the system's main memory (such as DDR4 memory). Data in host memory can be transferred to the GPU via Direct Memory Access (DMA) or other mechanisms.

[0099] Device memory refers to the local memory of the GPU, usually the GPU's video memory (such as GDDR6 or HBM). Device memory is memory directly accessed by the GPU and is used to store data required during the calculation process.

[0100] ;

[0101] Represents the original video frame data stored in the CPU memory;

[0102] The core data structure representing an image, including pixel values, dimensions (height, width), number of channels, and data type;

[0103] : Original frame data ;

[0104] : Frame data in GPU memory

[0105] cudaMemcpycudaMemcpy: memory copy function in the CUDA library, used to transfer data from CPU memory (Host) to GPU memory (Device)

[0106] : A three-dimensional tensor representing the spatial resolution and color channels of the image.

[0107] (2) GPU implementation of convolution

[0108] ;

[0109] Represents the output feature map Channels in space The value at

[0110] Represents the first Channels in space The value at

[0111] represents the convolution kernel weight, The convolution kernel corresponding to the output channel is The weights on the input channels, positions ;

[0112] represents the bias term, The offset value of each output channel;

[0113] Indicates the size of the convolution kernel and the spatial size of the convolution kernel;

[0114] (3) GPU utilization

[0115] .

[0116] Active Warps Indicates the number of active warps currently executing on the GPU;

[0117] Warp represents the basic unit of GPU scheduling and execution, usually consisting of 32 parallel threads;

[0118] Maximum Warps indicates the maximum number of simultaneously active thread warps supported by the GPU hardware, which is determined by the GPU architecture (e.g., NVIDIA A100 supports 64 warps per SM).

[0119] The present invention comprises the following steps:

[0120] S1: The user creates a task, and the system generates task metadata, where the priority follows the priority queue management strategy:

[0121] priority_score = w1*user_level + w2*time_sensitivity

[0122] Among them: priority_score indicates the calculation priority score;

[0123] The two weights w1 and w2 are used to adjust the relative importance of user level and time sensitivity in calculating the priority score. The larger the weight value, the greater the impact of the corresponding factor on the priority score;

[0124] User Level represents the user level, which is usually an integer, indicating the level or importance of the user. The higher the user level, the more important the user is, and the higher the priority score should be.

[0125] time_sensitivity represents time sensitivity, which is usually an integer or real number, indicating the time sensitivity of the task or request. The higher the time sensitivity, the faster the task or request needs to be processed, and the higher the priority score should be.

[0126] S2: The task scheduling system passes the frame extraction task to the video frame extraction system;

[0127] S3: The video frame extraction system sends a request to the streaming platform to obtain the video stream;

[0128] S4: The streaming platform returns the video stream address;

[0129] S5: The video frame extraction system extracts frames and adopts an adaptive frame extraction strategy;

[0130] Frame extraction interval formula:

[0131] (1);

[0132] in: is the gradient change between frames;

[0133] 、 and It is an adjustable parameter.

[0134] S6: The video frame extraction system uploads the image to the image resource pool and generates thumbnails using a space-saving strategy.

[0135] The thumbnails are:

[0136] (2);

[0137] in: is the original image;

[0138] The thumbnail to be generated;

[0139] is the convolution operation;

[0140] G ( s =1.5) is the Gaussian kernel with a standard deviation of 1.5;

[0141] This is a downsampling operation with a sampling interval of 4.

[0142] S7: The image resource pool publishes the image;

[0143] S8: The image resource pool returns the image address to the video frame extraction system;

[0144] S9: The video frame extraction system stores the video address information;

[0145] S10: The task scheduling system initiates a request to the video frame extraction system to obtain camera images;

[0146] S11: The video frame extraction system returns the camera image address;

[0147] S12: The real-time resource monitoring system dynamically monitors the load status of each node in the algorithm service cluster, and the task scheduling system preferentially assigns recognition tasks to nodes with lower current computing resource utilization.

[0148] Node load assessment model:

[0149] Load_score = 0.6*CPU_util + 0.3*MEM_util + 0.1*GPU_util(3);

[0150] Among them: CPU_util represents CPU utilization, that is, the percentage of CPU capacity used at a certain moment;

[0151] MEM_util represents memory utilization, which is the percentage of total memory (RAM) currently in use;

[0152] GPU_util represents GPU utilization, that is, the percentage of GPU capacity used at a certain moment;

[0153] The task allocation strategy adopts the improved softmax strategy:

[0154] (4);

[0155] in: Is the selection action probability;

[0156] It's action the value of No explanation

[0157] is the temperature parameter, which is used to control the smoothness of the probability distribution;

[0158] is the total number of actions.

[0159] S13: The algorithm service cluster generates an alarm and uploads the alarm information to the alarm receiving program;

[0160] S14: The alarm receiving program sends a notification to the alarm video caching system;

[0161] S15: The alarm video cache system initiates a video stream acquisition request to the streaming media platform;

[0162] Video caching uses the Least Recently Used-K Algorithm (LRU-K):

[0163] (5);

[0164] in: It is the cache elimination strategy;

[0165] Represents a data item The timestamp is usually the time when the data item was last accessed. The smaller the timestamp, the "older" the data item.

[0166] Represents a data item The access frequency, that is, the number of times the data item is accessed. The higher the access frequency, the hotter the data item.

[0167] is a tuning parameter used to balance the impact of timestamp and access frequency. The larger the value of , the more significant the impact of access frequency;

[0168] The logarithmic function is used to adjust the access frequency so that the impact of the access frequency is not too drastic. The logarithmic function can smooth the change of access frequency and avoid excessive impact of high-frequency data items on the elimination strategy.

[0169] The expression combines the effects of timestamp and access frequency; timestamp The smaller the value (the older the data item), the smaller the access frequency. The higher the logarithmic term The larger the value, the larger the subtraction, making the value of the entire expression smaller. Therefore, this expression tends to eliminate "old" and "not hot" data items;

[0170] Indicates that in all data items , select Make Expression The smallest data items are eliminated.

[0171] S16: The streaming platform returns the video stream address;

[0172] S17: The alarm video caching system intercepts, caches, and publishes the alarm video.

[0173] Calculation of cached video fragment duration:

[0174] (6);

[0175] in: Indicates the alarm time, that is, the time when the system issues an alarm;

[0176] Indicates the start time, that is, the time point when the system starts to perform a task or enter a certain state;

[0177] Use the maximum value function to compare 30 and the dynamic time length, and take the larger value as the cache expiration time; this means that the cache expiration time is at least 30, but may be longer, depending on the difference between the alarm time and the start time;

[0178] System key performance indicators:

[0179] End-to-end delay constraints: ; Indicates that the system should ensure that the response time for 95% of requests does not exceed 2 seconds;

[0180] Frame processing throughput :

[0181] (7);

[0182] in: Indicates the total number of frames processed;

[0183] They represent the rates of the three processing stages: network IO, decoding, and encoding; that is, the number of frames that can be processed per unit time in each stage;

[0184] (8);

[0185] Among them: MTBF (Mean Time Between Failures) represents the average time between two consecutive failures of the system. The longer the MTBF, the more stable the system and the lower the frequency of failures.

[0186] MTTR (Mean Time To Repair) indicates the average time it takes for a system to recover from a failure. The shorter the MTTR, the easier it is to repair the system and the smaller the impact of the failure on the system.

[0187] A (Availability) represents the probability of system availability, that is, the probability that the system will work normally at any time. The higher the availability, the more reliable the system is, and the smaller the impact of failures on the system.

[0188] The method is based on an efficient distributed algorithm task scheduling system, which consists of six subsystems: a task scheduling system, a video frame extraction system, an image resource pool, an alarm receiving program, an alarm video caching system, and a real-time resource monitoring system.

[0189] The task scheduling system is responsible for receiving and processing user-initiated requests and sending them to the algorithm service cluster in the form of tasks.

[0190] The video frame extraction system is responsible for extracting frames from the video obtained from the streaming platform into pictures.

[0191] The image resource pool is responsible for storing images extracted from the image resource pool.

[0192] The alarm receiving program is responsible for receiving the alarm information generated by the algorithm service cluster.

[0193] The alarm video cache system is responsible for caching alarm videos.

[0194] The real-time resource monitoring system is responsible for monitoring the operating pressure of the algorithm service cluster and coordinating the task scheduling system for load balancing.

[0195] The efficient distributed algorithm task scheduling system also includes a supporting system, which includes a streaming media platform and an algorithm service cluster.

[0196] The streaming platform is responsible for providing video service support, such as obtaining video streams.

[0197] The algorithm service cluster is usually composed of multiple servers containing GPUs, and has model computing capabilities, which can recognize the input image information and generate alarms for the corresponding algorithms.

[0198] The process innovation of using a distributed frame extraction system to extract video frames and then calling a third-party general video image recognition model (such as a vehicle illegal parking recognition model, a green space trampling recognition model, etc.) for calculation is the key protection point of this invention, and the process of using alarm video caching technology to extract the video at the time of the alarm is also a key innovative part of this system.

[0199] The above disclosure is only a specific embodiment of the present invention, but the present invention is not limited thereto. Any changes that can be conceived by those skilled in the art should fall within the scope of protection of the present invention.

Claims

1. An efficient distributed algorithm task scheduling method, characterized in that: The following steps are involved: S1: User creates a task; S2: The task scheduling system passes the frame extraction task to the video frame extraction system; S3: The video frame extraction system sends a request to the streaming platform to obtain the video stream; S4: The streaming platform returns the video stream address; S5: The video frame extraction system extracts frames and adopts an adaptive frame extraction strategy; S6: The video frame extraction system uploads the image to the image resource pool and generates thumbnails using a space-saving strategy. S7: The image resource pool publishes the image; S8: The image resource pool returns the image address to the video frame extraction system; S9: The video frame extraction system stores the video address information; S10: The task scheduling system initiates a request to the video frame extraction system to obtain camera images; S11: The video frame extraction system returns the camera image address; S12: The real-time resource monitoring system dynamically monitors the load status of each node in the algorithm service cluster, and the task scheduling system preferentially assigns recognition tasks to nodes with low current computing resource utilization; S13: The algorithm service cluster generates an alarm and uploads the alarm information to the alarm receiving program; S14: The alarm receiving program sends a notification to the alarm video caching system; S15: The alarm video cache system initiates a video stream acquisition request to the streaming media platform; S16: The streaming platform returns the video stream address; S17: The alarm video caching system intercepts, caches, and publishes the alarm video.

2. The efficient distributed algorithm task scheduling method according to claim 1, characterized in that: The frame extraction interval formula of the video frame extraction system is: （1）； in: is the gradient change between frames; 、 and It is an adjustable parameter.

3. The efficient distributed algorithm task scheduling method according to claim 1, characterized in that: The thumbnail of step S6 is: （2）； in: is the original image; The thumbnail to be generated; is the convolution operation; G ( σ =1.5) is a Gaussian kernel with a standard deviation of 1.5; This is a downsampling operation with a sampling interval of 4.

4. The efficient distributed algorithm task scheduling method according to claim 1, characterized in that: The node load evaluation model of the load condition of each node in step S12 is: （3）； Among them: CPU_util represents CPU utilization, that is, the percentage of CPU capacity used at a certain moment; MEM_util represents memory utilization, which is the percentage of total memory currently in use; GPU_util represents GPU utilization, that is, the percentage of GPU capacity used at a certain moment; The task allocation strategy adopts the improved softmax strategy: （4）； in: Is the selection action probability; It's action the value of is the temperature parameter; is the total number of actions.

5. The efficient distributed algorithm task scheduling method according to claim 1, characterized in that: The video stream acquisition request in step S15 adopts the least recently used-K algorithm: （5）； in: It is the cache elimination strategy; Represents a data item timestamp; Represents a data item Frequency of visits; is a tuning parameter; To adjust the access frequency using a logarithmic function; The expression combines the effects of timestamp and access frequency; Indicates that in all data items , select Make Expression The smallest data items are eliminated.

6. The efficient distributed algorithm task scheduling method according to claim 1, characterized in that: The calculation of the cached video segment duration in step S17 is: （6）； in: Indicates the alarm time; Indicates the start time; Use the maximum value function to compare 30 and the dynamic time length, and take the larger value as the cache expiration time; System key performance indicators: End-to-end delay constraint: ; Indicates that the system should ensure that the response time for 95% of requests does not exceed 2 seconds; Frame processing throughput : （7）； in: Indicates the total number of frames processed; They represent the rates of the three processing stages: network IO, decoding, and encoding rates; （8）； Where: MTBF represents the mean time between two consecutive failures of the system; MTTR refers to the average time from when a system fails to when it returns to normal; A represents the probability of system availability.

7. The efficient distributed algorithm task scheduling method according to claim 1, characterized in that: The efficient distributed algorithm task scheduling system consists of six subsystems: task scheduling system, video frame extraction system, image resource pool, alarm receiving program, alarm video caching system and real-time resource monitoring system.

8. The efficient distributed algorithm task scheduling method according to claim 7, characterized in that: The efficient distributed algorithm task scheduling system also includes a supporting system, which includes a streaming media platform and an algorithm service cluster.

Images